Abstract

Speaker normalization is one of the crucial aspects of an Automatic speech recognition system (ASR). Speaker normalization is employed to reduce the performance drop in ASR due to speaker variabilities. Traditional speaker normalization methods are mostly linear transforms over the input data estimated per speaker, such transforms would be efficient with sufficient data. In practical scenarios, only a single utterance from the test speaker is accessible. The present study explores speaker normalization methods for end-to-end speech recognition systems that could efficiently be performed even when single utterance from the unseen speaker is available. In this work, it is hypothesized that by suitably providing information about the speaker’s identity while training an end-to-end neural network, the capability to normalize the speaker variability could be incorporated into an ASR system. The efficiency of these normalization methods depends on the representation used for unseen speakers. In this work, the identity of the training speaker is represented in two different ways viz. i) by using a one-hot speaker code, ii) a weighted combination of all the training speakers identities. The unseen speakers from the test set are represented using a weighted combination of training speakers representations. Both the approaches have reduced the word error rate (WER) by 0.6, 1.3% WSJ corpus

Abstract

In this paper, a combination of excitation source information and vocal tract system information is explored for the task of language identification (LID). The excitation source information is represented by features extracted from linear prediction (LP) residual signal called the residual cepstral coefficients (RCC). Vocal tract system information is represented by the mel frequency cepstral coefficients (MFCC). In order to incorporate additional temporal information, shifted delta cepstra (SDC) are computed. An LID system is built using SDC over both MFCC and RCC features individually and evaluated based on their equal error rate (EER). Experiments have been performed on a dataset consisting of 13 Indian languages with about 115 h for training and 30 h for testing using a deep neural network (DNN), DNN with attention (DNN-WA) and a state-ofthe-art i-vector system. DNN-WA outperforms the baseline i-vector system. An EER of 9.93 and 6.25% are achieved using RCC and MFCC features respectively. By combining evidence from both features using a late fusion mechanism, an EER of 5.76% is obtained. This result indicates the complementary nature of the excitation source information to that of the widely used vocal tract system information for the task of LID.

Abstract

A degradation in the performance of automatic speech recognition systems (ASR) is observed in mismatched training and testing conditions. One of the reasons for this degradation is due to the presence of emotions in the speech. The main objective of this work is to improve the performance of ASR in the presence of emotional conditions using prosody modification. The influence of different emotions on the prosody parameters is exploited in this work. Emotion conversion methods are employed to generate the word level non-uniform prosody modified speech. Modification factors for prosodic components such as pitch, duration and energy are used. The prosody modification is done in two ways. Firstly, emotion conversion is done at the testing stage to generate the neutral speech from the emotional speech. Secondly, the ASR is trained with the generated emotional speech from the neutral speech. In this work, the presence of emotions in speech is studied for the Telugu ASR systems. A new database of IIIT-H Telugu speech corpus is collected to build the large vocabulary neutral Telugu speech ASR system. The emotional speech samples from IITKGP-SESC Telugu corpus are used for testing it. The emotions of anger, happiness and compassion are considered during the evaluation. An improvement in the performance of ASR systems is observed in the prosody modified speech.

Abstract

A dominating set of a small size is useful in several settings including wireless networks, document summarization, secure system design, and the like. In this paper, we start by studying three distributed algorithms that produce a small sized dominating sets in a few rounds. We interpret these algorithms in the natural shared memory setting and experiment with these algorithms on a multi-core CPU. Based on the observations from these experimental results, we propose variations to the three algorithms and also show how the proposed variations offer interesting trade-offs with respect to the size of the dominating set produced and the time taken.

Abstract

Recent developments in deep learning methods have greatly influenced the performances of speech recognition systems. In a Hidden Markov model-Deep neural network (HMM-DNN) based speech recognition system, DNNs have been employed to model senones (context dependent states of HMM),where HMMs capture the temporal relations among senones. Due to the use of more deeper networks significant improvement in the performances has been observed and developing deep learning methods to train more deeper architectures has gained a lot of scientific interest. Optimizing a deeper network is more complex ask than to optimize a less deeper network, but recently residual network have exhibited a capability to train a very deep neural network architectures and are not prone to vanishing/exploding gradient problems. In this work, the effectiveness of residual networks have been explored for of speech recognition. Along with the depth of the residual network, the criticality of width of the residual network has also been studied. It has been observe dthat at higher depth, width of the networks is also a crucial parameter for attaining significant improvements. A 14-hoursubset of WSJ corpus is used for training the speech recognition systems, it has been observed that the residual networks have shown much ease in convergence even with a depth much higher than the deep neural network. In this work, using residual networks an absolute reduction of 0.4 in WER error rates (8%reduction in the relative error) is attained compared to the best performing deep neural network.

Abstract

Significant developments in deep learning methods have been achieved with the capability to train more deeper networks. The performance of speech recognition system has been greatly improved by the use of deep learning techniques. Most of the developments in deep learning are associated with the development of new activation functions and the corresponding initializations. The development of Rectified linear units (ReLU)has revolutionized the use of supervised deep learning methods for speech recognition. Recently there has been a great deal of research interest in the development of activation functionsLeaky-ReLU (LReLU), Parametric-ReLU (PReLU), Exponentia lLinear units (ELU) and Parametric-ELU (PELU). This work isaimed at studying the influence of various activation functions onspeech recognition system. In this work, a hidden Markov model-Deep neural network (HMM-DNN) based speech recognitionis used, where deep neural networks with different activationfunctions have been employed to obtain the emission probabilitiesof hidden Markov model. In this work, two datasets i.e., TIMITand WSJ are employed to study the behavior of various speechrecognition systems with different sized datasets. During thestudy, it is observed that the performance of ReLU-networksis superior compared to the other networks for the smaller sizeddataset (i.e., TIMIT dataset). For the datasets of sufficientlylarger size (i.e., WSJ) performance of ELU-networks is superiorto the other networks.

Abstract

Automatic speaker verification systems are more vulnerable to spoofing attacks. Recently, various countermeasures have been developed for detecting high technology attacks such as speech synthesis and voice conversion. However, there is a wide gap in dealing with replay attacks. In this paper, we propose a new feature for replay attack detection based on single frequency filtering (SFF), which provides high temporal and spectral resolution at each instant. Single frequency filtering cepstral coefficients (SFFCC) with Gaussian mixture model classifier is used for the experimentation on the standard BTAS-2016 corpus. The previously reported best result, which is based on constant Q cepstral coefficients (CQCC) has achieved a half total error rate of 0.67% on this data-set. Our proposed method outperforms the state of the art (CQCC) with a half total error rate of 0.0002%.

Abstract

The ASVspoof 2017 challenge is about the detection of replayed speech from human speech. The proposed system makes use of the fact that when the speech signals are replayed, they pass through multiple channels as opposed to original recordings. This channel information is typically embedded in low signal to noise ratio regions. A speech signal processing method with high spectro-temporal resolution is required to extract robust features from such regions. The single frequency filtering (SFF) is one such technique, which we propose to use for replay attack detection. While SFF based feature representation was used at front-end, Gaussian mixture model and bi-directional long short-term memory models are investigated at the backend as classifiers. The experimental results on ASVspoof 2017 dataset reveal that, SFF based representation is very effective in detecting replay attacks. The score level fusion of back end classifiers further improved the performance of the system which indicates that both classifiers capture complimentary information

Abstract

Language identification (LID) is vital frontend for spoken dialogue systems operating in diverse linguistic settings to reduce recognition and understanding errors. Existing LID systems which use low-level signal information for classification do not scale well due to exponential growth of parameters as the classes increase. They also suffer performance degradation due to the inherent variabilities of speech signal. In the proposed approach, we model the language-specific phonotactic information in speech using recurrent neural network for developing an LID system. The input speech signal is tokenized to phone sequences by using a common language-independent phone recognizer with varying phonetic coverage. We establish a causal relationship between phonetic coverage and LID performance. The phonotactics in the observed phone sequences are modeled using statistical and recurrent neural network language models to predict language-specific symbol from a universal phonetic inventory. Proposed approach is robust, computationally light weight and highly scalable. Experiments show that the convex combination of statistical and recurrent neural network language model (RNNLM) based phonotactic models significantly outperform a strong baseline system of Deep Neural Network (DNN) which is shown to surpass the performance of i-vector based approach for LID. The proposed approach outperforms the baseline models in terms of mean F1 score over 176 languages. Further we provide significant information-theoretic evidence to analyze the mechanism of the proposed approach.

Abstract

Recent improvement in usage of digital media has led people to share their opinions about specific entity through audio. In this paper, an approach to detect the sentiment of an online spoken reviews based on relative prosody features is presented. Most of the existing systems for audio based sentiment analysis use conventional audio features, but they are not problem specific features to extract the sentiment. In this work, relative prosody features are extracted from normal and stressed regions of audio signal to detect the sentiment. Stressed regions are identified using the strength of excitation. Support Vector Machine (SVM) and Gaussian Mixture Model (GMM) classifiers are used to build the sentiment models. MOUD database is used for the proposed study. Experimental results show that, the rate of detecting the sentiment is improved with relative prosody features compared with the prosody and Mel Frequency Cepstral Coefficients (MFCC) because the relative prosody features has more sentiment specific discrimination compared to prosody features.