haga click aquí - Amprae

EasyAlign Spanish: an (semi-)automatic segmentation
tool under Praat
Goldman, Jean-Philippe; Schwab, Sandra
Université de Genève
The purpose of phonetic alignment is to determine the time position of phone
boundaries in a speech corpus on the basis of the audio recording and its
orthographic transcription. Aligned corpora are widely used in various speech
applications like automatic speech recognition, speech synthesis, as well as
prosodic and phonetic research. Although manual segmentation constitutes the
more accurate method, it requires a large amount of time for the human labeller.
Thus, various automatic methods are now used as they are not only much
quicker, but their results are also reproducible and consistent throughout a large
corpus.
EasyAlign has been developed within Praat in order to provide an ergonomic
automatic segmentation tool, easy to use for computer science non-specialists.
It is freely distributed as a self-installable plug-in and it is available for French,
English, Brazilian Portuguese and recently for Castilian Spanish. EasyAlign
consists in a group of tools which successively perform three automatic steps
with some minor manual verifications and adjustments to ensure better quality:
utterance segmentation, grapheme-to-phoneme conversion and phonetic
segmentation. From the orthographic transcription, the utterance segmentation
process -which is language-independent- generates a TextGrid with a unique
tier, in which each interval encloses one utterance. Then, the graphemephoneme
conversion step creates a second tier with the phonetic transcription
of the utterances. For this language-specific process, we used for Castilian
Spanish the phonetizer SAGA (Moreno & Mariño 1998) which provides a
detailed phonetic transcription in SAMPA. Finally, in the phonetic segmentation
step, the Viterbi-based HVite tool (within HTK) is called to align each utterance
to its phonetic sequence. For Castilian Spanish, the acoustic models were
trained on the basis of about 360 minutes of unaligned multi-speaker speech for
which a verified phonetic transcription was provided. The phonetic
segmentation process simultaneously generates a phone tier and a word tier.
Additionally, a syllable tier is created on the basis of sonority-based rules for
syllable segmentation.
EasyAlign performances have been evaluated on the basis of a corpus of 12
minutes (one minute of 12 speakers: 6 "internal" speakers from the training
corpus and 6 new "external" speakers), which was manually annotated by
phonetic experts and compared to the automatic alignement. Evaluation was
performed according to three approaches. Firstly, a boundary-based evaluation
showed that 60% of the differences between automatic and manual boundaries
lie within 10ms (and 86% within 20ms). Little difference was observed when
comparing internal and external speakers, which shows a good generalization
of the training. Secondly, a duration-based evaluation revealed that the
difference of automatic/manual phone durations, which has a standard deviation
of 20ms, is similar in internal and external speakers. Finally, a segment-based
evaluation showed that the median of the so-called "overlapping-rate" -a
speech-rate independent measure- reaches 0.74, with little difference between
internal (0.75) and external (0.72) speakers.
65