Data-Driven Enhancement of State Mapping-Based Cross-Lingual Speaker Adaptation

The thesis work was motivated by the goal of developing personalized speech-to-speech translation and focused on one of its key component techniques – cross-lingual speaker adaptation for text-to-speech synthesis. A personalized speech-to-speech translator enables a person’s spoken input to be translated into spoken output in another language while maintaining his/her voice identity. Before addressing any technical issues, work in this thesis set out to understand human perception of speaker identity. Listening tests were conducted in order to determine whether people could differentiate between speakers when they spoke different languages. The results demonstrated that differentiating between speakers across languages was an achievable task. However, it was difficult for listeners to differentiate between speakers across both languages and speech types (original recordings versus synthesized samples). The underlying challenge in cross-lingual speaker adaptation is how to apply speaker adaptation techniques when the language of adaptation data is different from that of synthesis models. The main body of the thesis work was devoted to the analysis and improvement of HMM state mapping-based cross-lingual speaker adaptation. Firstly, the effect of unsupervised cross-lingual adaptation was investigated, as it relates to the application scenario of personalized speech-to-speech translation. The comparison of paired supervised and unsupervised systems shows that the performance of unsupervised cross-lingual speaker adaptation is comparable to that of the supervised fashion, even if the average phoneme error rate of the unsupervised systems is around 75%. Then the effect of the language mismatch between synthesis models and adaptation data was investigated. The mismatch is found to transfer undesirable language information from adaptation data to synthesis models, thereby limiting the effectiveness of generating multiple regression class-specific transforms, using larger quantities of adaptation data and estimating adaptation transforms iteratively. Thirdly, in order to tackle the problems caused by the language mismatch, a data-driven adaptation framework using phonological knowledge is proposed. Its basic idea is to group HMM states according to phonological knowledge in a data-driven manner and then to map each state to a phonologically consistent counterpart in a different language. This framework is also applied to regression class tree construction for transform estimation. It is found that the proposed framework alleviates the negative effect of the language mismatch and gives consistent improvement compared to previous state-of-the-art approaches. Finally, a two-layer hierarchical transformation framework is developed, where one layer captures speaker characteristics and the other compensates for the language mismatch. The most appropriate means to construct the hierarchical arrangement of transforms was investigated in an initial study. While early results show some promise, further in-depth investigation is needed to confirm the validity of this hierarchy.

Unified Framework Of Feature Based Adaptation For Statistical Speech Synthesis And Recognition

Lakshmi Babu Saheer

The advent of statistical parametric speech synthesis has paved new ways to a unified framework for hidden Markov model (HMM) based text to speech synthesis (TTS) and automatic speech recognition (ASR). The techniques and advancements made in the field of ASR can now be adopted in the domain of synthesis. Speaker adaptation is a well-advanced topic in the area of ASR, where the adaptation data from a target speaker is used to transform the canonical model parameters to represent a speaker specific model. Feature adaptation techniques like vocal tract length normalization (VTLN) perform the same task by transforming the features; this can be shown to be equivalent to model transformation. The main advantage of VTLN is that it can demonstrate noticeable improvements in performance with very little adaptation data and can be classified as a rapid adaptation technique. VTLN is a widely used technique in ASR, and can be used in TTS to improve the rapid adaptation performance. In TTS, the task is to synthesize speech that sounds like a particular target speaker. %Most of this work was performed in the context of Effective multilingual interactions in mobile environments (EMIME), personalized speech-to-speech translation project. Using VTLN for TTS is found to make the output synthesized speech sound quite similar to the target speaker from his very first utterance. An all-pass filter based bilinear transform was implemented for the mel-generalized cepstral (MGCEP) features of the HMM-based speech synthesis system (HTS). The initial implementation was using a grid search approach that selects the best warping factor for the speech spectrum from a grid of available values using the maximum likelihood criterion. VTLN was shown to give performance improvements in the rapid adaptation framework where the number of adaptation sentences from the target speaker was limited. But, this technique involves huge time and space complexities and the rapid adaptation demands for an efficient implementation of the VTLN technique. To this end, an efficient expectation maximization (EM) based VTLN approach was implemented for HTS using Brent's search. Unlike the ASR features, MGCEP does not use a filter bank (in order to facilitate the speech reconstruction) and this provides equivalence to the model transformation for the EM implementation. This facilitates the estimation of warping factors to be embedded in the HMM training using the same sufficient statistics as in constrained maximum likelihood linear regression (CMLLR). This work addresses a lot of challenges faced in the process of adopting VTLN for synthesis due to the higher dimensionality of the cepstral features used in the TTS models. The main idea was to unify the theory and practise in the implementation of VTLN for both ASR and TTS. Several techniques have been proposed in this thesis, in order to find the best feasible warping factor estimation procedure. Estimation of the warping factor using the lower order cepstral features representing the spectral envelope is demonstrated to be the best approach. Different evaluations on standard databases are performed in this work to illustrate the performance improvements and perceptual challenges involved in the VTLN adaptation. VTLN has only a single parameter to represent the speaker characteristics and hence, has the limitation of not scaling to the performance of other linear transform based adaptation methods with the availability of large amounts of adaptation data. Several techniques are demonstrated in this work to combine the model based adaptation like constrained structural maximum a posteriori linear regression (CSMAPLR) with VTLN, one such technique being using VTLN as the prior transform at the root node of the tree structure of the CSMAPLR system. Thus, along with rapid adaptation, the performance scales with the availability of more adaptation data. These techniques although developed for TTS, can also be effectively used in ASR. It was also shown to give improvements in ASR especially for scenarios like noisy speech conditions. Other improvements to rapid adaptation including a bias term for VTLN, multiple transform based VTLN using regression classes and VTLN prior for non-structural MAPLR adaptation are also proposed. These techniques also demonstrated both ASR and TTS performance improvements. Also, a few special scenarios, specifically cross-lingual speech, cross-gender speech, child speech and noisy speech evaluations are presented where the rapid adaptation methods presented in this work was shown to be highly beneficial. Most of these methods will be published as extensions to the open-source HTS toolkit.

Ecole Polytechnique Federale de Lausanne (EPFL)2012

Unified Framework of Feature Based Adaptation for Statistical Speech Synthesis and Recognition

Lakshmi Babu Saheer

The advent of statistical parametric speech synthesis has paved new ways to a unified framework for hidden Markov model (HMM) based text to speech synthesis (TTS) and automatic speech recognition (ASR). The techniques and advancements made in the field of ASR can now be adopted in the domain of synthesis. Speaker adaptation is a well-advanced topic in the area of ASR, where the adaptation data from a target speaker is used to transform the canonical model parameters to represent a speaker specific model. Feature adaptation techniques like vocal tract length normalization (VTLN) perform the same task by transforming the features ; this can be shown to be equivalent to model transformation. The main advantage of VTLN is that it can demonstrate noticeable improvements in performance with very little adaptation data and can be classified as a rapid adaptation technique. VTLN is a widely used technique in ASR, and can be used in TTS to improve the rapid adaptation performance. In TTS, the task is to synthesize speech that sounds like a particular target speaker. Using VTLN for TTS is found to make the output synthesized speech sound quite similar to the target speaker from his very first utterance. An all-pass filter based bilinear transform was implemented for the mel-generalized cepstral (MGCEP) features of the HMM-based speech synthesis system (HTS). The initial implementation was using a grid search approach that selects the best warping factor for the speech spectrum from a grid of available values using the maximum likelihood criterion. VTLN was shown to give performance improvements in the rapid adaptation framework where the number of adaptation sentences from the target speaker was limited. But, this technique involves huge time and space complexities and the rapid adaptation demands for an efficient implementation of the VTLN technique. To this end, an efficient expectation maximization (EM) based VTLN approach was implemented for HTS using Brent’s search. Unlike the ASR features, MGCEP does not use a filter bank (in order to facilitate the speech reconstruction) and this provides equivalence to the model transformation for the EM implementation. This facilitates the estimation of warping factors to be embedded in the HMM training using the same sufficient statistics as in constrained maximum likelihood linear regression (CMLLR). This work addresses a lot of challenges faced in the process of adopting VTLN for synthesis due to the higher dimensionality of the cepstral features used in the TTS models. The main idea was to unify the theory and practise in the implementation of VTLN for both ASR and TTS. Several techniques have been proposed in this thesis, in order to find the best feasible warping factor estimation procedure. Estimation of the warping factor using the lower order cepstral features representing the spectral envelope is demonstrated to be the best approach. Different evaluations on standard databases are performed in this work to illustrate the performance improvements and perceptual challenges involved in the VTLN adaptation. VTLN has only a single parameter to represent the speaker characteristics and hence, has the limitation of not scaling to the performance of other linear transform based adaptation methods with the availability of large amounts of adaptation data. Several techniques are demonstrated in this work to combine the model based adaptation like constrained structural maximum a posteriori linear regression (CSMAPLR) with VTLN, one such technique being using VTLN as the prior transform at the root node of the tree structure of the CSMAPLR system. Thus, along with rapid adaptation, the performance scales with the availability of more adaptation data. These techniques although developed for TTS, can also be effectively used in ASR. It was also shown to give improvements in ASR especially for scenarios like noisy speech conditions. Other improvements to rapid adaptation including a bias term for VTLN, multiple transform based VTLN using regression classes and VTLN prior for non-structural MAPLR adaptation are also proposed. These techniques also demonstrated both ASR and TTS performance improvements. Also, a few special scenarios, specifically cross-lingual speech, cross-gender speech, child speech and noisy speech evaluations are presented where the rapid adaptation methods presented in this work was shown to be highly beneficial. Most of these methods will be published as extensions to the open-source HTS toolkit.

EPFL2013

Robust Audio Segmentation

Jitendra Ajmera, Hervé Bourlard

Audio segmentation, in general, is the task of segmenting a continuous audio stream in terms of acoustically homogenous regions, where the rule of homogeneity depends on the task. This thesis aims at developing and investigating efficient, robust and unsupervised techniques for three important tasks related to audio segmentation, namely speech/music segmentation, speaker change detection and speaker clustering. The speech/music segmentation technique proposed in this thesis is based on the functioning of a HMM/ANN hybrid ASR system where an MLP estimates the posterior probabilities of different phonemes. These probabilities exhibit a particular pattern when the input is a speech signal. This pattern is captured in the form of feature vectors, which are then integrated in a HMM framework. The technique thus segments the audio data in terms of {\it recognizable} and {\it non-recognizable} segments. The efficiency of the proposed technique is demonstrated by a number of experiments conducted on broadcast news data exhibiting real-life scenarios (different speech and music styles, overlapping speech and music, non-speech sounds other than music, etc.). A novel distance metric is proposed in this thesis for the purpose of finding speaker segment boundaries (speaker change detection). The proposed metric can be seen as special case of Log Likelihood Ratio (LLR) or Bayesian Information Criterion (BIC), where the number of parameters in the two models (or hypotheses) is forced to be equal. However, the advantage of the proposed metric over LLR, BIC and other metric based approaches is that it achieves comparable performance without requiring an adjustable threshold/penalty term, hence also eliminating the need for a development dataset. Speaker clustering is the task of unsupervised classification of the audio data in terms of speakers. For this purpose, a novel HMM based agglomerative clustering algorithm is proposed where, starting from a large number of clusters, {\it closest} clusters are merged in an iterative process. A novel merging criterion is proposed for this purpose, which does not require an adjustable threshold value and hence the stopping criterion is also automatically met when there are no more clusters left for merging. The efficiency of the proposed algorithm is demonstrated with various experiments on broadcast news data and it is shown that the proposed criterion outperforms the use of LLR, when LLR is used with an optimal threshold value. These tasks obviously play an important role in the pre-processing stages of ASR. For example, correctly identifying {\it non-recognizable} segments in the audio stream and excluding them from recognition saves computation time in ASR and results in more meaningful transcriptions. Moreover, researchers have clearly shown the positive impact of further clustering of identified speech segments in terms of speakers (speaker clustering) on the transcription accuracy. However, we note that this processing has various other interesting and practical applications. For example, this provides characteristic information about the data (metadata), which is useful for the indexing of audio documents. One such application is investigated in this thesis which extracts this metadata and combines it with the ASR output, resulting in Rich Transcription (RT) which is much easier to understand for an end-user. In a further application, speaker clustering was combined with precise location information available in scenarios like smart meeting rooms to segment the meeting recordings jointly in terms of speakers and their locations in a meeting room. This is useful for automatic meeting summarization as it enables answering of questions like ``who is speaking and where''. This could be used to access, for example, a specific presentation made by a particular speaker or all the speech segments belonging to a particular speaker.

IDIAP2004

Unified Framework of Feature Based Adaptation for Statistical Speech Synthesis and Recognition

Lakshmi Babu Saheer

The advent of statistical parametric speech synthesis has paved new ways to a unified framework for hidden Markov model (HMM) based text to speech synthesis (TTS) and automatic speech recognition (ASR). The techniques and advancements made in the field of ASR can now be adopted in the domain of synthesis. Speaker adaptation is a well-advanced topic in the area of ASR, where the adaptation data from a target speaker is used to transform the canonical model parameters to represent a speaker specific model. Feature adaptation techniques like vocal tract length normalization (VTLN) perform the same task by transforming the features ; this can be shown to be equivalent to model transformation. The main advantage of VTLN is that it can demonstrate noticeable improvements in performance with very little adaptation data and can be classified as a rapid adaptation technique. VTLN is a widely used technique in ASR, and can be used in TTS to improve the rapid adaptation performance. In TTS, the task is to synthesize speech that sounds like a particular target speaker. Using VTLN for TTS is found to make the output synthesized speech sound quite similar to the target speaker from his very first utterance. An all-pass filter based bilinear transform was implemented for the mel-generalized cepstral (MGCEP) features of the HMM-based speech synthesis system (HTS). The initial implementation was using a grid search approach that selects the best warping factor for the speech spectrum from a grid of available values using the maximum likelihood criterion. VTLN was shown to give performance improvements in the rapid adaptation framework where the number of adaptation sentences from the target speaker was limited. But, this technique involves huge time and space complexities and the rapid adaptation demands for an efficient implementation of the VTLN technique. To this end, an efficient expectation maximization (EM) based VTLN approach was implemented for HTS using Brent’s search. Unlike the ASR features, MGCEP does not use a filter bank (in order to facilitate the speech reconstruction) and this provides equivalence to the model transformation for the EM implementation. This facilitates the estimation of warping factors to be embedded in the HMM training using the same sufficient statistics as in constrained maximum likelihood linear regression (CMLLR). This work addresses a lot of challenges faced in the process of adopting VTLN for synthesis due to the higher dimensionality of the cepstral features used in the TTS models. The main idea was to unify the theory and practise in the implementation of VTLN for both ASR and TTS. Several techniques have been proposed in this thesis, in order to find the best feasible warping factor estimation procedure. Estimation of the warping factor using the lower order cepstral features representing the spectral envelope is demonstrated to be the best approach. Different evaluations on standard databases are performed in this work to illustrate the performance improvements and perceptual challenges involved in the VTLN adaptation. VTLN has only a single parameter to represent the speaker characteristics and hence, has the limitation of not scaling to the performance of other linear transform based adaptation methods with the availability of large amounts of adaptation data. Several techniques are demonstrated in this work to combine the model based adaptation like constrained structural maximum a posteriori linear regression (CSMAPLR) with VTLN, one such technique being using VTLN as the prior transform at the root node of the tree structure of the CSMAPLR system. Thus, along with rapid adaptation, the performance scales with the availability of more adaptation data. These techniques although developed for TTS, can also be effectively used in ASR. It was also shown to give improvements in ASR especially for scenarios like noisy speech conditions. Other improvements to rapid adaptation including a bias term for VTLN, multiple transform based VTLN using regression classes and VTLN prior for non-structural MAPLR adaptation are also proposed. These techniques also demonstrated both ASR and TTS performance improvements. Also, a few special scenarios, specifically cross-lingual speech, cross-gender speech, child speech and noisy speech evaluations are presented where the rapid adaptation methods presented in this work was shown to be highly beneficial. Most of these methods will be published as extensions to the open-source HTS toolkit.

EPFL2013

Unified Framework Of Feature Based Adaptation For Statistical Speech Synthesis And Recognition

Lakshmi Babu Saheer

Ecole Polytechnique Federale de Lausanne (EPFL)2012

Robust Audio Segmentation

Jitendra Ajmera, Hervé Bourlard

IDIAP2004