Segmentation, Diarization and Speech Transcription: Surprise Data Unraveled

M.A.H. Huijbregts

Research output: ThesisPhD Thesis - Research UT, graduation UTAcademic

332 Downloads (Pure)

Abstract

In this thesis, research on large vocabulary continuous speech recognition for unknown audio conditions is presented. For automatic speech recognition systems based on statistical methods, it is important that the conditions of the audio used for training the statistical models match the conditions of the audio to be processed. Any mismatch will decrease the accuracy of the recognition. If it is unpredictable what kind of data can be expected, or in other words if the conditions of the audio to be processed are unknown, it is impossible to tune the models. If the material consists of `surprise data' the output of the system is likely to be poor. In this thesis methods are presented for which no external training data is required for training models. These novel methods have been implemented in a large vocabulary continuous speech recognition system called SHoUT. This system consists of three subsystems: speech/non-speech classification, speaker diarization and automatic speech recognition. The speech/non-speech classification subsystem separates speech from silence and unknown audible non-speech events. The type of non-speech present in audio recordings can vary from paper shuffling in recordings of meetings to sound effects in television shows. Because it is unknown what type of non-speech needs to be detected, it is not possible to train high quality statistical models for each type of non-speech sound. The speech/non-speech classification subsystem, also called the speech activity detection subsystem, does not attempt to classify all audible non-speech in a single run. Instead, first a bootstrap speech/silence classification is obtained using a standard speech activity component. Next, the models for speech, silence and audible non-speech are trained on the target audio using the bootstrap classification. This approach makes it possible to classify speech and non-speech with high accuracy, without the need to know what kinds of sound are present in the audio recording. Once all non-speech is filtered out of the audio, it is the task of the speaker diarization subsystem to determine how many speakers occur in the recording and exactly when they are speaking. The speaker diarization subsystem applies agglomerative clustering to create clusters of speech fragments for each speaker in the recording. First, statistical speaker models are created on random chunks of the recording and by iteratively realigning the data, retraining the models and merging models that represent the same speaker, accurate speaker models are obtained for speaker clustering. This method does not require any statistical models developed on a training set, which makes the diarization subsystem insensitive for variation in audio conditions. Unfortunately, because the algorithm is of complexity $O(n^3)$, this clustering method is slow for long recordings. Two variations of the subsystem are presented that reduce the needed computational effort, so that the subsystem is applicable for long audio recordings as well. The automatic speech recognition subsystem developed for this research, is based on Viterbi decoding on a fixed pronunciation prefix tree. Using the fixed tree, a flexible modular decoder could be developed, but it was not straightforward to apply full language model look-ahead efficiently. In this thesis a novel method is discussed that makes it possible to apply language model look-ahead effectively on the fixed tree. Also, to obtain higher speech recognition accuracy on audio with unknown acoustical conditions, a selection from the numerous known methods that exist for robust automatic speech recognition is applied and evaluated in this thesis. The three individual subsystems as well as the entire system have been successfully evaluated on three international benchmarks. The diarization subsystem has been evaluated at the NIST RT06s benchmark and the speech activity detection subsystem has been tested at RT07s. The entire system was evaluated at N-Best, the first automatic speech recognition benchmark for Dutch.
Original languageUndefined
Awarding Institution
  • University of Twente
Supervisors/Advisors
  • de Jong, Franciska M.G., Supervisor
  • Ordelman, Roeland J.F., Advisor
Thesis sponsors
Award date21 Nov 2008
Place of PublicationEnschede
Publisher
Print ISBNs978-90-365-2712-5
DOIs
Publication statusPublished - 21 Nov 2008

Keywords

  • Automatic Speech Recognition
  • IR-60130
  • speech decoding
  • METIS-255485
  • Speech segmentation
  • SHoUT
  • EWI-15023
  • EC Grant Agreement nr.: FP6/506811
  • Speaker diarization

Cite this

Huijbregts, M.A.H.. / Segmentation, Diarization and Speech Transcription: Surprise Data Unraveled. Enschede : University of Twente, 2008. 172 p.
@phdthesis{7a38d6ab2c664a9f832d909036708f52,
title = "Segmentation, Diarization and Speech Transcription: Surprise Data Unraveled",
abstract = "In this thesis, research on large vocabulary continuous speech recognition for unknown audio conditions is presented. For automatic speech recognition systems based on statistical methods, it is important that the conditions of the audio used for training the statistical models match the conditions of the audio to be processed. Any mismatch will decrease the accuracy of the recognition. If it is unpredictable what kind of data can be expected, or in other words if the conditions of the audio to be processed are unknown, it is impossible to tune the models. If the material consists of `surprise data' the output of the system is likely to be poor. In this thesis methods are presented for which no external training data is required for training models. These novel methods have been implemented in a large vocabulary continuous speech recognition system called SHoUT. This system consists of three subsystems: speech/non-speech classification, speaker diarization and automatic speech recognition. The speech/non-speech classification subsystem separates speech from silence and unknown audible non-speech events. The type of non-speech present in audio recordings can vary from paper shuffling in recordings of meetings to sound effects in television shows. Because it is unknown what type of non-speech needs to be detected, it is not possible to train high quality statistical models for each type of non-speech sound. The speech/non-speech classification subsystem, also called the speech activity detection subsystem, does not attempt to classify all audible non-speech in a single run. Instead, first a bootstrap speech/silence classification is obtained using a standard speech activity component. Next, the models for speech, silence and audible non-speech are trained on the target audio using the bootstrap classification. This approach makes it possible to classify speech and non-speech with high accuracy, without the need to know what kinds of sound are present in the audio recording. Once all non-speech is filtered out of the audio, it is the task of the speaker diarization subsystem to determine how many speakers occur in the recording and exactly when they are speaking. The speaker diarization subsystem applies agglomerative clustering to create clusters of speech fragments for each speaker in the recording. First, statistical speaker models are created on random chunks of the recording and by iteratively realigning the data, retraining the models and merging models that represent the same speaker, accurate speaker models are obtained for speaker clustering. This method does not require any statistical models developed on a training set, which makes the diarization subsystem insensitive for variation in audio conditions. Unfortunately, because the algorithm is of complexity $O(n^3)$, this clustering method is slow for long recordings. Two variations of the subsystem are presented that reduce the needed computational effort, so that the subsystem is applicable for long audio recordings as well. The automatic speech recognition subsystem developed for this research, is based on Viterbi decoding on a fixed pronunciation prefix tree. Using the fixed tree, a flexible modular decoder could be developed, but it was not straightforward to apply full language model look-ahead efficiently. In this thesis a novel method is discussed that makes it possible to apply language model look-ahead effectively on the fixed tree. Also, to obtain higher speech recognition accuracy on audio with unknown acoustical conditions, a selection from the numerous known methods that exist for robust automatic speech recognition is applied and evaluated in this thesis. The three individual subsystems as well as the entire system have been successfully evaluated on three international benchmarks. The diarization subsystem has been evaluated at the NIST RT06s benchmark and the speech activity detection subsystem has been tested at RT07s. The entire system was evaluated at N-Best, the first automatic speech recognition benchmark for Dutch.",
keywords = "Automatic Speech Recognition, IR-60130, speech decoding, METIS-255485, Speech segmentation, SHoUT, EWI-15023, EC Grant Agreement nr.: FP6/506811, Speaker diarization",
author = "M.A.H. Huijbregts",
note = "10.3990/1.9789036527125",
year = "2008",
month = "11",
day = "21",
doi = "10.3990/1.9789036527125",
language = "Undefined",
isbn = "978-90-365-2712-5",
publisher = "University of Twente",
address = "Netherlands",
school = "University of Twente",

}

Segmentation, Diarization and Speech Transcription: Surprise Data Unraveled. / Huijbregts, M.A.H.

Enschede : University of Twente, 2008. 172 p.

Research output: ThesisPhD Thesis - Research UT, graduation UTAcademic

TY - THES

T1 - Segmentation, Diarization and Speech Transcription: Surprise Data Unraveled

AU - Huijbregts, M.A.H.

N1 - 10.3990/1.9789036527125

PY - 2008/11/21

Y1 - 2008/11/21

N2 - In this thesis, research on large vocabulary continuous speech recognition for unknown audio conditions is presented. For automatic speech recognition systems based on statistical methods, it is important that the conditions of the audio used for training the statistical models match the conditions of the audio to be processed. Any mismatch will decrease the accuracy of the recognition. If it is unpredictable what kind of data can be expected, or in other words if the conditions of the audio to be processed are unknown, it is impossible to tune the models. If the material consists of `surprise data' the output of the system is likely to be poor. In this thesis methods are presented for which no external training data is required for training models. These novel methods have been implemented in a large vocabulary continuous speech recognition system called SHoUT. This system consists of three subsystems: speech/non-speech classification, speaker diarization and automatic speech recognition. The speech/non-speech classification subsystem separates speech from silence and unknown audible non-speech events. The type of non-speech present in audio recordings can vary from paper shuffling in recordings of meetings to sound effects in television shows. Because it is unknown what type of non-speech needs to be detected, it is not possible to train high quality statistical models for each type of non-speech sound. The speech/non-speech classification subsystem, also called the speech activity detection subsystem, does not attempt to classify all audible non-speech in a single run. Instead, first a bootstrap speech/silence classification is obtained using a standard speech activity component. Next, the models for speech, silence and audible non-speech are trained on the target audio using the bootstrap classification. This approach makes it possible to classify speech and non-speech with high accuracy, without the need to know what kinds of sound are present in the audio recording. Once all non-speech is filtered out of the audio, it is the task of the speaker diarization subsystem to determine how many speakers occur in the recording and exactly when they are speaking. The speaker diarization subsystem applies agglomerative clustering to create clusters of speech fragments for each speaker in the recording. First, statistical speaker models are created on random chunks of the recording and by iteratively realigning the data, retraining the models and merging models that represent the same speaker, accurate speaker models are obtained for speaker clustering. This method does not require any statistical models developed on a training set, which makes the diarization subsystem insensitive for variation in audio conditions. Unfortunately, because the algorithm is of complexity $O(n^3)$, this clustering method is slow for long recordings. Two variations of the subsystem are presented that reduce the needed computational effort, so that the subsystem is applicable for long audio recordings as well. The automatic speech recognition subsystem developed for this research, is based on Viterbi decoding on a fixed pronunciation prefix tree. Using the fixed tree, a flexible modular decoder could be developed, but it was not straightforward to apply full language model look-ahead efficiently. In this thesis a novel method is discussed that makes it possible to apply language model look-ahead effectively on the fixed tree. Also, to obtain higher speech recognition accuracy on audio with unknown acoustical conditions, a selection from the numerous known methods that exist for robust automatic speech recognition is applied and evaluated in this thesis. The three individual subsystems as well as the entire system have been successfully evaluated on three international benchmarks. The diarization subsystem has been evaluated at the NIST RT06s benchmark and the speech activity detection subsystem has been tested at RT07s. The entire system was evaluated at N-Best, the first automatic speech recognition benchmark for Dutch.

AB - In this thesis, research on large vocabulary continuous speech recognition for unknown audio conditions is presented. For automatic speech recognition systems based on statistical methods, it is important that the conditions of the audio used for training the statistical models match the conditions of the audio to be processed. Any mismatch will decrease the accuracy of the recognition. If it is unpredictable what kind of data can be expected, or in other words if the conditions of the audio to be processed are unknown, it is impossible to tune the models. If the material consists of `surprise data' the output of the system is likely to be poor. In this thesis methods are presented for which no external training data is required for training models. These novel methods have been implemented in a large vocabulary continuous speech recognition system called SHoUT. This system consists of three subsystems: speech/non-speech classification, speaker diarization and automatic speech recognition. The speech/non-speech classification subsystem separates speech from silence and unknown audible non-speech events. The type of non-speech present in audio recordings can vary from paper shuffling in recordings of meetings to sound effects in television shows. Because it is unknown what type of non-speech needs to be detected, it is not possible to train high quality statistical models for each type of non-speech sound. The speech/non-speech classification subsystem, also called the speech activity detection subsystem, does not attempt to classify all audible non-speech in a single run. Instead, first a bootstrap speech/silence classification is obtained using a standard speech activity component. Next, the models for speech, silence and audible non-speech are trained on the target audio using the bootstrap classification. This approach makes it possible to classify speech and non-speech with high accuracy, without the need to know what kinds of sound are present in the audio recording. Once all non-speech is filtered out of the audio, it is the task of the speaker diarization subsystem to determine how many speakers occur in the recording and exactly when they are speaking. The speaker diarization subsystem applies agglomerative clustering to create clusters of speech fragments for each speaker in the recording. First, statistical speaker models are created on random chunks of the recording and by iteratively realigning the data, retraining the models and merging models that represent the same speaker, accurate speaker models are obtained for speaker clustering. This method does not require any statistical models developed on a training set, which makes the diarization subsystem insensitive for variation in audio conditions. Unfortunately, because the algorithm is of complexity $O(n^3)$, this clustering method is slow for long recordings. Two variations of the subsystem are presented that reduce the needed computational effort, so that the subsystem is applicable for long audio recordings as well. The automatic speech recognition subsystem developed for this research, is based on Viterbi decoding on a fixed pronunciation prefix tree. Using the fixed tree, a flexible modular decoder could be developed, but it was not straightforward to apply full language model look-ahead efficiently. In this thesis a novel method is discussed that makes it possible to apply language model look-ahead effectively on the fixed tree. Also, to obtain higher speech recognition accuracy on audio with unknown acoustical conditions, a selection from the numerous known methods that exist for robust automatic speech recognition is applied and evaluated in this thesis. The three individual subsystems as well as the entire system have been successfully evaluated on three international benchmarks. The diarization subsystem has been evaluated at the NIST RT06s benchmark and the speech activity detection subsystem has been tested at RT07s. The entire system was evaluated at N-Best, the first automatic speech recognition benchmark for Dutch.

KW - Automatic Speech Recognition

KW - IR-60130

KW - speech decoding

KW - METIS-255485

KW - Speech segmentation

KW - SHoUT

KW - EWI-15023

KW - EC Grant Agreement nr.: FP6/506811

KW - Speaker diarization

U2 - 10.3990/1.9789036527125

DO - 10.3990/1.9789036527125

M3 - PhD Thesis - Research UT, graduation UT

SN - 978-90-365-2712-5

PB - University of Twente

CY - Enschede

ER -