Speaker and Language Recognition 3
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
Robust Speaker Verification Insensitive to Session-dependent Utterance Variation and Handset-dependent Distortion
Authors:
Tomoko Matsui, NTT Human Interface Labs. (Japan)
Kiyoaki Aikawa, NTT Human Interface Labs. (Japan)
Page (NA) Paper number 714
Abstract:
This paper investigates a method of creating robust speaker models that are not sensitive to session-dependent (SD) utterance-variation and handset-dependent (HD) distortion for HMM-based speaker verification systems in a real telephone network. We recently reported a method of creating session-independent (SI) speaker-HMMs that are not sensitive to SD utterance-variation. In that method, the distortion function that transforms SI speaker-HMMs to SD speaker-HMMs is introduced, and the parameters in the function and the speaker-HMM parameters are jointly estimated using a speaker adaptive training algorithm. This paper proposes a method that is less sensitive to SD utterance-variation and HD distortion than the previous method. This new idea focuses on different difficulties in estimating parameters in distortion functions for SD utterance-variation and HD distortion. In text-independent verification experiments using telephone speech data, the error reduction rate of the improved method compared with that of the conventional method of cepstral mean normalization is 24%.
SL980714.PDF (From Author)
SL980714.PDF (Rasterized)
TOP
A Comparative Evaluation of Variance Flooring Techniques in HMM-based Speaker Verification
Authors:
Haakan Melin, KTH, Dept. of Speech, Music and Hearing (Sweden)
Johan W. Koolwaaij, KUN, Dept. of Language & Speech (The Netherlands)
Johan Lindberg, KTH, Dept. of Speech, Music and Hearing (Sweden)
Frédéric Bimbot, IRISA / CNRS & INRIA - Sigma2 (France)
Page (NA) Paper number 467
Abstract:
The problem of how to train variance parameters on scarce data is addressed in the context of text-dependent, HMM-based, automatic speaker verification. Three variations of variance flooring is explored as a means to prevent over-fitting. With the best performing one, the floor to a variance vector of a client model is proportional to the corresponding variance vector in a non-client multi-speaker model. It is also found that creating a client-model by adapting the means and mixture weights from the non-client model while keeping variances constant works comparably to variance flooring and is much simpler. Comparisons are made on three large telephone quality corpora: Gandalf, SESP and Polycost.
SL980467.PDF (From Author)
SL980467.PDF (Rasterized)
TOP
Text-Independent Speaker Verification Using Automatically Labelled Acoustic Segments
Authors:
Dijana Petrovska-Delacrétaz, CIRC-EPFL (Switzerland)
Jan Cernocky, Technical University of Brno (Czech Republic)
Jean Hennebert, CIRC-EPFL (Switzerland)
Gérard Chollet, ENST, Paris (France)
Page (NA) Paper number 536
Abstract:
Most of the current text-independent speaker veri,cation techniques are based on modelling the global probability distribution function of speakers in the acoustic vector space. We present an alternative approach based on class-dependent veri,cation systems using automatically determined segmental units, obtained with temporal decomposition and labelled through unsupervised clustering. The core of the system is a set of multi-layer perceptrons (MLP) trained to discriminate between client and an independent set of world speakers. Each MLP is dedicated to work with data segments that are previously selected as belonging to a particular class. Issues and potential advantages of the segmental approach are presented. Performances of global and segmental approaches are tested on the NIST'98 database (250 female and 250 male speakers), showing promising results for the proposed new segmental approach. Comparison with a state of the art system, based on Gaussian Mixture Modelling is also included.
SL980536.PDF (From Author)
SL980536.PDF (Rasterized)
TOP
A Fast Decoding Algorithm Based on Sequential Detection of the Changes in Distribution
Authors:
Qi Li, Bell Labs, Lucent Technologies (USA)
Page (NA) Paper number 815
Abstract:
A fast algorithm for left-to-right HMM decoding is proposed in this paper. The algorithm is developed based on a sequential detection scheme which is asymptotically optimal in the sense of detecting a possible change in distribution as reliably and quickly as possible. The scheme is extended to HMM decoding in determining the state segmentations for likelihood or other score computations. As a sequential scheme, it can determine a state boundary in a few time steps after it occurs. The examples in this paper show that the proposed algorithm is 5 to 9 times faster than the Viterbi algorithm while it still can provide the same or similar decoding results. The proposed algorithm can be applied to speaker recognition, audio segmentation, voice/silence detection, and many other applications, where an assumption of the algorithm is usually satisfied.
SL980815.PDF (From Author)
SL980815.PDF (Rasterized)
TOP
Speaker Verification With Ensemble Classifiers Based On Linear Speech Transforms
Authors:
Jesper Østergaard Olsen, Aalborg University (Denmark)
Page (NA) Paper number 334
Abstract:
For most classifier architectures realistic training schemes only allow classifiers corresponding to local optima of the training criteria to be constructed. One way of dealing with this problem is to work with classifier ensembles: multiple classifiers are trained for the same classification problem and combined into one ``super'' classifier. The problem addressed in this paper is text prompted speaker verification by means of phoneme dependent Radial Basis Function networks trained by gradient descent error minimisation. In this context ensemble techniques are introduced by combining different classifiers that classify feature vectors, which have been pre-processed using different linear transforms. Four different types of linear transforms are studied: the Fisher transform, the LDA transform, the PCA transform and the cosine transform. The verification system is evaluated on the Gandalf database, where the equal error rate is reduced from 3.6% to 3.2% when ensemble techniques are introduced.
SL980334.PDF (From Author)
SL980334.PDF (Rasterized)
TOP
Speaker Recognition Based On Discriminative Projection Models
Authors:
Jesper Østergaard Olsen, Aalborg University (Denmark)
Page (NA) Paper number 335
Abstract:
A new discriminant speaker model is introduced in this paper. The model is text dependent and relies on characterising speakers in terms of the angular distance between ``projection vectors'', which allow good discrimination between individual speakers. The projection models require only little enrollment data to be available per target speaker, but at the same time require a set of ``cohort speakers'' to be available for which a relatively large amount of training speech is available per cohort speaker. The projection model technique is evaluated on the Gandalf database and compared to conventional Gaussian Mixture Models (GMMs). It is found that the projection models require less storage per target speaker, while at the same time achieving lower error rates, particularly when applied for speaker identification and recognition under mismatched conditions.