Speaker Adaptation 2
HomeFull List of Titles
1: ICSLP'98 Proceedings
Keynote SpeechesText-To-Speech Synthesis 1Spoken Language Models and Dialog 1Prosody and Emotion 1Hidden Markov Model Techniques 1Speaker and Language Recognition 1Multimodal Spoken Language Processing 1Isolated Word RecognitionRobust Speech Processing in Adverse Environments 1Spoken Language Models and Dialog 2Articulatory Modelling 1Talking to Infants, Pets and LoversRobust Speech Processing in Adverse Environments 2Spoken Language Models and Dialog 3Speech Coding 1Articulatory Modelling 2Prosody and Emotion 2Neural Networks, Fuzzy and Evolutionary Methods 1Utterance Verification and Word Spotting 1 / Speaker Adaptation 1Text-To-Speech Synthesis 2Spoken Language Models and Dialog 4Human Speech Perception 1Robust Speech Processing in Adverse Environments 3Speech and Hearing Disorders 1Prosody and Emotion 3Spoken Language Understanding Systems 1Signal Processing and Speech Analysis 1Spoken Language Generation and Translation 1Spoken Language Models and Dialog 5Segmentation, Labelling and Speech Corpora 1Multimodal Spoken Language Processing 2Prosody and Emotion 4Neural Networks, Fuzzy and Evolutionary Methods 2Large Vocabulary Continuous Speech Recognition 1Speaker and Language Recognition 2Signal Processing and Speech Analysis 2Prosody and Emotion 5Robust Speech Processing in Adverse Environments 4Segmentation, Labelling and Speech Corpora 2Speech Technology Applications and Human-Machine Interface 1Large Vocabulary Continuous Speech Recognition 2Text-To-Speech Synthesis 3Language Acquisition 1Acoustic Phonetics 1Speaker Adaptation 2Speech Coding 2Hidden Markov Model Techniques 2Multilingual Perception and Recognition 1Large Vocabulary Continuous Speech Recognition 3Articulatory Modelling 3Language Acquisition 2Speaker and Language Recognition 3Text-To-Speech Synthesis 4Spoken Language Understanding Systems 4Human Speech Perception 2Large Vocabulary Continuous Speech Recognition 4Spoken Language Understanding Systems 2Signal Processing and Speech Analysis 3Human Speech Perception 3Speaker Adaptation 3Spoken Language Understanding Systems 3Multimodal Spoken Language Processing 3Acoustic Phonetics 2Large Vocabulary Continuous Speech Recognition 5Speech Coding 3Language Acquisition 3 / Multilingual Perception and Recognition 2Segmentation, Labelling and Speech Corpora 3Text-To-Speech Synthesis 5Spoken Language Generation and Translation 2Human Speech Perception 4Robust Speech Processing in Adverse Environments 5Text-To-Speech Synthesis 6Speech Technology Applications and Human-Machine Interface 2Prosody and Emotion 6Hidden Markov Model Techniques 3Speech and Hearing Disorders 2 / Speech Processing for the Speech and Hearing Impaired 1Human Speech ProductionSegmentation, Labelling and Speech Corpora 4Speaker and Language Recognition 4Speech Technology Applications and Human-Machine Interface 3Utterance Verification and Word Spotting 2Large Vocabulary Continuous Speech Recognition 6Neural Networks, Fuzzy and Evolutionary Methods 3Speech Processing for the Speech-Impaired and Hearing-Impaired 2Prosody and Emotion 72: SST Student Day
SST Student Day - Poster Session 1SST Student Day - Poster Session 2Author Index
A B C D E F G H I
J K L M N O P Q R
S T U V W X Y Z
Multimedia Files
Eigenvoices for Speaker Adaptation
Authors:
Roland Kuhn, Panasonic Technologies Inc., Speech Technology Laboratory (USA)
Patrick Nguyen, Panasonic Technologies Inc., Speech Technology Laboratory (USA)
Jean-Claude Junqua, Panasonic Technologies Inc., Speech Technology Laboratory (USA)
Lloyd Goldwasser, Panasonic Technologies Inc., Speech Technology Laboratory (USA)
Nancy Niedzielski, Panasonic Technologies Inc., Speech Technology Laboratory (USA)
Steven Fincke, Panasonic Technologies Inc., Speech Technology Laboratory (USA)
Ken Field, Panasonic Technologies Inc., Speech Technology Laboratory (USA)
Matteo Contolini, Panasonic Technologies Inc., Speech Technology Laboratory (USA)
Page (NA) Paper number 303
Abstract:
We have devised a new class of fast adaptation techniques for speech recognition. These techniques are based on prior knowledge of speaker variation, obtained by applying Principal Component Analysis (PCA) or a similar technique to T vectors of dimension D derived from T speaker-dependent models. This offline step yields T basis vectors called ``eigenvoices''. We constrain the model for new speaker S to be located in the space spanned by the first K eigenvoices. Speaker adaptation involves estimating the K eigenvoice coefficients for the new speaker; typically, K is very small compared to D. We conducted mean adaptation experiments on the Isolet database. With a large amount of supervised adaptation data, most eigenvoice techniques performed slightly better than MAP or MLLR; with small amounts of supervised adaptation data or for unsupervised adaptation, some eigenvoice techniques performed much better. We believe that the eigenvoice approach would yield rapid adaptation for most speech recognition systems.
SL980303.PDF (From Author)
SL980303.PDF (Rasterized)
TOP
Speaker Clustering Using Direct Maximisation of the MLLR-Adapted Likelihood
Authors:
Sue E. Johnson, Cambridge University (U.K.)
Philip C. Woodland, Cambridge University (U.K.)
Page (NA) Paper number 726
Abstract:
In this paper speaker clustering schemes are investigated in the context of improving unsupervised adaptation for broadcast news transcription. The various techniques are presented within a framework of top-down split-and-merge clustering. Since these schemes are to be used for MLLR-based adaptation, a natural evaluation metric for clustering is the increase in data likelihood from adaptation. Two types of cluster splitting criteria have been used. The first minimises a covariance-based distance measure and for the second we introduce a two-step E-M type procedure to form clusters which directly maximise the likelihood of the adapted data. It is shown that the direct maximisation technique produces a higher data likelihood and also gives a reduction in word error rate.
SL980726.PDF (From Author)
SL980726.PDF (Rasterized)
TOP
Incremental On-Line Speaker Adaptation in Adverse Conditions
Authors:
Olli Viikki, Nokia Research Center, Speech and Audio Systems Laboratory (Finland)
Kari Laurila, Nokia Research Center, Speech and Audio Systems Laboratory (Finland)
Page (NA) Paper number 313
Abstract:
In this paper, we examine the use of speaker adaptation in adverse noise conditions. In particular, we focus on incremental on-line speaker adaptation since it, in addition to its other advantages, enables joint speaker and environment adaptation. First, we show that on-line adaptation is superior to off-line adaptation when realistic changing noise conditions are considered. Next, we show that a conventional left-to-right HMM structure is not well suited for on-line adaptation in variable noise conditions due to unreliable state-frame alignments of noisy utterances. To overcome this problem, we suggest the use of state duration constrained HMMs. Our experimental results indicate that the performance gain due to adaptation is much greater with duration constrained HMMs than obtained with conventional left-to-right HMMs. In addition to the appropriate model structure, we point out that in long-term adaptation, such as incremental on-line adaptation, the supervised approach is a necessity.
SL980313.PDF (From Author)
SL980313.PDF (Rasterized)
TOP
Cluster Adaptive Training for Speech Recognition
Authors:
Mark J.F. Gales, IBM Almaden Research Center (USA)
Page (NA) Paper number 375
Abstract:
When performing speaker adaptation there are two conflicting requirements. The transform must be powerful enough to model the speaker. Second, the transform should be rapidly estimated for any particular speaker. Recently the most popular adaptation schemes have used many parameters to adapt the models. This paper examines an adaptation scheme requiring few parameters to adapt the models, cluster adaptive training. It may be viewed as a simple extension to speaker clustering. A linear interpolation of the cluster means is used as the mean of the particular speaker. This scheme naturally falls into an adaptive training framework. Maximum likelihood estimates of the interpolation weights are given. Furthermore, re-estimation formulae for cluster means, represented both explicitly and by sets of transforms of some canonical mean, are given. On a speaker-independent task CAT reduced the word error rate using very little adaptation data compared to a standard system. a speaker independent model set.