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

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

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

Designing various learning strategies has been gaining a lot of scientific interest during the recent progress of deep learning methodologies. Curriculum learning is a learning strategy aimed at training the neural network model by presenting the samples in a specific meaningful order rather than randomly sampling the training examples from the data distribution. In this work, we have explored starting small paradigm of curriculum learning technique for speech recognition. The starting small paradigm of curriculum learning is performed by a two step learning strategy. Training dataset is re-organized as a set of easily classifiable examples followed by the actual training dataset and the model is trained on the re-organized dataset. We hypothesize that by following the starting small learning paradigm the learning gets initialized in a better way and progresses to attain a better convergence. We propose to rank the toughness of the training example based on the posterior probabilities obtained using a previously trained model. Apart from re-arranging the training corpus starting small paradigm of curriculum learning is applied at model level. We consider the broad manner-class classification objective function as the smoother version of the phone class classification objective function. The model initially trained for broad class classification is later adapted for phone classification.In this work, we have used TIMIT and a subset of WallStreet Journal (WSJ) corpus to validate the experiments, both the learning strategies have shown consistently better performances across the two datasets compared to the baseline system trained by randomly sampling the dataset.

Abstract

Source and system features are extensively used in building Speaker Recognition (SR) systems. In this paper, we investigate the influence of source and system features on the performance of the SR system in emotional conditions. The Linear Prediction Residual Cepstral Coefficients (LPRCC) which corresponds to source features, and Mel Frequency Cepstral Coefficients (MFCC) and Linear Prediction Cepstral Coefficients (LPCC) that correspond to system features are used for modeling SR system. A maximum-likelihood classifier based on Gaussian mixture density functions is used and experiments are carried out on 3 standard emotional speech databases (Indian Institute of Technology-Simulated Emotion speech corpus (IITKGP-SESC): Hindi, IITKGP-SESC: Telugu and German Emotional Speech Database (EMO-DB)). The speaker models are trained with neutral utterances and tested with 3 types (anger, happy and sad) of emotional speech utterances. Based on experimental results, the performance degradation was observed in mismatched (trained with one emotion and tested with other emotion) case and the average percentage of degradation in SR task using source features is approximately 16% more compared to system features. The performance degradation of SR system is nullified when trained and tested with the same emotion.

Abstract

The present study focuses on the text-independent speaker recognition in emotional conditions. In this paper, both system and source features are considered to represent speaker specific information. At the model level, Gaussian Mixture Models (GMMs), Gaussian Mixture Model-Universal Background Model (GMM-UBM) and Deep Neural Networks (DNN) are explored. The experiments are performed using 3 emotional databases, i.e. German emotional speech database (EMO-DB), IITKGP-SESC: Hindi and IITKGP-SESC: Telugu databases. The emotions considered in the present study are neutral, anger, happy and sad. The results show that, the performance of a speaker recognition system trained with clean speech is degrading while testing with emotional data irrespective of feature used or model used to build the system. The best results are obtained for the score level fusion of system and source features based systems when speakers are modeled with DNNs.

Abstract

The main focus of this study is to analyze the performance of Automatic Speech Recognition (ASR) in different emotional environments using prosody modification. The majority of ASR systems are trained using neutral speech and the performance of such systems degrade when tested with the emotional speech. In this paper, the various components of speech that contribute to the emotion characteristics are studied. The prosody features of the source emotional utterancesare modified according to the target neutral utterances using Flexible Analysis Synthesis Tool (FAST). In the FAST, Dynamic Time Warping (DTW) is used to align the source emotional and target neutral utterances. Components of the prosody such as intonation, duration and excitation source are manipulated to incorporate the desired features into the source utterance. The modified (source emotional) utterances are then used for testing the ASR system which is trained using neutral speech. In this study, three emotions (compassion, happiness and anger) are considered for the analysis. Experimental results indicate an average improvement in the speech recognition system performance by considering prosody modified speech.

Abstract

Automatic detection of shout prosody in continuous speech signal involves examining changes in its production characteristics. Our recent study of electroglottograph signals highlighted that significant changes occur in the glottal excitation source characteristics during production of shouted speech, especially in the vowel contexts. But the differences between normal and shouted speech, in the production features derived over utterances or word segments, may be masked sometimes by pauses or unvoiced regions related variations. Also, for such a real-time system, these vowel regions need to be found automatically. In this paper, changes in the shout production features are examined in the automatically detected vowel regions. Production of a vowel involves periodic impulse-like excitation and relatively high signal energy. Hence, the knowledge of epochs using zero-frequency filtering, and accurate vowel onset points can be used for detecting these regions. Changes in two excitation source features, the instantaneous fundamental frequency and strength of excitation, and in a vocal tract filter feature the dominant frequency, are examined for five steady vowel regions. Larger changes in these distinguishing features are observed in the automatically found vowel regions, than in word segments. This approach can help improving the systems for automatic detection of shout regions in continuous speech, and in paralinguistic applications that involve detection of prosody or emotions

Abstract

Shouted speech signals have been studied mostlyfor utterances or word segments. In the production features derived over these segments, the differences between normal and shouted speech may sometimes be masked by variations due to pauses and unvoiced regions. Also, our recent study of electroglottograph signals has highlighted the usefulness of examining changes in the glottal excitation source characteristics during production of shouted speech. In this paper we examine changes in the shout production features in the vowel regions,that are found automatically by using the knowledge of accurate vowel onset points and epochs. Vowel regions are produced by periodic impulse-like excitation and contain relatively high signal energy. Hence the source features the instantaneous fundamental frequency and strength of excitation are examined, along with the filter feature the dominant frequency, for five different steady vowel regions. Changes in these features, that distinguish between normal and shouted speech, are more prominent in the automatically found vowel regions than for the entire word segments. This insight can help improving the automatic shout detection and other paralinguistic applications where acoustic cues need to be found in the vowel regions, e.g., detection of emotion categories.

Abstract

The objective of this work is to develop a rule-based emotion conversion method for a better emotional perception. In this work, performance of emotion conversion using the linear modification model is improved by using vowel-based non-uniform prosody modification. In the present approach, attempts were made to integrate features like position and identity for addressing the non-uniformity in prosody generated due to the emotional state of the speaker. We mainly concentrate on the parameters such as strength, duration and pitch contour of vowels at different parts of the sentence. The influence of emotions on the above parameters is exploited to convert the speech from neutral emotion to the target emotion. Non-uniform prosody modification factors for emotion conversion are based on the position of vowels in the word, and the position of the word in the sentence. This study is carried out by using Indian Institute of Technology-Simulated Emotion speech corpus. Evaluation of the proposed algorithm is carried out by a subjective listening test. From the listening tests, it is observed that the performance of the proposed approach is better than the existing approaches