Data-Efficient Semi-Supervised Few-Shot Speaker Verification via Prototype Space Optimization

This study proposes an i-Vector based approach to gender identification. Gender-labeled utterances from the Fisher English (FE) corpus are used to formulate an i-Vector extraction framework, and a Probabilistic Linear Discriminant Analysis (PLDA) back-end is employed to compute the scores for gender identification. A novel duration mismatch compensation strategy is also presented that offers very little degradation in identification accuracy even with a large reduction in the duration of the test-segment. The proposed method is shown to consistently outperform a GMM-UBM based gender-identification scheme on several test-sets created from a held-out portion of the FE corpus, and is able to achieve an identification accuracy of up to 97.63%. On the severely distorted and multilingual DARPA-RATS (Robust Automatic Transcription of Speech) corpora, the proposed approach achieves an identification accuracy of 76.48% using only the FE data in training. Next, a novel unsupervised domain adaptation strategy is also presented that utilizes only unlabeled RATS data to adapt the out-of-domain PLDA parameters derived from the FE training data. The strategy is able to offer a 6.8% relative improvement in identification accuracy, and a 14.75% relative reduction in Equal Error Rate (EER) compared to using the out-of-domain PLDA model on the RATS test-utterances. These improvements are significant since: 1) RATS test-utterances are severely distorted, 2) No labeled data of any kind is used for 4 of the 5 languages present in the test-utterances.

Stochastic vector mapping-based feature enhancement using prior-models and model adaptation for noisy speech recognition

It is widely known that the quality of confidence measure is critical for speech applications. In this paper, we present our recent work on improving word confidence scores by calibrating them using a small set of calibration data when only the recognized word sequence and associated raw confidence scores are made available. The core of our technique is the maximum entropy model with distribution constraints which naturally and effectively make use of the word distribution, the raw confidence-score distribution, and the context information. We demonstrate the effectiveness of our approach by showing that it can achieve relative 38% mean square error (MSE), 39% negative normalized likelihood (NNLL), and 23% equal error rate (EER) reduction on a voice mail transcription data set and relative 35% MSE, 45% NNLL, and 35% EER reduction on a command and control data set.

Optimization of the parameters characterizing sigmoidal rate-level functions based on acoustic features

In this paper appropriate confidence measures (CMs) are investigated for Mandarin command word recognition, both in the so-called target region and non-target region, respectively. Here the target region refers to the recognized speech part of command word while the non-target region refers to the recognized silence part. It shows that exploiting extra information in the non-target region can effectively complement the traditional CM which usually focus on the target region. Furthermore, when analyzing the non-target region in a more theoretical way, where Bayesian information criterion (BIC) is employed to locate more precise boundary in the non-target region, even more improvement is achieved. In two different Mandarin telephone command word tasks, more than 20% relative reduction of equal error rate (EER) is obtained.

Most of the speaker recognition systems use system features for speaker recognition which are mostly spectral in nature. Recently, there has been significant work on using source features, viz., prosodies and pitch dynamics, glottal flow derivative, Linear Prediction (LP) residual and its phase, wavelet-domain representation of LP residual, etc for speaker recognition. In this paper, a new source-like feature set, viz., Teager Energy Operator (TEO) Phase is developed for speaker recognition. Proposed TEO Phase has several salient features as compared to that of LP residual phase, viz., less number of spurious peaks in Hilbert envelope plot, and better perceptual and sample correlation with speech signal. Furthermore, this avoids use of voiced/unvoiced detection, preprocessing techniques such as windowing and pre-emphasis and LP residual computation. Experiments have been carried out for speaker recognition task using TEO Phase and LP residual phase with polynomial classifier of 2 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">nd</sup> order approximation. For speaker verification, a reduction in equal error rate (EER) by 2.62 % over LP residual phase is achieved for proposed feature set.

We discuss the limitations of the i-vector representation of speech segments in speaker recognition and explain how Joint Factor Analysis (JFA) can serve as an alternative feature extractor in a variety of ways. Building on the work of Zhao and Dong, we implemented a variational Bayes treatment of JFA which accommodates adaptation of universal background models (UBMs) in a natural way. This allows us to experiment with several types of features for speaker recognition: speaker factors and diagonal factors in addition to i-vectors, extracted with and without UBM adaptation in each case. We found that, in text-independent speaker verification experiments on NIST data, extracting i-vectors with UBM adaptation led to a 10% reduction in equal error rates although performance did not improve consistently over the whole DET curve. We achieved a further 10% reduction (with a similar inconsistency) by using speaker factors extracted with UBM adaptation as features. In text-dependent speaker recognition experiments on RSR2015 data, we were able to achieve very good performance using a JFA model with diagonal factors but no speaker factors as a feature extractor. Contrary to standard practice, this JFA model was configured so as to model speakerphrase combinations (rather than speakers) and it was trained on utterances of very short duration (rather than whole recording sessions). We also present a variant of the length normalization trick inspired by uncertainty propagation which leads to substantial gains in performance over the whole DET curve.

Nonparametric Bayesian models use a Bayesian framework to learn the model complexity automatically from the data and eliminate the need for a complex model selection process. The Hierarchical Dirichlet Process hidden Markov model (HDP-HMM) is the nonparametric Bayesian equivalent of an HMM. However, HDP-HMM is restricted to an ergodic topology and uses a Dirichlet Process Model (DPM) to achieve a mixture distribution-like model. For applications such as speech recognition, where we deal with ordered sequences, it is desirable to impose a left-to-right structure on the model to improve its ability to model the sequential nature of the speech signal. In this paper, we introduce three enhancements to HDP-HMM: (1) a left-to-right structure: needed for sequential decoding of speech, (2) non-emitting initial and final states: required for modeling finite length sequences, (3) HDP mixture emissions: allows sharing of data across states. The latter is particularly important for speech recognition because Gaussian mixture models have been very effective at modeling speaker variability. Further, due to the nature of language, some models occur infrequently and have a small number of data points associated with them, even for large corpora. Sharing allows these models to be estimated more accurately. We demonstrate that this new HDP-HMM model produces a 15% increase in likelihoods and a 15% relative reduction in error rate on a phoneme classification task based on the TIMIT Corpus.

The automatic speaker verification (ASV) has achieved significant progress in recent years. However, it is still very challenging to generalize the ASV technologies to new, unknown and spoofing conditions. Most previous studies focused on extracting the speaker information from natural speech. This paper attempts to address the speaker verification from another perspective. The speaker identity information was exploited from singing speech. We first designed and released a new corpus for speaker verification based on singing and normal reading speech. Then, the speaker discrimination was compared and analyzed between natural and singing speech in different feature spaces. Furthermore, the conventional Gaussian mixture model, the dynamic time warping and the state-of-the-art deep neural network were investigated. They were used to build text-dependent ASV systems with different training-test conditions. Experimental results show that the voiceprint information in the singing speech was more distinguishable than the one in the normal speech. More than relative 20% reduction of equal error rate was obtained on both the gender-dependent and independent 1 s-1 s evaluation tasks.

In this paper, we propose a dynamic in-search data selection method to diagnose competing information automatically from speech data. In our method, the Viterbi beam search is used to decode all training data. During decoding, all partial paths within the beam are examined to identify the so-called competing-token and true-token sets for each individual hidden Markov model (HMM). In this work, the collected data tokens are used for acoustic modeling and utterance verification as two specific examples. In acoustic modeling, the true-token sets are used to adapt HMMs with a sequential maximum a posteriori adaptation method, while a generalized probabilistic descent-based discriminative training method is proposed to improve HMMs based on competing-token sets. In utterance verification, under the framework of likelihood ratio testing, the true-token sets are employed to train positive models for the null hypothesis and the competing-token sets are used to estimate negative models for the alternative hypothesis. All the proposed methods are evaluated in Bell Laboratories communicator system. Experimental results show that the new acoustic modeling method can consistently improve recognition performance over our best maximum likelihood estimation models, roughly 1% absolute reduction in word error rate. The results also show the new verification models can significantly improve the performance of utterance verification over the conventional anti models, almost relatively 30% reduction of equal error rate when identifying misrecognized words from the recognition results.

Long Short-Term Memory (LSTM) is the primary recurrent neural networks architecture for acoustic modeling in automatic speech recognition systems. Residual learning is an efficient method to help neural networks converge easier and faster. In this paper, we propose several types of residual LSTM methods for our acoustic modeling. Our experiments indicate that, compared with classic LSTM, our architecture shows more than 8% relative reduction in Phone Error Rate (PER) on TIMIT tasks. At the same time, our residual fast LSTM approach shows 4% relative reduction in PER on the same task. Besides, we find that all this architecture could have good results on THCHS-30, Librispeech and Switchboard corpora.

Automatic speech recognition (ASR) systems rely almost exclusively on short-term segment-level features (MFCCs), while ignoring higher level suprasegmental cues that are characteristic of human speech. However, recent experiments have shown that categorical representations of prosody, such as those based on the Tones and Break Indices (ToBI) annotation standard, can be used to enhance speech recognizers. However, categorical prosody models are severely limited in scope and coverage due to the lack of large corpora annotated with the relevant prosodic symbols (such as pitch accent, word prominence, and boundary tone labels). In this paper, we first present an architecture for augmenting a standard ASR with symbolic prosody. We then discuss two novel, un-supervised adaptation techniques for improving, respectively, the quality of the linguistic and acoustic components of our categorical prosody models. Finally, we implement the augmented ASR by enriching ASR lattices with the adapted categorical prosody models. Our experiments show that the proposed unsupervised adaptation techniques significantly improve the quality of the prosody models; the adapted prosodic language and acoustic models reduce binary pitch accent (presence versus absence) classification error rate by 13.8% and 4.3%, respectively (relative to the seed models) on the Boston University Radio News Corpus, while the prosody-enriched ASR exhibits a 3.1% relative reduction in word error rate (WER) over the baseline system.

A smoothing method for minimum phone error linear regression (MPELR) is proposed in this paper. We show that the objective function for minimum phone error (MPE) can be combined with a prior mean distribution. When the prior mean distribution is based on maximum likelihood (ML) estimates, the proposed method is the same as the previous smoothing technique for MPELR. Instead of ML estimates, maximum a posteriori (MAP) parameter estimate is used to define the mode of prior mean distribution to improve the performance of MPELR. Experiments on a large vocabulary speech recognition task show that the proposed method can obtain 8.4% relative reduction in word error rate when the amount of data is limited, while retaining the same asymptotic performance as conventional MPELR. When compared with discriminative maximum a posteriori linear regression (DMAPLR), the proposed method shows improvement except for the case of limited adaptation data for supervised adaptation.

We present the AMIDA 2009 system for participation in the NIST RT’2009 STT evaluations. Systems for close-talking, far field and speaker attributed STT conditions are described. Im- provements to our previous systems are: segmentation and diar- isation; stacked bottle-neck posterior feature extraction; fMPE training of acoustic models; adaptation on complete meetings; improvements to WFST decoding; automatic optimisation of decoders and system graphs. Overall these changes gave a 6- 13% relative reduction in word error rate while at the same time reducing the real-time factor by a factor of five and using con- siderably less data for acoustic model training.

Fast and accurate acoustic modelling with semi-continuous HMMs