1 - LumenVox

Table of Contents
About LumenVox
Why Choose LumenVox
Discover LumenVox
Speech Recognition Engine
Speech enables any application with its flexible API, powering every solution that
LumenVox provides.
Speech Platform
Allows you to develop and deploy your speech application: with just a few steps,
your application can go from conception to reality.
Speech Driven Assistant
Integrates seamlessly with Vertical Communication's TeleVantage IP-PBX, permitting
you to speech-enable the name directory, contact list, voicemail, IVR, and email.
Speech Tuner
Maintains and tests existing applications, ensuring that any speech recognition
application, including those driven by Nuance and ScanSoft⎯continue to work well.
LumenVox Training
Describes classes and instructors available to help you learn about the speech
industry, application design, and tuning with LumenVox’s products.
Application Development Overview
Provides insight to help you develop high-quality and effective applications for your
customer base.
Tuning Guide
Gives a basic overview of the steps required when tuning and improving your speech
applications.
2
4
6
8
14
22
26
36
40
52
1

About LumenVox
LumenVox is a speech recognition company with over a decade of
telephony experience. We utilize a business and technology approach
that allows businesses, corporations, resellers, platform providers, and
service providers access to the complex speech recognition industry.
Our revolutionary speech recognition software products have gained
industry recognition, winning over 17 awards for innovation, technical
excellence, and user's choice.
Whether your organization wants to quickly and easily speech-enable current applications, or
maintain existing ones, LumenVox provides all the necessary tools: our state-of-the-art Speech
Recognition Engine, Speech Platform, Speech Driven Assistant, and Speech Tuner.
Tired of extra costs for adaptations, application updates, and new deployments? At LumenVox, we
believe that you know your company's needs best: our tools allow you to develop speech
applications on your own terms, at your company's pace, without service fees for each new update.
With LumenVox's software suite, you remain in control.
LumenVox Team
Expert. Dedicated. Innovative.
LumenVox's development team⎯speech scientists, speech user interface designers, and experienced
sales and marketing personnel⎯has over 30 years of practical experience in the development and
integration of speech recognition systems, telephony, database design, and hardware integration.
At LumenVox, we know the challenges and opportunities that come with integrating speech into
your application. We offer guidelines, tips, and best practices, and our design team can assist in
any phase of your development cycle. We understand speech.
LumenVox is committed to providing the most powerful, flexible, and useful speech products and
services with excellent customer service to clients of any size.
Recent Awards
LumenVox offers an impressive suite of
versatile, world-class speech recognition
technologies.
Dr. Danny Lange
CEO of Vocomo Software
2 3

Why Choose LumenVox?
Rethinking the Business of Speech
Traditionally, speech recognition technology providers keep the
development and maintenance of speech applications under wraps.
Instead, LumenVox believes in empowering the users and developers of
our speech recognition technology.
Other technology providers create, deploy, and maintain their own proprietary applications for their
clients. For this, the price tag of the final application can be in the millions, due to the extensive
and on-going professional service fees. Traditional companies also have tiered pricing and
functionality for their core Speech Recognition technology. So depending on what you need or want,
the pricing can and will vary greatly. This greatly limits the development and expansion of speech
recognition, preventing smaller businesses from utilizing these products.
LumenVox Provides:
Exceptional customer and technical support
Hardware independent Automatic Speech Recognizer
(ASR) with a distributed client-server architecture
that runs on both Windows and Linux
Extensive logging of audio, grammars, results, and
scores, which allow you to recreate every call
State-of-the-art testing and post-deployment tools to
constantly improve your LumenVox, Nuance, or
ScanSoft speech application
Support and development of current and emerging
industry standards
We provide the tools and education needed when creating applications. We aspire to make speech
recognition widely available, so our business model is structured on a single per port charge, as
opposed to tiered pricing models.
At LumenVox, we believe in the power of speech recognition to help revolutionize many industries,
including businesses that are currently excluded from using speech by tiered pricing and costly
professional service fees. Contact us to learn more about how speech recognition can work for your
business.
I think your product is the tops, and
I’ve been around the block. The MAJOR
KEY TO ME is that I can use it effectively
right out of the box BUT ALSO create
custom speech-activated (true-IVR)
applications right down to the
TeleVantage API level!
Evan Klayman President of Brainstem
4 5

Discover LumenVox
LumenVox’s Products
The speech industry
contains many
different technologies,
platforms, and types
of applications. This
chart enables you to
quickly view the
components of a final
speech application, as
well as the products
LumenVox offers.
Tuning
The process of
changing an
application to improve
performance.
Applications
Speech applications
allow callers to access
any database to get
account information,
perform order entry,
take customer surveys,
or check order status.
Speech Tuner
Complete maintenance tool to
tune and test any speech
application using various ASRs
including LumenVox, Nuance,
and ScanSoft.
Speech Assistant
Complete solution, currently
for TeleVantage owners to
speech-enable their name
directories, contact lists,
voicemails, IVRs, and emails.
Professional Services
Experts in the speech industry
that provide a variety of
services including speech
application design,
development, and tuning.
Pre-Packaged
Pre-packaged applications can
offer features such as email
and calendar management.
Custom Built
Custom speech applications
can address any vertical or
horizontal market need
where touch-tone
applications can not.
Platform
Defines what type of
applications can be
run, and how those
applications are
allowed to operate.
Speech Platform
Complete platform with a
toolkit to design, and deploy
speech applications. The Call
Handler supports any phone
system via analog, digital,
ISDN, or PRI.
VoiceXML
Standards-based speech
application programming
language supported by many
platforms.
Others
Many platforms are available
that use languages, other than
VoiceXML, for creating and
running speech applications.
Core Speech
Technology
These technologies
form the basis for
building any speech
application.
Speech Engine
Automatic Speech Recognizer
(ASR) that supports SRGS,
MRCP, and SISR. Integrated
in various VoiceXML and
proprietary platforms.
ASR
Technology used for
interpreting audio data from
phone, web, microphone, or 2-
way radio.
Other Core Technology
Text-to-Speech is used to
produce audio from text. Voice
Verification is used to identify
individual speakers for security
purposes.
6 7

Speech Engine
LumenVox's Speech Recognition Engine is a flexible
API that performs speech recognition on audio data
from any audio source.
The Speech Engine is speaker and hardware
independent: it supports SRGS and SISR on both
Windows and Linux platforms.
How the Speech Engine works
The Speech Engine provides speech application developers with an
efficient development and runtime platform, allowing for dynamic
language, grammar, audio format, and logging capabilities to customize
every step of their applications. Grammars are entered as a simple list of
words or pronunciations, or in the industry standard Speech Recognition
Grammar Specification (SRGS), as defined by the W3C.
Grammars
Just 13 lines of code (8 calls to the
Speech Engine) will implement a
simple "yes-no" speech recognition
system. The system must provide the
audio and the audio data length for
SoundData and SoundDataLength.
LumenVox’s technology provides us
with some of the very best in speech
technologies today, at the right price for
our customers.
Mark Kelley President of Parallax
Sample Code
void RecognizeSpeech (void* SoundData, int SoundDataLength)
{
const char* GrammarString =
"#ABNF 1.0\n"
"language en-US;\n"
"mode voice;\n"
"tag-format ;\n"
"$yes = (yes | yeah | okay):'true';\n"
"$no = (nope | no):'false';\n";
LVSpeechPort Port;
Port.OpenPort ();
Port.LoadGrammarFromBuffer (0, GrammarString);
Port.LoadVoiceChannel (0, SoundData,SoundDataLength, ULAW_8KHZ);
Port.Decode (0, 0, LV_DECODE_SEMANTIC_INTERPRETATION |
LV_DECODE_BLOCK );
int NumInterpretations = Port.GetNumberOfInterpretations (0);
for (int i = 0; i < NumInterpretations; ++i)
cout

Supporting Standards
LumenVox supports the W3C's Speech Recognition Grammar
Specification (SRGS), part of the VoiceXML 2.0 and SALT
specifications. Companies that track these specifications are
dedicated to the future of speech, and to integrating with
other companies committed to promoting speech recognition.
The LumenVox SRGS implementation is backward compatible
with the existing LumenVox BNF grammar format; current
deployments will leverage the power of the SRGS system
immediately and transparently.
Both companies (LumenVox and
SandCherry) have focused on simplifying
the development, integration, and
deployment of speech services while
maintaining affordability.
Charles Corfield
President and CEO of
SandCherry
LumenVox recognizes that the speech community will need to
work together to develop solutions for businesses, and as
such, LumenVox applications complement the following
technologies:
VXML
VoiceXML (VXML) is a mark-up language designed to code speech
applications with many of the same architectural components as HTML.
VoiceXML platforms connect to a combination of speech recognition engines,
text-to-speech synthesis, telephony interfaces and VoiceXML interpreter
software to process the call. In order to interface VXML with any speech
engine, the engine must understand SRGS and SISR.
LumenVox's Speech Engine is compliant with what VXML expects, and our
engine powers the speech recognition portion of several VXML platforms.
SALT
Speech Application Language Tags (SALT) is similar to VoiceXML but also
adds support for multi-modal systems. SALT extends existing mark-up
languages such as XHTML, XML, and HTML. Similar to our work with VXML,
the LumenVox Speech Recognition Engine conforms to SALT specifications.
Semantic Interpretation
Engine Features and Functionality:
Streaming audio
LumenVox has implemented the W3C's Semantic Interpretation for Speech
Recognition (SISR) working draft, also part of the VoiceXML 2.0 specification.
SISR allows grammar authors to embed snippets of JavaScript code into
Supports English, Latin American Spanish, and Canadian French
their SRGS grammars, to automatically transform what a speaker says into a
format understandable to an application. With LumenVox's Semantic Tags,
Flexible API easily integrates into current OA&M, billing, provisioning, and debugging
callers can say, "September thirteenth two thousand four," and your
systems
application will understand "2004-09-13."
LumenVox is committed to supporting the W3C's working draft. As the draft
evolves, we will support both new and old drafts, so application developers
can be confident that their grammars and tags will perform to specification.
Client/Server architecture distributes speech-processing load
Run time defined grammars entered as simple text, BNF, raw phonetic spelling, or SRGS
Advanced dynamic barge-in adapts to each call in real-time
SDK includes documentation and a demonstration C/C++ application
10 Flexible error recovery through the use of confidence scores and n-best results
11

Advanced Features
Noise Reduction Module
When noise is present, it will degrade the performance of any speech recognition system.
Quality noise reduction improves the accuracy of Voice Activity Detection and Core
Recognition, both essential parts of a speech recognition system.
To improve application robustness in noisy environments, LumenVox implemented a Noise
Reduction Module (NRM) into our Speech Recognition Engine. The NRM automatically
adapts to the acoustic environment, and dynamically updates its estimate of noise levels.
The adaptive algorithm enables the NRM to reduce the effects of noise.
The waveforms below demonstrate the power of LumenVox's Noise Reduction Module. In
the original audio [Fig. 1], a truck driver speaks on a cell phone while driving. In addition
to noise from the truck engine and blowing wind, another vehicle engine starts in the middle
of the recording. Although traditional noise reduction implementations often fail to adapt to
such dramatic changes, LumenVox's NRM adjusts to the new noise characteristics rapidly
and automatically. [Fig.2]
Fig. 1 Original Audio
Truck’s engine noise Another vehicle starting
Fig. 2 Audio after noise cancellation
Reduced truck’s noise Another vehicle starting Adapting to new engine noise
Voice Activity Detection
Voice Activity Detection (VAD), also referred to as barge-in and/or End-Of-Speech (EOS) detection,
identifies when a person begins speaking, finishes speaking, or pauses while speaking.
LumenVox's VAD implementation delivers high performance despite challenging conditions: hisses,
pops, abrupt changes in background noise, telephone line echo, and squawks from two-way radio
communication.
The Voice Activity Detection module is highly configurable and can be adapted to work equally well
within telephone, VoIP, or microphone-based applications.
We are delighted that LumenVox is
extending the capabilities of our
platform.
Media Resource Control Protocol (MRCP)
Speech synthesizers…Audio recorders…DTMF recognizers…Speech
recognizers…Speech verifiers…a fully functioning, media-rich application
needs lots of components to work together. Until now, all of these
components had to be provided by a single vendor, or required extensive
custom programming to integrate them. MRCP changes all this. The
Media Resource Control Protocol allows you to seamlessly manage
diverse media resources. MRCP provides a common language to speak
to all of these devices.
With MRCP, vendors can compete on the basis of their strengths, rather
than attempting to create an all-inclusive, yet mediocre package.
Customers can take the best product from each vendor, creating a
speech application package that is tailored to their particular needs.
For detailed information visit:
http://www.ietf.org/internet-drafts/draft-ietf-speechsc-mrcpv2-06.txt
n-best Results
Instead of returning only the top scoring result, you can
instruct the engine to return several of the highest scoring,
most likely answers, often called n-best results. Returning
n-best results is particularly effective when callers need to
spell names, street addresses, or e-mail addresses.
Without n-best results, if a caller spells a name beginning
with "N," but the engine returns a low confidence score, the
caller would be asked to repeat the letter-and given how
similar "N" is to "M," it's likely that the second answer
would have a similarly low confidence score. With n-best
results, the system can prompt the caller using several of
the likely results, such as "Did you mean 'M,' as in 'Mary'?"
When the caller responds, "No," the system goes to its next
option, "Perhaps you meant 'N' as in 'Nancy'?"
Returning n-best results improves the caller's experience:
instead of asking the caller to simply repeat an answer that
received a low confidence score, the system can confirm the
caller's intention using several likely choices.
Server-Side Grammar
LumenVox offers even more efficient support for
large grammars, by allowing clients to pre-load
grammars onto the server, allowing users to send
the grammar prior to the decode requests.
Typically, the grammar itself accompanies each
decode request, but in the case of large grammars,
sending the grammar to the server prior to
decoding is more efficient⎯reducing network traffic.
John Hibel Vocalocity’s Vice President
of Marketing and Business Development
12 13

Speech Platform
The Speech Platform is an intuitive GUI-based toolkit to quickly design,
develop, and deploy any speech application or IVR. By connecting to
almost any phone system and database, the Platform can easily be
designed for a Speech Driven Technical Support, Call Router, Customer
Service, Dealer Locator, Auto-Attendant, or any other speech
application.
Platform Features:
English, Latin American Spanish, and Canadian French Support
Client/Server functionality
Database Connectivity through Custom Action DLLs
Call Bridging and Outbound Dialing through Custom Action DLLs
Support for Intel Dialogic Dx1E, JCT, Dx2, JCT, and DMV Series cards
Enterprise level distribution
User-created sophisticated grammars
Loop start / Analog or T1 / PRI ISDN / Digital Switch
Barge-In capability
Live updates without rebooting system
DTMF and speech input
Detailed Call Flow logging
Live runtime GUI status monitoring
Complete Call Flow handling
Flash-hook transfer capability
Flexible Call Job definitions
Carrier-grade application-ready
Full SRGS, SISR support
On-the-fly project switching
File or SQL/MSDE Database-based projects
Flexible line setting
Assign each phone line to a different database, file project, or CAPI only mode
Noise Reduction Module
Speech Tuner included
Speech Recognition Engine included
LumenVox’s Speech Platform is a
clear leader in the speech recognition
sector.
Nadii Tehrani Chairman of TMC
14 15

Everything You Need
LumenVox's Speech Platform includes all of the components you need to
produce, adjust, and maintain your speech applications. Designed from
the outset to work together, the Platform's components operate
seamlessly.
Platform Component Descriptions:
The Platform Designer allows you to construct the framework for your speech
application in a GUI environment.
Platform Extensions are used to handle any situation that the Platform Designer
cannot support internally.
The Speech Engine recognizes what the caller says, returning the results to the Call
Handler.
The Call Handler works with the Platform’s Designer, Extensions, and Speech
Recognition Engine, executing the logic of your application, and directing calls
appropriately.
1
2
3
The Call Flow View allows the designer of the speech
application to see all of the modules that are associated with
the application and how they relate to each other. This is
where the application dialog flow is created and can be
tracked to see how callers will flow through the speech
application or IVR.
The Objects Panel allows application designers to drag and
drop Modules that perform a specific function, or to add
notes with Annotations to provide quick information or
reminders on the Call Flow itself. Actions Lists, Actions,
and Grammars are drag and drop icons to add within each
individual Module's properties.
The Properties Window has six tabs: Project, Modules, Audio
Library, Quick Audio, To Do, and Notes. Project shows
theglobal and project specific properties. Modules is a list of
modules you have created so far for the application. The
Audio Library and Quick Audio displays all audio prompts
contained within the project. To Do reminds you that there
are objects that have not been finished within the whole
project. Notes provides a place to enter comments or
thoughts concerning the applications.
The Speech Tuner provides a comprehensive window into your application: it allows
you to quickly note where your application performs well, and determine which
areas need improvement. The Tuner also lets you simulate changes to your
application, using audio from past calls, to determine the effectiveness of each
change.
2
LumenVox allows us to have control
over the development of call flow,
grammars and tuning, while keeping the
costs at a respectable level.
Brian Lauzon President of TelASK
3
1
16 17

Platform Extensions
The Speech Platform can be extended, allowing you to create a
number of speech applications, by accessing pre-built libraries and
controlling the call flow. These Platform Extensions can be written
as either a Visual Basic ActiveX exe or C/C++ DLL.
Examples of Platform Extensions:
Connect to a live, or regularly updated, website or RSS
feed
Connect to a database via an ADO (ActiveX Data
Object)
Use SRGS for grammars and enable semantic
interpretation
Direct callers to different modules within the
Application Designer project based on external
decision trees
Example Applications:
Pin code and account number capture
Survey systems
Automated billing
Collecting demographic information
Auto-attendant
Transaction completion
Appointment scheduling
Email reader
Voicemail access
By utilizing LumenVox’s Speech
Platform our callers actually enjoy the
experience of the phone call, helping us
to build a good relationship with our
customers.
Derek Henry CEO of 1-800-US-LOTTO
Corporation
VB ActiveX Platform Extension Example
18 C/C++ Platform Extension Example 19

Call Handler
The Speech Platform's Call Handler runs the speech application. The Platform's Settings, and
more specifically Line Settings, inform the Call Handler as to which Project to use.
CallHandler
The Call Handler supports a variety of Intel Dialogic telephony
cards, and is designed to make hardware setup as easy as possible;
this allows more time for you to develop and test your speech
application.
When developing your speech application with the Speech Platform's Designer, run tests by
clicking on the Test button in the Toolbar. The Call Handler allows you to use speakers and a
microphone, to completely test your application in-house⎯before releasing it to customers.
Handler
Design Tips
Build the application for the expected case, anticipating the caller's situation and the
application's goals. Focus on these goals as you design the call flow and dialogues.
Keep prompts and grammars concise in your initial design, then expand based on
callers' interactions, learned from the tuning process.
Give thought and time to determine the persona and brand image. Verify that the
voice talent creates the proper tone and perception for your company.
Make the system match the user. Listen to the way a user responds and interacts
with the system.
Apply the strengths of speech recognition to your application: remember that many
of the hierarchical structures used in DTMF or touch tone call flows are not
appropriate in speech applications!
The reason we chose LumenVox’s
speech technology was simply because of
the flexibility it provides us, as developers
Never disguise a list question as a yes/no by adding unnecessary pauses, as in,
"Would you like Red (pause), Blue, or Black?" This leads to caller confusion.
Brian Lauzon
President of TelASK
Adjust the "No Input" timeout to match the complexity of the question.
Make explicit decisions with yes/no.
Make it easy for the caller to leave the speech application and reach a live person.
20 21

Speech Driven Assistant
LumenVox's Speech Driven Assistant for TeleVantage is a complete
turn-key program to speech-enable name directories, contact lists,
voicemail, IVRs, and emails, providing all callers hands-free phone
interactions. The system also allows for alternate names, nicknames,
and various spellings and pronunciations to be recognized.
LumenVox’s Speech Driven Assistant
for TeleVantage has become the most
valued component in our operation of a
speech-activated hotline, 1-800-US-LOTTO.
Derek Henry
CEO of 1-800-US-LOTTO
Corporation
Remote Access to:
Speech enabled outbound dialing
Users can say a name from their personal contact lists when they want to place a call
Speech enabled name directory
Callers can speak the name of the person or department they want to reach
Support for workgroups and multiple company configurations
Transfer fax calls to specified extension
Add alternate names, spellings, and pronunciations
Speech enabled voicemail access
Access, control, and manage voicemail with speech
Reply directly to caller's phone number or extension
Forward messages to other users
Speech enabled IVR (LumenVox's Speech Platform)
GUI-based development tool to create a variety of IVR applications
Supports both Speech and DTMF input
External database access
Speech enabled access to email
Access POP3, Exchange Server, and IMAP email
Reply to email in an attached recorded audio format or with user predefined text messages
NeoSpeech Text-to-Speech (TTS)
Play prompts of un-recorded proper names, dynamic IVR applications, and emails
Assistant Features:
Speech activated dialing to contact list
Transfer to another extension from voicemail
Navigate, forward, save, and delete voicemail
IMAP, POP3 and Exchange server email access
DTMF and Speech input
Configure Speech Driven Assistant database remotely
Barge-in capability with CSP compatible Intel Dialogic cards
Live run-time GUI status monitoring
Speech Platform included
Speech Tuner included
NeoSpeech Text-to-Speech included
22 23

Configuration Tool
User’s email account information and pre-defined
text message replies are easily managed.
LumenVox’s Configuration tool is where all modifications to the Speech
Driven Assistant are made.
The Configuration program allows
administrators to modify user
information, maintain email
servers, configure speech
applications, and check the
server’s telephony hardware
compatibility.
Administrators can easily select and
change any user’s information, which is
continuously synchronized with
TeleVantage’s database.
The Configuration tool
can also discover and
display all Intel Dialogic
cards installed. All
features associated with
the telephony cards are
shown as a quick
reference of its
capabilities.
Alternate names or departments
associated with a particular user can
be added quickly to the live system.
For commonly mispronounced
names, alternate
pronunciations can be added.
The Phonetic Speller provides
alternate pronunciations as
well as the chance to hear
how it sounds, using Text-to-
Speech.
LumenVox’s Speech Driven Assistant
has provided the best integration to
TeleVantage of any other speech
recognition product we tested.
John Gagliardi
President of GTI Solutions
24 25

Speech Tuner
LumenVox's Speech Tuner is a complete
maintenance tool for end-users, valueadded
resellers, and platform providers.
It’s designed to perform tuning and
transcription, as well as parameter,
grammar, and version upgrade testing of
any speech application.
With this GUI-based tool, companies
developing speech applications on
various ASR platforms (including
Nuance and ScanSoft) can bring speech
application tuning in-house and avoid
professional service fees.
LumenVox is on the cutting edge of
speech technologies and customer
satisfaction by supporting not only their
own speech engine, but other leaders in the
industry as well.
Bruce Balentine
Executive Vice President
and Chief Scientist at EIG
Why Do Companies Need the
LumenVox Speech Tuner?
No untuned speech application
survives contact with actual
customers. Tuning is an absolute
requirement for every speech application
deployment. Our Tuner allows you to
quickly assess changes and upgrades.
Tuner Capabilities:
Evaluate and improve the speech recognition
application
Analyze each stage in the call process
Transcribe audio data, make pinpoint adjustments,
and immediately measure the effects on performance
Test changes against actual calls immediately
Analyze data collected using different ASR engines
Test design and development decisions of new
applications, using data from deployed applications
26 27

Tuner Overview
The Speech Tuner is comprised of several key functions and windows.
SRE Custom Tags
displays log
information supplied
by the application’s
developers.
The Call Log displays the list calls and controls the display
of information in the rest of the Tuner. Each interaction
(or turn) in a call is marked with an event type, such as:
: Beginning of the call
: Speech event
: Touch tone event
: 'No input' event; the application
did not detect any speech or touch tones
: 'Unknown' event; these
events are typically ASR-specific
: End of the call
The Transcription tool allows a
transcriber to type the text of the
caller's speech. Transcriptions
are automatically evaluated and
stored in the database for use in
performance evaluations.
The Statistics window
displays performance
statistics directly related
to the calls in the Call Log
window.
The Answer window shows
recognition results. Full
n-best support is available,
as well as the semantic
interpretations, and actual
words recognized.
28
Listen to the application prompt, and the caller's pre- and postprocessed
speech. The recognized words are displayed where the
ASR found them within the caller's speech. Vertical bars, which are
useful for detecting problems, indicate the beginning and end of each
word.
29

Tuning Processes
LumenVox's Speech Tuner provides full support for LumenVox's Speech
Recognition Engine, Nuance 8.5, ScanSoft OSR 2, and other ASRs. The
Speech Tuner allows you to work with any supported ASR via a single
interface.
LumenVox is an active supporter of the Tools committee in the VXML
Forum, and is working to help define standard logging information, to
help ease the tuning process.
The tuning process involves three easy steps:
Import Data.
1
2
3
The basic process is simple. Users import call log data into
the Speech Tuner database. All information stored by the call
log is available in the Speech Tuner. In most cases, log fields
between ASR engines are very similar; when the information
differs, every effort is made to preserve the original data.
Each special case is fully documented.
Transcribe Speech.
Transcribers can type the text of the caller's speech directly
into the Speech Tuner. Once the audio is transcribed, the
Tuner compares audio transcripts with the speech engine
results to determine accuracy, greatly reducing errors
associated with hand evaluations. If semantic interpretations
are available, the transcriber can also mark whether the
semantic interpretation was correct or incorrect. The
transcripts are evaluated using the actual decode grammar,
producing measurements such as word-error-rate, in- and
out-of-grammar rates, and semantic error rates.
Test Immediately.
Selecting an interaction in the Call Log automatically loads
the associated audio and grammar into the Tester. The
grammar can be edited, speech engine parameters set, and
individual recognition tests generated. The Speech Tuner
natively supports industry standard SRGS grammars. Once
a set of possible changes is identified, users can batch test
audio to evaluate performance, using those changes.
The Speech Tuner assumes the user possesses licensed
versions of the relevant ASR, that the ASR platform is up and
running, and that the platform is able to accept connections.
LumenVox Speech Tuner Database
The Speech Tuner communicates with an open-source, freeware database called SQLite
(www.sqlite.org). The Speech Tuner manages call log importing, searching, and exporting⎯so
users can focus on the task of tuning, not log management. The database is contained in a single
file, is easy to back up and transport, and can be queried using SQL-92 (see the SQLite website for
full details) from a variety of exterior tools. Other speech engine vendors are free to convert their
native logs to ones the engine understands. The format, content, and semantics of the LumenVox
Speech Tuner database are published.
The database maintains all the information contained in the original call log. The Speech Tuner
includes not only the decode grammar and ASR results, but also the decode platform, parameter
settings, alternative results, prompt audio, and pre- and post-processed audio.
Depending on the platform logging capabilities, the database can provide more advanced
information, such as ASR result alignments within the audio; the list of phonemes used in the
decode; and word, utterance, and semantic interpretation confidence measurements.
In addition, the Tuner stores all transcripts and evaluations within the call log. As transcripts are
entered into the Speech Tuner, they are automatically evaluated against the decode grammar.
These transcripts, and any notes or additional information, are stored directly into the database.
Individual scores⎯such as word error rate, semantic error rate, and in- and out-of-grammar
measurements⎯are stored along with their alignments, as well as information about how the scores
were reached.
Users can generate a variety of reports from these results, including error rate by grammar or
dialog, confusion matrices, transcription progress, and confidence thresholds for confirmation or
rejection settings.
In the future, LumenVox's Speech Tuner will also support back-end database replacement, for use in
enterprise level systems, where multiple users will be analyzing the same data simultaneously.
Companies who use an ODBC-capable database can replace, with certain SQL changes, the diskbased
SQLite system with an enterprise system such as MS SQL Server 2000, MySQL, PostgreSQL,
and/or Oracle.
LumenVox has created speech
recognition products that are easy to
code with and GUI-based tools, such as
the new Speech Tuner that greatly
simplifies post-deployment
maintenance.
Vern Baker
President of enGenic
Corporation
30 31

Taking Out the Guesswork
Make changes to grammars, parameters, or ASR engines, secure in the
knowledge that those changes will make the application better, faster,
and more accurate. The Speech Tuner uses historical information to
validate your changes, ensuring your success.
Grammar Tester
Most 'tuning' tools are passive log viewers, requiring that changes be
made in the live speech application and retested over a period of time
with live callers. With LumenVox's Tuner, we send the changes to the
Speech Engine, simulating the recognition process and evaluating
changes instantly. Instead of slow, non-interactive, static tuning, the
Speech Tuner enables on-the-fly, highly interactive, dynamic tuning.
Make a change, do the test, get the results!
The Grammar Tester is a dynamic
testing component. You can switch
ASR engines, grammars, and engine
search parameters on-the-fly, and test
changes in single or batch tests.
Grammar Evaluation
Evaluate speech and grammar sets against the speech engine, as they
took place during the actual call. Adjust grammars and instantly
re-test and re-score to evaluate improvements in performance. With
LumenVox's Speech Tuner, you can instantly determine whether adding
a new phrase to the grammar will improve your accuracy.
Parameter Evaluation
Setting parameters optimizes the speech engine performance, further
improving the caller's experience. Traditionally, changing ASR
parameters is a difficult and time-consuming task, often requiring long
delays between changing a parameter, and evaluating its effects on
performance. Our Speech Tuner can dramatically shorten the process.
The dynamic test capability of the LumenVox Speech Tuner allows the
user to quickly make and test parameter changes: now, ASR engine
parameters such as search optimizations, speech end-pointing, and
n-best result processing can be easily adjusted, and immediately
re-tested and re-scored from within the Speech Tuner.
Performance Measurements
The Speech Tuner rates performance against commonly accepted
measures like WER (Word Error Rate), Grammar Coverage, and
Semantic Interpretation matching. This helps give an accurate picture
of details such as average confidence scores, correct versus incorrect
responses, and In-Grammar versus Out-of-Grammar performance.
Assessing Upgrades
Installing new versions of platforms and ASR engines entails a certain
risk with each new upgrade. On occasion, new default settings, search
routines, changes to acoustic models, and so on will actually worsen
the caller's experience, until the application is re-tuned. But using the
LumenVox Speech Tuner, you can perform baseline testing with the old
version to establish the minimum acceptable performance. Then,
using the upgraded version of the ASR engine, you can easily re-test
all existing data and compare the results to the baseline. The new
performance, judged against the baseline, gives you the information
you need to make a decision, and deploy an upgrade with confidence.
32 33

Tuner Reports
The LumenVox Speech Tuner defines several pre-built queries for most
common reports. The reports are generated using SQL queries against
the Tuner database, with results produced in a pre-defined XML format.
The format, content, and semantics of the LumenVox Speech Tuner
database are published: if you need to extract data in logs that are not
provided in the Tuner interface by default, you can easily produce
custom reports by writing SQL commands.
Tuner Tips
When a problem occurs with a transaction,
determine if the fault lies primarily with
prompts or grammars.
Never make a change in your call flow or
grammar for just one failed call.
Train acoustic models with environment
noise in play.
Train acoustic models to allow for caller
dialects and regional pronunciations.
Tune grammars, prompts, and all system
parameters.
Remember that accurate transcriptions
need to account for noise.
Speech Understood
Speech Understood
The Speech Tuner is an excellent fully
integrated tool for improving speech
applications.
Bruce Balentine Executive Vice President
and Chief Scientist at EIG
34 35

LumenVox Training Courses
The LumenVox Team, a group of knowledgeable
professionals with extensive development and
technical support experience, provides our
courses. We will help you get the most out of
your LumenVox products.
Courses include Speech Application Design, API Development,
Speech Application Tuning, and many others. These courses give
developers and business personnel opportunities to learn about
speech development and sales opportunities on our premises, with
the assistance and advice of the LumenVox Team.
Who Should Attend
People responsible for developing, maintaining,
marketing, and/or selling LumenVox speech
applications will benefit from LumenVox training.
Classes also benefit anyone with an interest in
designing, developing, tuning, testing, or maintaining
any speech telephony system.
What to Expect
LumenVox training is key to accelerating your learning curve. Through a combination of
presentations and hands-on exercises, our courses provide the details of creating and maintaining
applications. In these courses, you will learn solutions to real problems encountered during actual
application design, development, deployment, tuning, marketing, and selling.
LumenVox training will give you the guidance you need to successfully design, develop, deploy, and
refine your applications. We tailor our trainings to meet your particular needs.
About the Instructors
Our team of expert instructors is committed to your success. Every LumenVox instructor has a
background in computer telephony, application development, and speech recognition. We are
familiar with the development challenges you will encounter on a daily basis; we will offer solutions
to routine problems, as well as creative approaches to not-so-routine problems.
LumenVox Support Services: Ensuring Your Success
At LumenVox, we recognize that high quality, cost effective technical support is a crucial component
of successful application development. With proper support, subscribers gain a deeper product
understanding, resulting in enhanced productivity, and ultimately in greater customer satisfaction.
With this in mind, LumenVox offers a simplified technical support system designed to meet varying
customer needs. LumenVox technical support plans are available for VARs, Distributors, and End
Users with ongoing projects/support needs. The key component of LumenVox technical support is
the Customer Hotline. Two additional avenues are also available: Fax Support and Email Support.
Whichever method you choose, know that LumenVox will work efficiently to answer your questions
and resolve the problem.
Our LumenVox technical support team is made up of knowledgeable
professionals with extensive LumenVox development and support experience.
We are well versed in computer telephony technology and are available to
assist you with:
General LumenVox Technical Assistance
Timely Problem Resolution
Product Installation Assistance
LumenVox Database/Host Connection Assistance
Intel Dialogic Hardware Optimization
I really appreciate the cooperation
and assistance that we received from the
LumenVox engineers. They are an easy
group to work with.
Chris Riggenbach CXM Product
Development Manager
36 37

Our Partners Include...
Keeping People Connected
LumenVox and SandCherry Speech Enable 2-way Radio Networks
SandCherry's Voice4 Radio Message System (RMS) dramatically improves workgroup efficiency for
teams with 2-way radio, phone, and Web users by providing easy-to-use voice and data messaging.
The ability to offer messaging to 2-way radios⎯one of the most widely used tools for mobile work
forces⎯adds an entirely new dimension to workgroup communications.
Leaving messages for radio users when they are unavailable, and providing access to the same
system for phone and Web users to leave and retrieve messages, provides the critical bridge
between what had been independent communications networks. The system also offers phone
users a patch capability to connect to the radio network for real-time communication with radio
users. Voice4 RMS equips a workgroup to improve its efficiency and performance using tools and
processes already in place, without relying on a dispatcher.
Using LumenVox's Speech Engine, SandCherry's Voice4 RMS provides an unparalleled level of
voice-driven functionality for mobile workgroups. Voice4 RMS is easy to install and use, and
provides a cost-effective, robust communications solution.
Making Travel Easier
Los Angeles World Airports (LAWA) using LumenVox Engine and enGenic technology to
speech-enable travel information
Los Angeles World Airports, including Los Angeles International Airport (LAX), has implemented the
first fully speech enabled voice response system for a major airport network. In compliance with
Homeland Security regulations, LAWA helps callers access the most up-to-date flight information
directly from the airport hotline, rather than calling individual airline carriers.
Now callers can check the status of their flights, get information about parking, ground
transportation, services for people with disabilities, inquire about lost and found items, contact
administrative offices, and get directions to the airport⎯all through an automated speech system.
The hotline was created using enGenic's development and engineering tools, and LumenVox's
Speech Engine. Jim Coulter, CEO of enGenic, states, "LumenVox provides us with fast and accurate
recognition of speech grammars, in an ever-changing world environment. Callers around the world,
with different accents, can use simple English commands to obtain information from all four
airports, over 185 airlines, 40 different taxi and limousine services, and 20 different airport
departments."
The new system will enable us to
continue our rapid growth and at the
same time improve the efficiency of our
ordering system without adding
additional staff.
Mike Gilson
Vice President of ATA Retail
Services
Helping Doctors Treat Patients
LumenVox Speech Engine Used in Innovative Patient Follow-up System
The University of Ottawa Heart Institute in Ottawa, Canada has implemented an innovative
automated patient follow-up system developed by LumenVox partner TelASK Technologies Inc.
(www.TelASK.com). The TelASK System incorporates the LumenVox Speech Engine, allowing the
Ottawa Heart Institute to closely monitor the progress and recovery of surgical patients by placing
an automated call to their homes on day three, and again on day ten, after the patients have been
discharged. The speech enabled outbound dialing system automatically phones the patient and
delivers a pre-set list of questions. Each question is a strong indicator of the patient's progress.
Using a specially designed algorithm, patients' answers are grouped as either requiring an
immediate call back, contact within a day, or progressing normally. In the event of a response that
may indicate a problem, the system will hold callers on the line and connect them to an Advance
Practice Nurse at the Heart Institute for immediate attention. The University of Ottawa Heart
Institute will also use this system to monitor patients with Acute Coronary Syndrome, and patients
participating in their Smoking Cessation programs.
38
39

Types of Speech Recognition
Speech recognition is used in a wide range of applications, from
automated commercial phone systems to enhancing personal
productivity. This technology appeals to anyone who needs or wants a
hands-free approach to computing tasks.
There are two main types of speaker models: speaker independent and speaker dependent.
Speaker independent models recognize the speech patterns of a large group of people. Speaker
dependent models recognize speech patterns from only one person. Both models use mathematical
and statistical formulas to yield the best word match for speech. A third variation of speaker
models is now emerging, called speaker adaptive. Speaker adaptive systems usually begin with a
speaker independent model and adjust these models more closely to each individual during a brief
training period.
Leveraging the Power of Speech
Although many companies' first instincts are to simply speech enable their existing DTMF
applications, doing so does not leverage the power and strengths of speech: speech enabling a
DTMF application will not make the system smoother, faster or easier-to-use.
The combination speech/DTMF system lengthens already complex menus, by adding the "press or
say" routine⎯"For Checking, press or say 'one,'" or even worse, "For Checking, press one or say
'checking'." The typical combination speech/DTMF system requires the caller to remember too
much, putting undue burdens on the caller.
Migrating a DTMF application and its prompt design does not fully utilize the conversational aspect
of speech. For speech applications to perform well, the call flow and dialog design is crucial.
Designers must study the user patterns of the existing system, so they can redesign prompts,
menus, and change the steps of the call flow to make the experience faster and more pleasant for
callers.
Well-designed speech applications offer many advantages over the combination speech/DTMF
systems.
By using a speech enabled system, our
merchandisers realize significant time
savings through 24-hour-a-day telephone
access to information, from delivery status
to new delivery days and product
opportunities.
Mike Gilson Vice President of ATA Retail
Services
Speech is:
More Human
With speech, prompts are phrased as easy
questions, and callers can answer simply and
naturally⎯with their voices. Speech systems
provide a more natural interface than touch tone
menus.
Smooth and Fast
Good speech call flow designs permit callers to
get what they need faster, without having to wade
through cumbersome filter menus.
Easy-to-Use
Navigation is much simpler, and callers can use
the application with the interface mechanisms
they are most familiar with: their voices.
More Personal
Speech applications give the impression of the
ideal employee: attentive, empathic, alert, and
consistently agreeable, rather than an impersonal
string of numbers and tones.
When to Use DTMF
Sometimes, DTMF is appropriate: as an errorhandling
backup, or in special, security-sensitive
interactions, such as pin code or credit card entry.
In terms of customer satisfaction for most calls,
speech applications outperform DTMF. Rather
than speech-enabling an existing DTMF system,
design your application with a conversation in
mind⎯and learn to leverage the power of speech.
40 41

State of the Industry
To get an accurate understanding of the current state of the speech
industry, we must first look at the history of Interactive Voice Response
(IVR) systems.
Companies have attempted to handle customer interactions with touch tone IVR systems or live
agents. Yet, most customers become frustrated with ineffective DTMF interfaces, or hang up while
holding for a live agent. In order to support customer interactions more quickly and efficiently,
companies began to request speech recognition interfaces.
This movement towards speech provided the speech recognition industry with tremendous growth
potential, however, many companies consider speech to be in the early adopter stage of the market.
Why is that?
While people have been hearing about speech for decades, only in the past decade or so have
advances in the technology and supporting hardware allowed speech to finally become a viable
option, with most performing tasks with over 90 percent accuracy. During this period many Fortune
500 companies implemented speech recognition, and helped educate consumers on how to interact
with speech applications. These applications have become so advanced and mainstream, that
businesses⎯both large and small⎯now turn to speech solutions for everything from basic autoattendants
to more complex order-taking systems.
Vendor Selection Tips
Select a partner with the technological and business
expertise that best suits your company and future
projects. Ensure that the partner you choose will
provide all of the services and products you'll
need to be successful.
Search for tools that allow for every change and
adjustment to be automatically and rigorously
tested, with actual historical call data.
Include a tool in your development process that
verifies any dead ends or unfinished work.
Choose technologies that best fit with your application
requirements.
Select partners with deployment experience.
Verify the level of technical support provided⎯ensure that
you will receive the support you need.
Speech recognition offers a great solution for large and small businesses: it simplifies customer
interactions, increases efficiency, and reduces operating costs.
Analysts at Cahners In-Stat and Giga report that calls that are handled by live agents can have an
average cost-per-call from $2 to over $15. With speech recognition, the average cost-per-call can
be cut to $.20 per call or less.
LumenVox’s corporate and product
strategy is right in sync with us.
Since this technology appeals to anyone who needs or wants a hands-free approach to computing
tasks, it is becoming a standard software option. At LumenVox, we focus on developing tools that
will empower the user to build, customize, maintain, and refine their own applications.
Vern Baker
President of enGenic
Corporation
Speech recognition system development is still something that requires time to prepare and
monitor. With tools like LumenVox's Speech Platform and Speech Tuner, we are continually working
to simplify the aspects of speech application development⎯to help your business get the most out
of speech recognition.
42 43

Effective Design = Customer Satisfaction
Speech recognition applications come in many varieties, from simple
routers to complex ordering systems. What designers must remember is
ease-of-use. Even with a complex system, callers must be able to
navigate through the system easily.
Speech applications allow customers to accomplish their goals quickly and easily. Much of the
internal work is in the design phase: building the call flow, creating grammars, recording prompts,
and conducting usability testing. Speech application designers will modify each aspect throughout
design and internal testing phases.
But with all the speech applications on the market today, and most prominent speech companies
boasting recognition accuracy in the high 90%'s, why do so many people feel that "speech doesn't
work?"
Usually, it's because callers don't know what to say.
You can avoid common problems by carrying
out the following steps during the design phase:
1 Research Needs and Create Initial Design.
First, speak to the people who currently answer phone calls, and get their
input. What current questions and interactions could potentially be automated?
If the company already uses a DTMF application, how well does DTMF handle
these interactions and questions? Not all interactions match well with speech's
capabilities, so initial research is critical.
Next, sketch out a potential progression of the call flow, and share this with
others to make sure the progression makes sense and that callers can quickly
and easily navigate the system.
2Develop Prompts and Grammars.
Designers should decide how much of a "natural language" system callers need
or desire. "How may I help you?" only works if callers know precisely what
they want, and the designer can accurately predict their responses. Generally,
callers will need some guidance and cues as to what to say. A "How may I
help you?" question involves more extensive grammar development and testing
than a question like: "We offer three choices, A, B, or C. Which would you
prefer?"
The customer service that we have
received from the LumenVox team has
always been top of the line and we are
looking forward to continuing our
partnership for many years in the
future.
The application developer must keep the system as conversational as possible,
but prevent callers from treating the machine as a human. When callers think
the system "actually" understands, or think that it possesses a greater
vocabulary than it really does, they get lost⎯or make requests that are outside
the system's capabilities. These problems rapidly compound, resulting in caller
frustration and dissatisfaction. Effective, intentional, clearly designed prompts
and grammars help keep customers satisfied while managing cost.
3 Test with Real Customers.
The ultimate measure of an application is the first live deployment of the
system. The first live deployment must be a test version with actual users, not
the programmers who are intimately familiar with the application's design.
This will be the first time that assumptions about user behavior will be
seriously tested; the resulting data will allow designers to modify the
application to meet the caller's needs.
4 Tune with Real Data.
Test deployments permit designers to fine-tune the system, often resulting in
significant changes, if they review actual caller experiences. By refining
prompts, grammars, and call flow design, the application will become more
robust, error-free, clear, and effective⎯in short, an application that customers
will want to use.
To tune the application effectively, all of the components of initial
design⎯prompts, grammars, call flow, and the persona of the call
system⎯must be tested. Since these elements are often built separately,
Victor Salazar President of Technical
developers must ensure that all of the parts combine effectively in the testing
Support Systems
phase to achieve the desired effects. Properly tuning the speech application
involves a thorough assessment of initial design components and real caller
44
interactions.
45

Error Handling
Make your speech system more efficient and usable by optimizing error
handling.
Focus on the basics: Is the system accurate? Does the caller achieve call completion? Does the
caller like using the system, and want to continue doing business with your company?
An optimal application addresses these issues by combining technology with art. Mixing technical
aspects, like programming and testing, with aesthetic elements, like writing, casting, and coaching,
reduces errors and increases customer satisfaction. Give sufficient consideration to each part of
the development process.
Understand the speech recognizer that you are using⎯the confidence scores it returns allow you to
make good decisions about the call flow. Track confidence scores at the project, grammar, and
single call levels, to set both static and dynamic thresholds. This will permit the system to make a
good decision on whether or not to confirm.
Remember that it is better to confirm than to make a mistake. Although confirming can be
unpleasant for callers, it is preferable to the frustration of being lost. Figure out when you need to
confirm using the confidence scores, and try to make the confirmation prompts less complex than
the original prompt. If you choose your confirmations wisely, even though it takes a little longer,
users will not become irritated or impatient, and will get to where they need to be.
Help Callers Avoid Frustration…
Errors occur because callers get lost⎯or are unsure of what to say. In either case, this
is usually a prompt issue. Effective prompt writing guides callers to say what is in the
grammar. Another option is to make the grammar robust enough to handle reasonable
requests, even when the caller's phrasing is clumsy.
No matter the style, the prompt should always focus on the task at hand: moving the
call forward.
Sequential prompts should be connected with transitions, e.g., first, next, finally.
Prompts can also use related questions, as in "Please tell me your account
number…And your pin code."
Although long prompts should generally be avoided, sometimes a few extra words pay
off. Phrases like "Just to confirm," "Almost finished," and "So I can help you better"
create a forward mental model…in these cases, reassuring the caller is more beneficial
than the few seconds you gain by clipping the prompt.
Whatever your company's style or needs, LumenVox can help.
Solving Problems
When an error happens, fix it! And don't let it happen
again.
Sounds simple enough, but how? And what will
customers do in the meantime?
Figuring out why an error happens is the key to fixing
it. The feedback callers provide is often vague;
instead, go straight for empirical data. Examine the
context of the call to help pinpoint what caused the
error: too much background noise; not enough
volume; mispronunciation. Invariably, errors will
occur. To provide great customer service through
speech applications, we need to minimize errors, and
make the caller as comfortable as possible when errors
do occur.
Error-handling interfaces often increase caller
frustration. Recognize any of these?
Adversarial error responses:
"I need you to be more specific."
Generic responses:
"I'm sorry, I didn't understand."
Annoyingly enthusiastic responses:
"Let's try it again!"
When developing speech applications, most companies strive to
achieve a balance between saving money on live agents and
providing better service than traditional DTMF systems.
Unfortunately, callers are sometimes uncomfortable with speech
applications⎯they try to talk as they would to a human, or worse,
speak in a stilted way, because they think computers will better
process their requests. Successful experiences with your call
system will help increase caller confidence⎯and when errors
occur, callers will be more patient if they have had positive
experiences in the past.
Improving the caller experience is about good service.
Thinking carefully about the error interface, designing
effective prompts, testing the call system often, and
examining the context of errors can help improve
the caller's satisfaction with the speech
application…and with your business.
We are delighted that an industry
leader like LumenVox has met the market
demand with a product specialized for our
TeleVantage platform.
Rob Black Product Marketing Manager
of Vertical (formerly Artisoft)
46 47

Voice Matters
The voice of your speech application is your company's
first representative. Choose the voice wisely. You're not
just looking for a type of voice; you're looking for an
emissary. The voice needs to be able to explain, inspire,
soothe, excite, and above all else, sound sincere.
Assuming the Voice User Interface (VUI) design is good, the prompts must be
recorded to best represent that design. The four essential pieces for great
prompt recording include:
Great Casting
Great Directing
Great Concatenate Recording
Great "Voice"
Think about your desired call experience⎯and consider using professional
voice talent rather than your receptionist. Professionals will provide
better sounding prompts: they will know appropriate variations in
pitch, rhythm, speed, duration of pauses, and elongation of words.
In addition, professionals will avoid novice errors like wobble,
nervous tics, sloppy diction, and colloquial pronunciations. Most
importantly, talent is directable; professionals can respond to
instructions regarding persona ideas and desired inflections, to
create the proper tone for the application.
Voice Talents versus Voice Actors
Voice Talents are people who speak well, with good resonance
and intonation. Voice Actors are people who are Voice Talents,
but through the use of their voices alone, can also convey
character, humor, sincerity, and meaning. Voice actors can take
your application to the next level, providing a better experience
for your callers.
Voice Actors do not necessarily over-articulate everything. They
stress the most important words and concepts, which often
corresponds to what the system is trying to recognize. Voice Actors will
also know how to record concatenated prompts with consistency. If
feasible, using a Voice Actor will give your speech application polish⎯and
this polish, combined with a well-designed prompt and grammar, will lead to
more satisfied customers.
Prompt Tips
Never disguise the system as being a real person.
Ensure that prompts elicit a predictable response.
Offer the needed information at the right time,
and try not to frontload the application with too
much information.
Never use "Press or Say 1" style prompts.
List options from specific to general⎯so that
users do not choose a general category when a
more specific one will be offered later.
Allow more time for difficult tasks: the pace of
the recorded prompts dictates the pace at which
the callers will respond.
Keep the caller in the transaction by saying
"first…next…finally" or similar transition words, in
the appropriate places.
Insert pauses in your prompts to allow
experienced callers a turn-taking option to
interrupt and move forward in the application.
Provide audio rewards at the completion of
difficult transactions, such as "great, wonderful,
excellent..."
LumenVox's software proved to be
very effective in precisely interpreting
callers’ speech patterns.
Chris Riggenbach CXM Product
Development Manager
48 49

1
2
3
Alternate Pronunciations
One of the most useful speech applications today is the front-end call
router, however, it's also one of the most challenging applications
because the system must recognize names.
Sometimes names are derived from languages other than English, and the
pronunciation reflects rules from the other language. Often these names
contain sounds that are not apparent from the spelling, or the caller
stresses a syllable that differs from the common pronunciation.
Imagine a person looking at the name "Elicia." Is the initial sound a
soft 'AX' as in "about", an 'EH' as in "bed" or is it stressed heavily
with a long 'IY' as in "equal?" Is the third syllable pronounced with
an 's' sound or a 'SH'?" The speech application needs help to
determine this.
If a word or name is not in the dictionary, the Speech Recognition
Engine will try to figure out how that word is pronounced using a set
of phonetic rules, similar to how a person might try sounding out the
new word. Unfortunately, the Speech Engine is not always correct. A
good indicator is that if a person has trouble figuring out how to
pronounce a name, the speech engine will, too.
Steps for Developing a Smooth Call Router:
Figure out who the incoming callers are. Are they strangers, people who know the employees, or
a combination of both? In other words, will the callers know how to pronounce the name correctly,
or will other likely pronunciations need to be added? Do the callers refer to employees by first or
last name only, and are they familiar enough to know people's nicknames?
Find out what the Speech Engine thinks is the correct pronunciation, or whether other
pronunciations are needed. You can use the Phonetic Speller located in the Speech Platform to
see how the Speech Engine determines the pronunciations of the names, without having to run the
Speech Engine itself.
Add the new pronunciations for names into the grammar. Here's an illustration of this process:
The name "Paty" (pronounced "Patty") is not a common spelling and is not in the Speech Engine's
dictionary. When typing it into the Phonetic Speller, the system returns 'P EY DX IY' which sounds
something like "Paydee", instead of the correct 'P AE DX IY'. To add a new pronunciation, the
phonemes will need to be made within a set of curly braces. Adding a colon followed by the true
spelling of the name helps readability, so it's a good idea to include it. The final entry of {P AE DX
IY: Paty} as an alternative pronunciation should help the system's performance, since now callers
who say "Patty" will be more likely to be recognized, instead of the incorrect {P EY DX IY}.
Phonemes
The unit of sound the recognition engine actually recognizes is the phoneme. All phrase formats
are ultimately translated into phonetic spellings for decoding. These phonetic spellings can be
directly entered if surrounded by curly braces.
The phonetic alphabet used by the American English language model is below.
Phoneme Example 1 Phonetic Spelling 1 Example 2 Phonetic Spelling 2
Considering alternate pronunciations and spellings at the outset will help avoid errors
50 and frustration later!
51
Vowels
AA
AE
AH
AO
AW
AX
AXR
AY
EH
ER
EY
IH
IX
IY
OW
OY
UH
UW
Consonants
B
CH
D
DH
DX
F
G
HH
JH
K
L
M
N
NG
P
R
S
SH
T
TH
V
W
Y
Z
ZH
barn
bat
what
more
cow
about
butter
type
check
church
take
little
action
team
loan
hoist
book
flew
web
chair
reed
with
forty
four
peg
halt
cage
coin
late
lemon
night
ring
pay
rest
sit
blush
raft
three
van
swap
yes
arms
Asian
B AA R N
B AE T
W AH T
M AO R
C AW
AX B AW T
B AH DX AXR
T AY P
CH EH K
CH ER CH
T EY K
L IH DX AX L
AE K SH IX N
T IY M
L OW N
H OY S T
B UH K
F L UW
W EH B
CH EY R
R IY D
W IH DH
F AO R DX IY
F AO R
P EH G
HH AO L T
K EY JH
K OY N
L EY T
L EH M AH N
N AY T
R IH NG
P EY
R EH S T
S IH T
B L AH SH
R AE F T
TH R IY
V AE N
S W AA P
Y EH S
AA R M Z
EY ZH AH N
top
crab
cut
auto
house
dial
career
life
mess
bird
hail
rib
women
keep
robe
joy
look
who
bear
statue
dark
other
butter
graph
exam
Jose
Jack
back
really
mail
any
ankle
beep
prior
bass
sure
taped
youth
river
wing
year
blaze
genre
T AA P
K R AE B
K AH T
AO T OW
HH AW S
D AY AX L
K AXR IH R
K AY F
M EH S
B ER D
HH EY L
R IH B
W IH M IX N
K IY P
R OW B
JH OY
L UH K
HH UW
B EH R
S T AE CH UW
D AA R K
AH DH ER
B AH DX AXR
G R AE F
IH G Z AE M
HH OW Z EY
JH AE K
B AE K
R IH L IY
M EY L
EH N IY
AE NG K AH L
B IY P
P R AY ER
B AE S
SH UH R
Y EY P T
Y UW TH
R IH V AXR
W IH NG
Y IY R
B L EY Z
ZH AA N R AH

Practical Guide To Tuning
Untuned speech applications do not survive contact with customers:
whether your company has live speech applications in deployment
today, plans to implement one within the next three to six months, or is
only beginning to consider adding speech applications, you should
consider the importance of tuning. Tuning uses prompts, grammars,
call flow, and caller data to improve the speech application as a whole.
There are three ideas to keep in mind when approaching
the tuning task:
1
Tuning Takes Time.
Even the best of "best-practices" build on assumptions that might not hold
true after deployment⎯once you have callers, you must often readjust or
remove these assumptions to provide the quality experience callers expect.
To give an idea of how much time tuning can take, the speech industry
estimates that 40-50% of total development and deployment time should
be spent on the tuning process. Putting emphasis on tuning will help your
application run more smoothly, keeping callers happy.
2
Adapt the System to the Caller.
Start with Small Changes.
In general, you will not be able to make
users do anything in any particular way.
You can, and should, give as much
guidance for callers as possible, but
ultimately the caller dictates the
conversation. The trick is to provide
good cues and guidelines, so callers
choose the pathway you designed for
the application. Remember that if the
system fails to meet the caller's needs,
it's not the caller who has failed; it's
the speech application.
3
It is all too easy to get caught up in the
moment, expending hours of effort on a
seemingly enormous problem⎯for something
that really only affects a few out of several
hundred callers. Identify the issues that are the
easiest to resolve and provide the biggest
benefit. Making small changes to improve the
experience for most callers is preferable to
costly changes that only benefit a few.
Instead, try this process when tuning an application:
Familiarize Yourself with the Caller's Experiences.
1
2
Do this by listening to the calls, from start to finish. Compare
the ASR results with respect to the audio prompts and the
caller's speech. Transcribe the audio, so you can analyze
the accuracy and performance.
Use your ASR platform's reporting and analytical tools
to maximize your information. You can even use
LumenVox Speech Tuner on Nuance's 8.5 or
ScanSoft's OSR.
Above all, identify the key issues and prioritize them.
Solve the easiest dilemmas first, like typical grammar
problems. Then, move to prompt and dialogue
changes, and finally proceed to acoustic model
training and adaptations.
Test Changes Rigorously.
When you make a change, you must test it. You did
the transcripts, and so you have the grammar and
audio data: as much as possible, test under 'real'
conditions. Give yourself the assurance that any
change will help, and then test to find what solution
works best.
What you shouldn't do when tuning a speech application:
Don’t Make Changes
Based on One Instance.
This should be fairly obvious, but it still
happens. Making changes based on a single
instance usually results in fixing a problem that
doesn't really exist. There are numerous
'one-off' errors in speech recognition, many of
which are associated with noise, or transient
effects that won't be generally reproducible.
Real issues will arise multiple times, in multiple
places, with plenty of evidence to help you
decide how to solve them.
Don’t Make Changes on
Unanalyzed Reports.
Treat the report with respect: analyze the call,
compare it with other calls, see what really
happened⎯often, the system worked as
designed, but the design was flawed. Research
the problem carefully so that you avoid
unnecessary (and costly) changes.
The LumenVox support team is
always available and always willing to go
the extra mile to provide us with
excellent support.
Derek Henry
CEO of 1-800-US-LOTTO
Corporation
52 53

Tuning Grammars
There are many places to make effective changes, but generally, we have
found grammars to be the easiest and most effective place to start. In
this segment, we will look at how to detect errors and modify grammars
to optimize performance.
Grammar Terms
In-Grammar (IG) and Out-of-Grammar (OOG) are labels that look at whether the ASR matches
a path in the grammar with what the caller actually said. If it can, then the spoken words are
In-Grammar, if not, the spoken words are considered Out-of-Grammar.
Confidence scores indicate the ASR's certainty about the answer it returns.
Confirmations are dialog techniques to help the speech application avoid making a mistake, in
cases where the results are ambiguous.
Substitutions are a particular kind of error the ASR makes; this occurs when the result from the
ASR does not match the words the caller said.
LumenVox’s development and support
staff has been very responsive to our
requirements and issues.
Out-Of-Grammar Indicators
There are a number of ways to determine whether or not the error is due to an Out-of-Grammar
issue. The easiest, most efficient way to discover this is to use the tools provided by the platform
or ASR provider. Pre-configured reports will usually highlight these issues up front.
If Out-of-Grammar is a big problem, you will likely receive many customer complaints and low
completion and usage rates. Call lots will also have many "No Matches" or empty results. Finally,
when you do obtain results, the confidence score will be significantly lower than the rest of the
application.
As the call logs are transcribed, look for low accuracies. ASRs will usually get anything that is
In-Grammar, so if there is still low accuracy, look for a bunch of Out-of-Grammar speech.
The other good indicator is a high Out-of-Grammar rate. Generally, this is a direct indication that
callers are not saying things your grammar understands.
There are a few reasons for Out-of-Grammar problems, most of which are easy to resolve.
One common reason is that the grammar designer simply forgot to add items, but callers are asking
for them. Leaving "next" out of a navigation grammar, or forgetting to add a product name is not
uncommon.
Another common error is forgetting common synonyms, for example, 'copier', but not 'Xerox', or
different dialectal versions such as 'soda' in the west, and 'pop' in the South. Usually, you can just
add the missing items.
In-Grammar Indicators
Typical In-Grammar issues will often be oriented towards improving the confidence scores of
recognition, but you will also confront misrecognitions.
Doug Behl
President of Malibu Software
What kinds of issues arise frequently?
Regularly confused phrases are fairly common, and often result because two or more phrases sound
quite similar. Another common issue is the result of bad pronunciations in the grammar. ASRs
provide a methodology for arriving at pronunciations for words that aren't in the dictionary, but the
automatic pronunciations are not always the best.
So, how can we handle this?
For regularly confused phrases, differentiate them by choosing alternate ways to describe the
words. For bad pronunciations, you must add the pronunciations to the dictionary or grammar.
There are tools for helping with this task, although they are nearly always ASR-specific.
54 55

Failures and Fixes: Common
Prompt Tuning Issues
Effective prompt design takes time and practice⎯some errors will not
present themselves until the prompts are tested. There are, however,
some common issues that arise when tuning prompts, which should
help streamline your prompt design and tuning process.
Here's what you should do when callers…
…Give Long, Perplexing Answers
Callers continuously give full sentence answers instead of short, to the point answers. Typically,
this occurs because the prompt asks a very open-ended question, such as "How may I help
you?" Avoid these open-ended prompts; callers usually do not know what responses are
appropriate at particular points in the call. The only real solution when this error occurs is to
redesign the prompt to be more specific, or redesign the interaction to focus caller responses
on specific tasks.
…Answer with Out-of-Grammar (OOG) Phrases
Callers regularly use a particular phrase that is not in the grammar. Prompts are designed to
elicit particular pieces of information from the caller. Because of this, the prompts usually try to
lead the caller to using the correct words or phrases to minimize recognition errors and caller
confusion. When callers regularly use Out-of-Grammar phrases, it's usually because the prompt
leads them to the wrong phrases. Two choices are available: include the Out-of-Grammar
phrases, or revise the prompt to more obviously reflect available options.
…Answer Randomly, or 'Hunt' for the Right Phrase
Unclear, incomplete prompts force the caller to search unnecessarily for the correct response.
Adding clarifying information will generally fix this problem.
Transcription and Training
Humans are exceptionally good at speech processing.
We handle a variety of accents, speaking styles, pitch
differences, noise, and more, with a high degree of
accuracy. No speakers say the same word exactly the
same way. Needless to say, this represents a considerable
challenge for automated speech recognizers. Accurate
transcription and tuning is essential.
Every speech recognizer uses a statistical model of speech in
order to perform good recognition. These models are built
during training, where speech audio and text transcriptions are
combined with algorithms that 'learn' how speech sounds. The
models attempt to determine what 'average' speakers sound
like when they speak particular words, and apply that
knowledge to new incoming speech to determine what words
were spoken.
Words and speaking styles are different for every
application domain (i.e., the vocabulary for a travel
system is quite different from that of a financial
application). Speech applications benefit from
acoustic models specially trained with data
from their specific domains.
Transcribing audio data must be exact, word
for word, and include noise tags so the
system can learn the differences between
noise and speech. To do this, the data must
include as many speakers (both male and
female) as possible, so that the new acoustic
models accurately reflect the average speaker, and
not just one or two particular speakers.
With a larger volume of transcribed audio data,
new models will perform better. New acoustic
models will likely require a new round of
tuning, particularly with respect to
confirmation thresholds.
…Answer 'Yes' or 'No,' Instead of Expected Content
If callers respond 'yes' or 'no' when the prompt requests a content word, a poorly designed
prompt is responsible. For example, the prompt might ask a question like "Do you want..."
or "Would you like..." and pauses after the first choice. The pause can be long enough to
make the caller believe the desired answer is yes or no, rather than a list of choices.
Similarly, multi-item lists may pause too long at later points, as in, "Would you like
pizza, soft drinks, or side items?" Generally, we recommend that
you reserve "Do" and "Would" lead-offs for questions that require yes or no answers.
LumenVox support is very fast and
accessible.
Kelly Lumpkin CEO of Alternate Access
56 57

Standards and Systems Supported
Industry Standards
VXML/VoiceXML (Voice eXtensible Markup
Language)
SALT (Speech Application Language Tags)
MRCP (Media Resource Control Protocol)
SRGS (Speech Recognition Grammar
Specification)
SAPI 5 (Microsoft SAPI TTS)
Operating Systems Supported
Linux
Windows NT, 2000, XP, 2003
Telephony Standards
PSTN (Public Switched Telephone
Network)
SIP (Session Initiation Protocol) -
Signaling protocol for Internet
conferencing, telephony, and instant
messaging
VoIP (Voice Over Internet Protocol) -
Protocol to send audio and data
information in digital
DM3 - Intel® series of boards
HMP (Host Media Processing) -
Software that performs media
processing tasks based on Intel®
architecture
Global Call - Protocol for handling the
call control interface for Intel® cards
LumenVox allows us to have control
over the development of call flow,
grammars and tuning, while keeping the
costs at a respectable level.
Brian Lauzon
President of TelASK
University Research
Carnegie Mellon University
Performs research for all aspects of speech
recognition, including signal processing,
acoustic model training, language model
training, decoding, spoken language parsing
and interface building.
University of Colorado, Boulder
Focused on research and education in areas
of human communication technology.
Oregon University
Center for Spoken Language-Synthesis,
Recognition and Enhancement.
Stanford University
The Center for the Study of Language and
Information (CSLI) is an Independent
Research Center founded in 1983 by
researchers from Stanford University,
SRI International, and Xerox PARC.
UC Berkeley
The major application area researched in
the Speech Group at ICSI is speech
recognition, although some of this work has
led to basic research in auditory processing.
M.I.T.
Computer Science and Artificial Intelligence
Laboratory that focuses on Spoken
Language Systems.
Other Speech Groups
Carnegie Mellon Speech Group
Develops user interfaces that improve
human-computer and human-human
communication.
IEICE
The Institute of Electronics, Information and
Communication Engineers aims at the
investigation and exchange of knowledge on
the science and technology of electronics,
information, and communications.
Speech Publications
Speech Technology Magazine
Leading source of information promoting the
speech technology solutions that are
changing communications, and reports the
technology needs of organizations
worldwide.
Telephone Strategy News
Newsletter that includes full coverage of the
impact of the Voice User Interface on
telephony.
ASRNews
Monthly newsletter, which tracks the latest
development in the speech recognition and
text-to-speech marketplace.
Call Center And Telephony
Technology Marketing
Corporation
TMC publishes industry-leading print
magazines including Internet Telephony,
Customer Inter@ction Solutions, and
Communications Solutions.
CMP Media
Leading multimedia company that prints
numerous magazines, including Call
Center and Communications Convergence.
Business Communications
Review Magazine
Magazine for the enterprise network
manager and other communications
professionals.

Standards and Systems Supported
Industry Standards
VXML/VoiceXML (Voice eXtensible Markup
Language)
SALT (Speech Application Language Tags)
MRCP (Media Resource Control Protocol)
SRGS (Speech Recognition Grammar
Specification)
SAPI 5 (Microsoft SAPI TTS)
Operating Systems Supported
Linux
Windows NT, 2000, XP, 2003
Telephony Standards
PSTN (Public Switched Telephone
Network)
SIP (Session Initiation Protocol) -
Signaling protocol for Internet
conferencing, telephony, and instant
messaging
VoIP (Voice Over Internet Protocol) -
Protocol to send audio and data
information in digital
DM3 - Intel® series of boards
HMP (Host Media Processing) -
Software that performs media
processing tasks based on Intel®
architecture
Global Call - Protocol for handling the
call control interface for Intel® cards
LumenVox allows us to have control
over the development of call flow,
grammars and tuning, while keeping the
costs at a respectable level.
Brian Lauzon
President of TelASK
University Research
Carnegie Mellon University
Performs research for all aspects of speech
recognition, including signal processing,
acoustic model training, language model
training, decoding, spoken language parsing
and interface building.
University of Colorado, Boulder
Focused on research and education in areas
of human communication technology.
Oregon University
Center for Spoken Language-Synthesis,
Recognition and Enhancement.
Stanford University
The Center for the Study of Language and
Information (CSLI) is an Independent
Research Center founded in 1983 by
researchers from Stanford University,
SRI International, and Xerox PARC.
UC Berkeley
The major application area researched in
the Speech Group at ICSI is speech
recognition, although some of this work has
led to basic research in auditory processing.
M.I.T.
Computer Science and Artificial Intelligence
Laboratory that focuses on Spoken
Language Systems.
Other Speech Groups
Carnegie Mellon Speech Group
Develops user interfaces that improve
human-computer and human-human
communication.
IEICE
The Institute of Electronics, Information and
Communication Engineers aims at the
investigation and exchange of knowledge on
the science and technology of electronics,
information, and communications.
Speech Publications
Speech Technology Magazine
Leading source of information promoting the
speech technology solutions that are
changing communications, and reports the
technology needs of organizations
worldwide.
Telephone Strategy News
Newsletter that includes full coverage of the
impact of the Voice User Interface on
telephony.
ASRNews
Monthly newsletter, which tracks the latest
development in the speech recognition and
text-to-speech marketplace.
Call Center And Telephony
Technology Marketing
Corporation
TMC publishes industry-leading print
magazines including Internet Telephony,
Customer Inter@ction Solutions, and
Communications Solutions.
CMP Media
Leading multimedia company that prints
numerous magazines, including Call
Center and Communications Convergence.
Business Communications
Review Magazine
Magazine for the enterprise network
manager and other communications
professionals.