In this letter, we propose a new approach to SVM based speaker recognition, which utilizes a kind of novel phonotactic information as the feature for SVM modeling. Gaussian mixture models (GMMs) have been proven extremely successful for text-independent speaker recognition. The GMM universal background model (UBM) is a speaker-independent model, each component of which can be considered as modeling some underlying phonetic sound classes. We assume that the utterances from different speakers should get different average posterior probabilities on the same Gaussian component of the UBM, and the supervector composed of the average posterior probabilities on all components of the UBM for each utterance should be discriminative. We use these supervectors as the features for SVM based speaker recognition. Experiment results on a NIST SRE 2006 task show that the proposed approach demonstrates comparable performance with the commonly used systems. Fusion results are also presented.

In this paper, we presented a novel method for automatic pronunciation quality assessment. Unlike the popular “Goodness of Pronunciation” (GOP) method, this method does not map the decoding confidence into pronunciation quality score, but differentiates the different pronunciation quality utterances directly. In this method, the student's utterance need to be decoded for two times. The first-time decoding was for getting the time points of each phone of the utterance by a forced alignment using a conventional trained acoustic model (AM). The second-time decoding was for differentiating the pronunciation quality for each triphone using a specially trained AM, where the triphones in different pronunciation qualities were trained as different units, and the model was trained in discriminative method to ensure the model has the best discrimination among the triphones whose names were same but pronunciation quality scores were different. The decoding network in the second-time decoding included different pronunciation quality triphones, so the phone-level scores can be obtained from the decoding result directly. The phone-level scores were combined into the sentence-level scores using maximum entropy criterion. The experimental results shows that the scoring performance was increased significantly compared to the GOP method, especially in sentence-level.

In this letter, we adopt a new factor analysis of neighborhood-preserving embedding (NPE) for speaker verification. NPE aims at preserving the local neighborhood structure on the data and defines a low-dimensional speaker space called neighborhood-preserving embedding space. We compare the proposed method with the state-of-the-art total variability approach on the telephone-telephone core condition of the NIST 2008 Speaker Recognition Evaluation (SRE) dataset. The experimental results indicate that the proposed NPE method outperforms the total variability approach, providing up to 24% relative improvement.

Mel Frequency Cepstral Coefficients (MFCC) are the most popular acoustic features used in automatic speech recognition (ASR), mainly because the coefficients capture the most useful information of the speech and fit well with the assumptions used in hidden Markov models. As is well known, MFCCs already employ several principles which have known counterparts in the peripheral properties of human hearing: decoupling across frequency, mel-warping of the frequency axis, log-compression of energy, etc. It is natural to introduce more mechanisms in the auditory periphery to improve the noise robustness of MFCC. In this paper, a k-nearest neighbors based frequency masking filter is proposed to reduce the audibility of spectra valleys which are sensitive to noise. Besides, Moore and Glasberg's critical band equivalent rectangular bandwidth (ERB) expression is utilized to determine the filter bandwidth. Furthermore, a new bandpass infinite impulse response (IIR) filter is proposed to imitate the temporal masking phenomenon of the human auditory system. These three auditory perceptual mechanisms are combined with the standard MFCC algorithm in order to investigate their effects on ASR performance, and a revised MFCC extraction scheme is presented. Recognition performances with the standard MFCC, RASTA perceptual linear prediction (RASTA-PLP) and the proposed feature extraction scheme are evaluated on a medium-vocabulary isolated-word recognition task and a more complex large vocabulary continuous speech recognition (LVCSR) task. Experimental results show that consistent robustness against background noise is achieved on these two tasks, and the proposed method outperforms both the standard MFCC and RASTA-PLP.

The optimization of spoken dialog management policies is a non-trivial task due to the erroneous inputs from speech recognition and language understanding modules. The dialog manager needs to ground uncertain semantic information at times to fully understand the need of human users and successfully complete the required dialog tasks. Approaches based on reinforcement learning are currently mainstream in academia and have been proved to be effective, especially when operating in noisy environments. However, in reinforcement learning the dialog strategy is often represented by complex numeric model and thus is incomprehensible to humans. The trained policies are very difficult for dialog system designers to verify or modify, which largely limits the deployment for commercial applications. In this paper we propose a novel framework for optimizing dialog policies specified in human-readable domain language using genetic algorithm. We present learning algorithms using user simulator and real human-machine dialog corpora. Empirical experimental results show that the proposed approach can achieve competitive performance on par with some state-of-the-art reinforcement learning algorithms, while maintaining a comprehensible policy structure.

Introducing pronunciation models into decoding has been proven to be benefit to LVCSR. In this paper, a discriminative pronunciation modeling method is presented, within the framework of the Minimum Phone Error (MPE) training for HMM/GMM. In order to bring the pronunciation models into the MPE training, the auxiliary function is rewritten at word level and decomposes into two parts. One is for co-training the acoustic models, and the other is for discriminatively training the pronunciation models. On Mandarin conversational telephone speech recognition task, compared to the baseline using a canonical lexicon, the discriminative pronunciation models reduced the absolute Character Error Rate (CER) by 0.7% on LDC test set, and with the acoustic model co-training, 0.8% additional CER decrease had been achieved.

Although end-to-end based speech recognition research for Mandarin-English code-switching has attracted increasing interests, it remains challenging due to data scarcity. Meta-learning approach is popular with low-resource modeling using high-resource data, but it does not make full use of low-resource code-switching data. Therefore we propose a two-fold cross-validation training framework combined with meta-learning approach. Experiments on the SEAME corpus demonstrate the effects of our method.

Semantic concept in an utterance is obtained by a fuzzy matching methods to solve problems such as words' variation induced by automatic speech recognition (ASR), or missing field of key information by users in the process of spoken language understanding (SLU). A two-stage method is proposed: first, we adopt conditional random field (CRF) for building probabilistic models to segment and label entity names from an input sentence. Second, fuzzy matching based on similarity function is conducted between the named entities labeled by a CRF model and the reference characters of a dictionary. The experiments compare the performances in terms of accuracy and processing speed. Dice similarity and cosine similarity based on TF score can achieve better accuracy performance among four similarity measures, which equal to and greater than 93% in F1-measure. Especially the latter one improved by 8.8% and 9% respectively compared to q-gram and improved edit-distance, which are two conventional methods for string fuzzy matching.

This paper presents our developed decoder which adopts the idea of statically optimizing part of the knowledge sources while handling the others dynamically. The lexicon, phonetic contexts and acoustic model are statically integrated to form a memory-efficient state network, while the language model (LM) is dynamically incorporated on the fly by means of extended tokens. The novelties of our approach for constructing the state network are (1) introducing two layers of dummy nodes to cluster the cross-word (CW) context dependent fan-in and fan-out triphones, (2) introducing a so-called "WI layer" to store the word identities and putting the nodes of this layer in the non-shared mid-part of the network, (3) optimizing the network at state level by a sufficient forward and backward node-merge process. The state network is organized as a multi-layer structure for distinct token propagation at each layer. By exploiting the characteristics of the state network, several techniques including LM look-ahead, LM cache and beam pruning are specially designed for search efficiency. Especially in beam pruning, a layer-dependent pruning method is proposed to further reduce the search space. The layer-dependent pruning takes account of the neck-like characteristics of WI layer and the reduced variety of word endings, which enables tighter beam without introducing much search errors. In addition, other techniques including LM compression, lattice-based bookkeeping and lattice garbage collection are also employed to reduce the memory requirements. Experiments are carried out on a Mandarin spontaneous speech recognition task where the decoder involves a trigram LM and CW triphone models. A comparison with HDecode of HTK toolkits shows that, within 1% performance deviation, our decoder can run 5 times faster with half of the memory footprint.

Recently, automated recognition and analysis of human emotion has attracted increasing attention from multidisciplinary communities. However, it is challenging to utilize the emotional information simultaneously from multiple modalities. Previous studies have explored different fusion methods, but they mainly focused on either inter-modality interaction or intra-modality interaction. In this letter, we propose a novel two-stage fusion strategy named modality attention flow (MAF) to model the intra- and inter-modality interactions simultaneously in a unified end-to-end framework. Experimental results show that the proposed approach outperforms the widely used late fusion methods, and achieves even better performance when the number of stacked MAF blocks increases.

The introduction of deep neural networks (DNNs) leads to a significant improvement of the automatic speech recognition (ASR) performance. However, the whole ASR system remains sophisticated due to the dependent on the hidden Markov model (HMM). Recently, a new end-to-end ASR framework, which utilizes recurrent neural networks (RNNs) to directly model context-independent targets with connectionist temporal classification (CTC) objective function, is proposed and achieves comparable results with the hybrid HMM/DNN system. In this paper, we investigate per-dimensional learning rate methods, ADAGRAD and ADADELTA included, to improve the recognition of the end-to-end system, based on the fact that the blank symbol used in CTC technique dominates the output and these methods give frequent features small learning rates. Experiment results show that more than 4% relative reduction of word error rate (WER) as well as 5% absolute improvement of label accuracy on the training set are achieved when using ADADELTA, and fewer epochs of training are needed.

This paper proposes a new quantization framework on activation function of deep neural networks (DNN). We implement fixed-point DNN by quantizing the activations into powers-of-two integers. The costly multiplication operations in using DNN can be replaced with low-cost bit-shifts to massively save computations. Thus, applying DNN-based speech recognition on embedded systems becomes much easier. Experiments show that the proposed method leads to no performance degradation.

Short texts usually encounter the problem of data sparseness, as they do not provide sufficient term co-occurrence information. In this paper, we show how to mitigate the problem in short text classification through word embeddings. We assume that a short text document is a specific sample of one distribution in a Gaussian-Bayesian framework. Furthermore, a fast clustering algorithm is utilized to expand and enrich the context of short text in embedding space. This approach is compared with those based on the classical bag-of-words approaches and neural network based methods. Experimental results validate the effectiveness of the proposed method.

In this letter, we explored the usage of spatio-temporal information in one unified framework to improve the performance of multichannel speech recognition. Generalized cross correlation (GCC) is served as spatial feature compensation, and an attention mechanism across time is embedded within long short-term memory (LSTM) neural networks. Experiments on the AMI meeting corpus show that the proposed method provides a 8.2% relative improvement in word error rate (WER) over the model trained directly on the concatenation of multiple microphone outputs.

In this paper, the robustness of the posterior-based confidence measures is improved by utilizing entropy information, which is calculated for speech-unit-level posteriors using only the best recognition result, without requiring a larger computational load than conventional methods. Using different normalization methods, two posterior-based entropy confidence measures are proposed. Practical details are discussed for two typical levels of hidden Markov model (HMM)-based posterior confidence measures, and both levels are compared in terms of their performances. Experiments show that the entropy information results in significant improvements in the posterior-based confidence measures. The absolute improvements of the out-of-vocabulary (OOV) rejection rate are more than 20% for both the phoneme-level confidence measures and the state-level confidence measures for our embedded test sets, without a significant decline of the in-vocabulary accuracy.

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.

Weakly labelled semi-supervised audio tagging (AT) and sound event detection (SED) have become significant in real-world applications. A popular method is teacher-student learning, making student models learn from pseudo-labels generated by teacher models from unlabelled data. To generate high-quality pseudo-labels, we propose a master-teacher-student framework trained with a dual-lead policy. Our experiments illustrate that our model outperforms the state-of-the-art model on both tasks.

In this letter, a recently proposed clustering algorithm named affinity propagation is introduced for the task of speaker clustering. This novel algorithm exhibits fast execution speed and finds clusters with low error. However, experiments show that the speaker purity of affinity propagation is not satisfying. Thus, we propose a hybrid approach that combines affinity propagation with agglomerative hierarchical clustering to improve the clustering performance. Experiments show that compared with traditional agglomerative hierarchical clustering, the hybrid method achieves better performance on the test corpora.

Multi-task learning in deep neural networks has been proven to be effective for acoustic modeling in speech recognition. In the paper, this technique is applied to Mandarin-English code-mixing recognition. For the primary task of the senone classification, three schemes of the auxiliary tasks are proposed to introduce the language information to networks and improve the prediction of language switching. On the real-world Mandarin-English test corpus in mobile voice search, the proposed schemes enhanced the recognition on both languages and reduced the relative overall error rates by 3.5%, 3.8% and 5.8% respectively.

This paper presents our investigation into improving the performance of our previous automatic reading quality assessment system. The method of the baseline system is calculating the average value of the Phone Log-Posterior Probability (PLPP) of all phones in the voice to be assessed, and the average value is used as the reading quality assessment feature. In this paper, we presents three improvements. First, we cluster the triphones, and then calculate the average value of the normalized PLPP for each classification separately, and use this average values as the multi-dimensional assessment features instead of the original one-dimensional assessment feature. This method is simple but effective, which made the score difference of the machine scoring and manual scoring decrease by 30.2% relatively. Second, in order to assess the reading rhythm, we train Gaussian Mixture Models (GMM), which contain the information of each triphone's relative duration under standard pronunciation. Using the GMM, we can calculate the probability that the relative duration of each phone is conform to the standard pronunciation, and the average value of the probabilities is added to the assessment feature vector as a dimension of feature, which decreased the score difference between the machine scoring and manual scoring by 9.7% relatively. Third, we detect Filled Pauses (FP) by analyzing the formant curve, and then calculate the relative duration of FP, and add the relative duration of FP to the assessment feature vector as a dimension of feature. This method made the score difference between the machine scoring and manual scoring be further decreased by 10.2% relatively. Finally, when the feature vector extracted by the three methods are used together, the score difference between the machine scoring and manual scoring was decreased by 43.9% relatively compared to the baseline system.