Best Practices for Speech Corpora in Linguistic Research Workshop ...

select m.label as phoneme, k.label as canonic, o.label as
word, count(f.f1), avg(f.f1)::int as f1, avg(f.f2)::int as f2
from session ses
join signalfile sig on ses.session = substring(sig.filename, 3, 9)
join segment m on sig.id = m.signal_id and m.tier = ’MAU:’
join segment o on m.signal_id = o.signal_id and
m.ref_seg = o.ref_seg and o.tier = ’ORT:’
join segment k on k.signal_id = o.signal_id and
k.ref_seg = m.ref_seg and k.tier = ’KAN:’
join formant f on m.signal_id = f.signal_id and
f.time between (m.begin_seg + (m.dur_seg * 0.2)) and
(m.begin_seg + (m.dur_seg * 0.8))
join speaker spk on ses.session = spk.speaker_code and
ses.project = spk.project and ses.project = ’VOYS’
where m.label = ’E’
group by m.label, k.label, o.label
order by m.label, k.label, o.label;
Figure 3: Sample SQL query to retrieve phoneme segments,
their count and the average f1 and f2 values, grouped
by phoneme segment label, and word
select phoneme, avg(f1)::int, avg(f2)::int
from voys_data
where phoneme = ’E’
group by phoneme
Figure 4: Using a view (a predefined virtual table) to express
the same query as in Figure 3
administrator). The researcher can now access the database
directly from the spreadsheet or statistics software he or she
is using, perform the statistical analyses and display the results
in text format or diagrams (Figure 5).
1000 800 600 400 200
1000 800 600 400 200
i
e
I
all cities (F)
}
E
a
V
o
O
2500 2000 1500 1000
i
e
I
all cities (F)
}
E
a
V
o
O
2500 2000 1500 1000
1000 800 600 400 200
1000 800 600 400 200
i
all cities (M)
I
e
E
}
a
V
o
O
2500 2000 1500 1000
i
all cities (M)
I
e
E
}
a
V
o
O
2500 2000 1500 1000
Figure 5: Formant charts for Scottish English vowels from
the VOYS speech database
Using this workflow and the global corpus data model, similar
analyses or analyses of other corpora may now be performed
with little effort and can thus be used in education.
4. Case study 2: A perception experiment on
regional variants of sounds
Perception experiments are an essential part of phonetic,
linguistic and psychological research. Most such exper-
53
iments are now performed using a computer, and some
speech processing tools, e.g. Praat, directly support performing
perception experiments. However, with standalone
software, performing an experiment requires that the software
is installed on every computer on which the experiment
will run. Web-based online experiments overcome
this limitation and provide access to potentially large
groups of participants (Reips, 2002).
Currently, only a few tools or services exist that allow online
experiments with audio. Examples are WebExp (Keller
et al., 2009) and Percy (Draxler, 2011). WebExp is a flexible
and powerful online experiment software that uses Java
applets for media display, and which stores its result data in
XML files. Percy is based on HTML5 and stores its data in
a relational database on the server.
4.1. Perception experiments in the workflow
In principle, a perception experiment is not very different
from an annotation task – speech material is presented and
the participant has to enter his or her judgment. Hence, the
global corpus data model also covers online perception experiments.
Experiment results are simply considered as yet
another annotation tier. This allows the same data retrieval
mechanism to be used for experiment data, which greatly
simplifies further processing.
4.2. Running the experiment
A recent perception experiment on regional variation uses
the single digit items recorded in the Ph@ttSessionz
project. Participants were asked to judge whether a given
phoneme in a digit has certain properties, e.g. whether the
initial ”s” in the word ”sieben” was voiced or voiceless.
The question was formulated in colloquial terms so that
non-experts could participate in the experiment.
The experiment consists of three steps: the participant registers
and provides some personal and context information.
During the experiment he or she listens to the recorded audio
and enters a judgment, and when all items are done, the
experiment displays a map with the geographic locations
where the audio files were recorded (Figure 6).
4.3. Statistical analysis
With support from the statistics lab of LMU University, the
experiment input data was analysed using mixed-models
provided by the R software. Results show that a) sounds
effectively differ from one region to the other, and b) that
the perception of sound difference depends on the regional
background of the listener.
5. Discussion
A global corpus model is a suitable tool to support the
workflow in phonetic and linguistic research, development,
and education.
However, serious problems remain. The most important are
1. Missing data in the database may lead to broken workflows
or inconsistent or unexpected results.
2. New tools and services may not fit into the global corpus
model.