Noise robust ASR in reverberated multisource environments applying convolutive NMF and Long Short-Term Memory

This article proposes and evaluates various methods to integrate the concept of bidirectional Long Short-Term Memory (BLSTM) temporal context modeling into a system for automatic speech recognition (ASR) in noisy and reverberated environments. Building on recent advances in Long Short-Term Memory architectures for ASR, we design a novel front-end for context-sensitive Tandem feature extraction and show how the Connectionist Temporal Classification approach can be used as a BLSTM-based back-end, alternatively to Hidden Markov Models (HMM). We combine context-sensitive BLSTM-based feature generation and speech decoding techniques with source separation by convolutive non-negative matrix factorization. Applying our speaker adapted multi-stream HMM framework that processes MFCC features from NMF-enhanced speech as well as word predictions obtained via BLSTM networks and non-negative sparse classification (NSC), we obtain an average accuracy of 91.86% on the PASCAL CHiME Challenge task at signal-to-noise ratios ranging from ?6 to 9 dB. To our knowledge, this is the best result ever reported for the CHiME Challenge task.     

基于知识蒸馏和生成对抗网络的远场语音识别

为了进一步利用近场语音数据来提高远场语音识别的性能,提出一种基于知识蒸馏和生成对抗网络相结合的远场语音识别算法.该方法引入多任务学习框架,在进行声学建模的同时对远场语音特征进行增强.为了提高声学建模能力,使用近场语音的声学模型（老师模型）来指导远场语音的声学模型（学生模型）进行训练.通过最小化相对熵使得学生模型的后验概率分布逼近老师模型.为了提升特征增强的效果,加入鉴别网络来进行对抗训练,从而使得最终增强后的特征分布更逼近近场特征.AMI数据集上的实验结果表明,该算法的平均词错误率（WER）与基线相比在单通道的情况下,在没有说话人交叠和有说话人交叠时分别相对下降5.6%和4.7%.在多通道的情况下,在没有说话人交叠和有说话人交叠时分别相对下降6.2%和4.1%.TIMIT数据集上的实验结果表明,该算法获得了相对7.2%的平均词错误率下降.为了更好地展示生成对抗网络对语音增强的作用,对增强后的特征进行了可视化分析,进一步验证了该方法的有效性.     

基于非线性堆叠双向网络的端到端声纹识别

传统声纹识别方法过程繁琐且识别率低,现有的深度学习方法所使用的神经网络对语音信号没有针对性从而导致识别精度不够。针对上述问题,本文提出一种基于非线性堆叠双向LSTM的端到端声纹识别方法。首先,对原始语音文件提取出Fbank特征用于网络模型的输入。然后,针对语音信号连续且前后关联性强的特点,构建双向长短时记忆网络处理语音数据提取深度特征,为进一步增强网络的非线性表达能力,利用堆叠多层双向LSTM层和多层非线性层实现对语音信号更深层次抽象特征的提取。最后,使用SGD优化器优化训练方式。实验结果表明提出的方法能够充分利用语音序列信号特征,具有较强的时序全面性和非线性表达能力,所构造模型整体性强,比GRU和LSTM等模型具有更好的识别效果。     

Modelling non-stationary noise with spectral factorisation in automatic speech recognition

Speech recognition systems intended for everyday use must be able to cope with a large variety of noise types and levels, including highly non-stationary multi-source mixtures. This study applies spectral factorisation algorithms and long temporal context for separating speech and noise from mixed signals. To adapt the system to varying environments, noise models are acquired from the context, or learnt from the mixture itself without prior information. We also propose methods for reducing the size of the bases used for speech and noise modelling by 20–40 times for better practical applicability. We evaluate the performance of the methods both as a standalone classifier and as a signal-enhancing front-end for external recognisers. For the CHiME noisy speech corpus containing non-stationary multi-source household noises at signal-to-noise ratios ranging from +9 to ?6 dB, we report average keyword recognition rates up to 87.8% using a single-stream sparse classification algorithm.     

面向语音谎言检测的智能机器学习特征融合算法研究

针对目前语音谎言检测识别效果差、特征提取不充分等问题,提出了一种基于注意力机制的欺骗语音识别网络。首先,将双向长短时记忆与帧级声学特征相结合,其中帧级声学特征的维数随语音长度的变化而变化,从而有效提取声学特征。其次,采用基于时间注意增强卷积双向长短时记忆模型作为分类算法,使分类器能够从输入中学习与任务相关的深层信息,提高识别性能。最后,采用跳跃连接机制将时间注意增强卷积双向长短时记忆模型的底层输出直接连接到全连接层,从而充分利用了学习到的特征,避免了消失梯度的问题。实验阶段,与LSTM以及其他基准模型进行对比,所提模型性能最优。仿真结果进一步验证了所提模型对语音谎言检测领域发展及提升识别率提供了一定借鉴作用。     

A statistical model for robust integration of narrowband cues in speech

We investigate a statistical model for integrating narrowband cues in speech. The model is inspired by two ideas in human speech perception: (i) Fletcher’s hypothesis (1953) that independent detectors, working in narrow frequency bands, account for the robustness of auditory strategies, and (ii) Miller and Nicely’s analysis (1955) that perceptual confusions in noisy bandlimited speech are correlated with phonetic features. We apply the model to detecting the phonetic feature [ +  /  − sonorant] that distinguishes vowels, approximants, and nasals (sonorants) from stops, fricatives, and affricates (obstruents). The model is represented by a multilayer probabilistic network whose binary hidden variables indicate sonorant cues from different parts of the frequency spectrum. We derive the Expectation-Maximization algorithm for estimating the model’s parameters and evaluate its performance on clean and corrupted speech.     

Discriminative n-gram language modeling

This paper describes discriminative language modeling for a large vocabulary speech recognition task. We contrast two parameter estimation methods: the perceptron algorithm, and a method based on maximizing the regularized conditional log-likelihood. The models are encoded as deterministic weighted finite state automata, and are applied by intersecting the automata with word-lattices that are the output from a baseline recognizer. The perceptron algorithm has the benefit of automatically selecting a relatively small feature set in just a couple of passes over the training data. We describe a method based on regularized likelihood that makes use of the feature set given by the perceptron algorithm, and initialization with the perceptron’s weights; this method gives an additional 0.5% reduction in word error rate (WER) over training with the perceptron alone. The final system achieves a 1.8% absolute reduction in WER for a baseline first-pass recognition system (from 39.2% to 37.4%), and a 0.9% absolute reduction in WER for a multi-pass recognition system (from 28.9% to 28.0%).     

LSTM-Modeling of continuous emotions in an audiovisual affect recognition framework

Automatically recognizing human emotions from spontaneous and non-prototypical real-life data is currently one of the most challenging tasks in the field of affective computing. This article presents our recent advances in assessing dimensional representations of emotion, such as arousal, expectation, power, and valence, in an audiovisual human–computer interaction scenario. Building on previous studies which demonstrate that long-range context modeling tends to increase accuracies of emotion recognition, we propose a fully automatic audiovisual recognition approach based on Long Short-Term Memory (LSTM) modeling of word-level audio and video features. LSTM networks are able to incorporate knowledge about how emotions typically evolve over time so that the inferred emotion estimates are produced under consideration of an optimal amount of context. Extensive evaluations on the Audiovisual Sub-Challenge of the 2011 Audio/Visual Emotion Challenge show how acoustic, linguistic, and visual features contribute to the recognition of different affective dimensions as annotated in the SEMAINE database. We apply the same acoustic features as used in the challenge baseline system whereas visual features are computed via a novel facial movement feature extractor. Comparing our results with the recognition scores of all Audiovisual Sub-Challenge participants, we find that the proposed LSTM-based technique leads to the best average recognition performance that has been reported for this task so far.     

Spoken emotion recognition through optimum-path forest classification using glottal features

A new method for the recognition of spoken emotions is presented based on features of the glottal airflow signal. Its effectiveness is tested on the new optimum path classifier (OPF) as well as on six other previously established classification methods that included the Gaussian mixture model (GMM), support vector machine (SVM), artificial neural networks – multi layer perceptron (ANN-MLP), k-nearest neighbor rule (k-NN), Bayesian classifier (BC) and the C4.5 decision tree. The speech database used in this work was collected in an anechoic environment with ten speakers (5 M and 5 F) each speaking ten sentences in four different emotions: Happy, Angry, Sad, and Neutral. The glottal waveform was extracted from fluent speech via inverse filtering. The investigated features included the glottal symmetry and MFCC vectors of various lengths both for the glottal and the corresponding speech signal. Experimental results indicate that best performance is obtained for the glottal-only features with SVM and OPF generally providing the highest recognition rates, while for GMM or the combination of glottal and speech features performance was relatively inferior. For this text dependent, multi speaker task the top performing classifiers achieved perfect recognition rates for the case of 6th order glottal MFCCs.     

Random forests and the data sparseness problem in language modeling

Language modeling is the problem of predicting words based on histories containing words already hypothesized. Two key aspects of language modeling are effective history equivalence classification and robust probability estimation. The solution of these aspects is hindered by the data sparseness problem.Application of random forests (RFs) to language modeling deals with the two aspects simultaneously. We develop a new smoothing technique based on randomly grown decision trees (DTs) and apply the resulting RF language models to automatic speech recognition. This new method is complementary to many existing ones dealing with the data sparseness problem. We study our RF approach in the context of n-gram type language modeling in which n − 1 words are present in a history. Unlike regular n-gram language models, RF language models have the potential to generalize well to unseen data, even when histories are longer than four words. We show that our RF language models are superior to the best known smoothing technique, the interpolated Kneser–Ney smoothing, in reducing both the perplexity (PPL) and word error rate (WER) in large vocabulary state-of-the-art speech recognition systems. In particular, we will show statistically significant improvements in a contemporary conversational telephony speech recognition system by applying the RF approach only to one of its many language models.     

A hearing-inspired approach for distant-microphone speech recognition in the presence of multiple sources

This paper addresses the problem of speech recognition in reverberant multisource noise conditions using distant binaural microphones. Our scheme employs a two-stage fragment decoding approach inspired by Bregman's account of auditory scene analysis, in which innate primitive grouping ‘rules’ are balanced by the role of learnt schema-driven processes. First, the acoustic mixture is split into local time-frequency fragments of individual sound sources using signal-level primitive grouping cues. Second, statistical models are employed to select fragments belonging to the sound source of interest, and the hypothesis-driven stage simultaneously searches for the most probable speech/background segmentation and the corresponding acoustic model state sequence. The paper reports recent advances in combining adaptive noise floor modelling and binaural localisation cues within this framework. By integrating signal-level grouping cues with acoustic models of the target sound source in a probabilistic framework, the system is able to simultaneously separate and recognise the sound of interest from the mixture, and derive significant recognition performance benefits from different grouping cue estimates despite their inherent unreliability in noisy conditions. Finally, the paper will show that missing data imputation can be applied via fragment decoding to allow reconstruction of a clean spectrogram that can be further processed and used as input to conventional ASR systems. The best performing system achieves an average keyword recognition accuracy of 85.83% on the PASCAL CHiME Challenge task.     

Sparse imputation for large vocabulary noise robust ASR

An effective way to increase noise robustness in automatic speech recognition is to label the noisy speech features as either reliable or unreliable (‘missing’), and replace (‘impute’) the missing ones by clean speech estimates. Conventional imputation techniques employ parametric models and impute the missing features on a frame-by-frame basis. At low SNRs, frame-based imputation techniques fail because many time frames contain few, if any, reliable features. In previous work, we introduced an exemplar-based method, dubbed sparse imputation, which can impute missing features using reliable features from neighbouring frames. We achieved substantial gains in performance at low SNRs for a connected digit recognition task. In this work, we investigate whether the exemplar-based approach can be generalised to a large vocabulary task.Experiments on artificially corrupted speech show that sparse imputation substantially outperforms a conventional imputation technique when the ideal ‘oracle’ reliability of features is used. With error-prone estimates of feature reliability, sparse imputation performance is comparable to our baseline imputation technique in the cleanest conditions, and substantially better at lower SNRs. With noisy speech recorded in realistic noise conditions, sparse imputation performs slightly worse than our baseline imputation technique in the cleanest conditions, but substantially better in the noisier conditions.     

Speech emotion recognition: Features and classification models

To solve the speaker independent emotion recognition problem, a three-level speech emotion recognition model is proposed to classify six speech emotions, including sadness, anger, surprise, fear, happiness and disgust from coarse to fine. For each level, appropriate features are selected from 288 candidates by using Fisher rate which is also regarded as input parameter for Support Vector Machine (SVM). In order to evaluate the proposed system, principal component analysis (PCA) for dimension reduction and artificial neural network (ANN) for classification are adopted to design four comparative experiments, including Fisher + SVM, PCA + SVM, Fisher + ANN, PCA + ANN. The experimental results proved that Fisher is better than PCA for dimension reduction, and SVM is more expansible than ANN for speaker independent speech emotion recognition. The average recognition rates for each level are 86.5%, 68.5% and 50.2% respectively.     

混合多尺度卷积结合双层LSTM语音情感识别

针对深度学习算法在语音情感特征提取方面的不足以及识别准确率不高的问题,本文通过提取语音数据中有效的情感特征,并将特征进行多尺度拼接融合,构造语音情感特征,提高深度学习模型对特征的表现能力。传统递归神经网络无法解决语音情感识别长时依赖问题,本文采用双层LSTM模型来改进语音情感识别效果,提出一种混合多尺度卷积与双层LSTM模型相结合的模型。实验结果表明,在中科院自动化所汉语情感数据库（CASIA）和德国柏林情感公开数据集（Emo-DB）下,本文所提语音情感识别模型相较于其他情感识别模型在准确率方面有较大提高。     

Using geometric spectral subtraction approach for feature extraction for DSR front-end Arabic system

Noise robustness and Arabic language are still considered as the main challenges for speech recognition over mobile environments. This paper contributed to these trends by proposing a new robust Distributed Speech Recognition (DSR) system for Arabic language. A speech enhancement algorithm was applied to the noisy speech as a robust front-end pre-processing stage to improve the recognition performance. While an isolated Arabic word engine was designed, and developed using HMM Model to perform the recognition process at the back-end. To test the engine, several conditions including clean, noisy and enhanced noisy speech were investigated together with speaker dependent and speaker independent tasks. With the experiments carried out on noisy database, multi-condition training outperforms the clean training mode in all noise types in terms of recognition rate. The results also indicate that using the enhancement method increases the DSR accuracy of our system under severe noisy conditions especially at low SNR down to 10 dB.     

一种基于时频域特征融合的语音增强方法

为充分利用含噪语音特征来提升深度神经网络的语音增强性能,提出一种融合时频域特征的语音增强方法。以含噪语音的波形和纯净语音的对数功率谱分别作为训练特征和训练目标,获取含噪语音时域特征到纯净语音频域特征的映射关系。将含噪语音的波形和对数功率谱共同作为训练特征,构建融合含噪语音时域和频域特征的深度神经网络实现语音增强。实验结果表明,与单纯使用频域特征的语音增强方法相比,该方法能够明显提升增强语音的质量和可懂度,具有更好的语音增强性能。     

A new variance-based approach for discriminative feature extraction in machine hearing classification using spectrogram features

Machine hearing is an emerging research field that is analogous to machine vision in that it aims to equip computers with the ability to hear and recognise a variety of sounds. It is a key enabler of natural human–computer speech interfacing, as well as in areas such as automated security surveillance, environmental monitoring, smart homes/buildings/cities. Recent advances in machine learning allow current systems to accurately recognise a diverse range of sounds under controlled conditions. However doing so in real-world noisy conditions remains a challenging task. Several front–end feature extraction methods have been used for machine hearing, employing speech recognition features like MFCC and PLP, as well as image-like features such as AIM and SIF. The best choice of feature is found to be dependent upon the noise environment and machine learning techniques used. Machine learning methods such as deep neural networks have been shown capable of inferring discriminative classification rules from less structured front–end features in related domains. In the machine hearing field, spectrogram image features have recently shown good performance for noise-corrupted classification using deep neural networks. However there are many methods of extracting features from spectrograms. This paper explores a novel data-driven feature extraction method that uses variance-based criteria to define spectral pooling of features from spectrograms. The proposed method, based on maximising the pooled spectral variance of foreground and background sound models, is shown to achieve very good performance for robust classification.     

短文本分类的ResLCNN模型

短文本分类是互联网文本数据处理中的关键任务之一.长短时记忆网络LSTM（long short-term memory）和卷积神经网络CNN（convolutional neural network）是广泛应用于短文本分类任务的两种深度学习模型.在计算机视觉和语音识别领域的深度学习研究表明,深层次的神经网络模型具有较好的表达数据特征的能力.受此启发,面向文本深度学习分类问题,提出基于3层LSTM和CNN网络结构的ResLCNN（residual-LSTM-CNN）深度学习模型.该模型有效结合LSTM获取文本序列数据的长距离依赖特征和CNN通过卷积操作获取句子局部特征的优势,同时借鉴残差模型理论,在第1层LSTM层与CNN层之间加入恒等映射,构建残差层,缓解深层模型梯度消失问题.为了探究深层短文本分类中ResLCNN模型的文本分类能力,在多种数据集上将其与LSTM、CNN及其组合模型进行对比实验.结果表明,相比于单层LSTM与CNN组合模型,ResLCNN深层模型在MR、SST-2和SST-5数据集上分别提高了1.0%、0.5%、0.47%的准确率,取得了更好的分类效果.