Abstract

In this study, the acoustic-phonetic attributes of palatalization in the Kashmiri speech is investigated. It is a unique phonetic feature of Kashmiri in the Indian context. An automated approach is proposed to detect this unique phonetic feature from the continuous Kashmiri speech. The i-matra vowel has the impact of palatalizing the consonant connected to it. Therefore, these consonants investigated in synchronous with vowel regions, which are spotted using the instantaneous energy computed from the envelope-derivative of the speech signal. The resonating characteristics of the vocal-tract system framework that reflect the formant dynamics are used to differentiate palatalized consonants from the other consonants. In this regard, the Hilbert envelope of the numerator of the group-delay function that provides good time-frequency resolution used to extract formants. The palatalization detection experimentation carried out in various vowel contexts using the acoustic cues, and it produced a promising result with a detection accuracy of 92.46 %.

Abstract

This work focuses on the development of speech data comprising 23 Indian languages for developing language identification (LID) systems. Large data is a pre-requisite for developing state-of-the-art LID systems. With this motivation, the task of developing multilingual speech corpus for Indian languages has been initiated. This paper describes the composition of the data and the performances of various LID systems developed using this data. In this paper, Mel frequency cepstral feature representation is used for language identification. In this work, various state-of-the-art LID systems are developed using i-vectors, deep neural network (DNN) and deep neural network with attention (DNN-WA) models. The performance of the LID system is observed in terms of the equal error rate for i-vector, DNN and DNN-WA is 17.77%, 17.95%, and 15.18% respectively. Deep neural network with attention model shows a better performance over i-vector and DNN models

Abstract

Language identification (LID) systems, which can model highlevel information such as phonotactics have exhibited superior performance. State-of-the-art models use sequential models to capture the high-level information, but these models are sensitive to the length of the utterance and do not equally generalize over variable length utterances. To effectively capture this information, a feature that can model the long-term temporal context is required. This study aims to capture the long-term temporal context by appending successive shifted delta cepstral (SDC) features. Deep neural networks have been explored for developing LID systems. Experiments have been performed using AP17-OLR database. LID systems developed by stacking SDC features have shown significant improvement compared to the system trained with SDC features. The proposed feature with residual connections in the feed-forward networks reduced the equal error rate from 21.04, 18.02, 16.45 to 14.42, 11.14 and 10.11 on the 1-second, 3-seconds and > 3-second test utterances respectively.

Abstract

India being a multilingual society, a multilingual automatic speech recognition system (ASR) is widely appreciated. Despite different orthographies, Indian languages share same phonetic space. To exploit this property, a joint acoustic model has been trained for developing multilingual ASR system using a common phone-set. Three Indian languages namely Telugu, Tamil and, Gujarati are considered for the study. This work studies the amenability of two different acoustic modeling approaches for training a joint acoustic model using common phone-set. Sub-space Gaussian mixture models (SGMM), and recurrent neural networks (RNN) trained with connectionst temporal classification (CTC) objective function are explored for training joint acoustic models. From the experimental results, it can be observed that the joint acoustic models trained with RNN-CTC have performed better than SGMM system even on 120 hours of data (approx 40 hrs per language). The joint acoustic model trained with RNN-CTC has performed better than monolingual models, due to an efficient data sharing across the languages. Conditioning the joint model with language identity had a minimal advantage. Sub-sampling the features by a factor of 2 while training RNN-CTC models has reduced the training times and has performed better. Index Terms: Speech recognition, Joint acoustic model, lowresource, common phone set, Indian languages, RNN-CTC, SGMM

Abstract

The main focus of this paper is towards the development of a large vocabulary Telugu speech database. Telugu is a low resource language where there exists no standardized database for building the speech recognition system (ASR). The database consists of neutral speech samples collected from 100 speakers for building the Telugu ASR system and it was named as IIIT-H Telugu speech corpus. The speech and text corpus design and the procedure followed for the collection of the database have been discussed in detail. The preliminary ASR system results for the models built in this database are reported. The architectural choices of deep neural networks (DNNs) play a crucial role in improving the performance of ASR systems. ASR trained with hybrid DNNs (DNNHMM) with more hidden layers have shown better performance over the conventional GMMs (GMM-HMM). Kaldi tool kit is used for building the acoustic models required for the ASR system.

Abstract

A mismatch in training and operating environments causes a performance degradation in speech recognition systems (ASR). One major reason for this mismatch is due to the presence of expressive (emotive) speech in operational environments. Emotions in speech majorly inflict the changes in the prosody parameters of pitch, duration and energy. This work is aimed at improving the performance of speech recognition systems in the presence of emotive speech. This work focuses on improving the speech recognition performance without disturbing the existing ASR system. The prosody modification of pitch, duration and energy is achieved by tuning the modification factors values for the relative differences between the neutral and emotional data sets. The neutral version of emotive speech is generated using uniform and non-uniform prosody modification methods for speech recognition. During the study, IITKGP-SESC corpus is used for building the ASR system. The speech recognition system for the emotions (anger, happy and compassion) is evaluated. An improvement in the performance of ASR is observed when the prosody modified emotive utterance is used for speech recognition in place of original emotive utterance. An average improvement around 5% in accuracy is observed due to the use of non-uniform prosody modification methods.

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.

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

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

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%.