Combining evidences from excitation source and vocal tract system features for Indian language identification using deep neural networks

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-of-the-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.     

Characterization and recognition of emotions from speech using excitation source information

This paper explores the excitation source features of speech production mechanism for characterizing and recognizing the emotions from speech signal. The excitation source signal is obtained from speech signal using linear prediction (LP) analysis, and it is also known as LP residual. Glottal volume velocity (GVV) signal is also used to represent excitation source, and it is derived from LP residual signal. Speech signal has high signal to noise ratio around the instants of glottal closure (GC). These instants of glottal closure are also known as epochs. In this paper, the following excitation source features are proposed for characterizing and recognizing the emotions: sequence of LP residual samples and their phase information, parameters of epochs and their dynamics at syllable and utterance levels, samples of GVV signal and its parameters. Auto-associative neural networks (AANN) and support vector machines (SVM) are used for developing the emotion recognition models. Telugu and Berlin emotion speech corpora are used to evaluate the developed models. Anger, disgust, fear, happy, neutral and sadness are the six emotions considered in this study. About 42 % to 63 % of average emotion recognition performance is observed using different excitation source features. Further, the combination of excitation source and spectral features has shown to improve the emotion recognition performance up to 84 %.     

Spotting glottal stop in Amharic in continuous speech

Glottal stop sounds in Amharic are produced due to abrupt closure of the glottis without any significant gesture in the accompanying articulatory organs in the vocal tract system. It is difficult to observe the features of the glottal stop through spectral analysis, as the spectral features emphasize mostly the features of the vocal tract system. In order to spot the glottal stop sounds in continuous speech, it is necessary to extract the features of the source of excitation also, which may require some non-spectral methods for analysis. In this paper the linear prediction (LP) residual is used as an approximation to the excitation source signal, and the excitation features are extracted from the LP residual using zero frequency filtering (ZFF). The glottal closure instants (GCIs) or epoch are identified from the ZFF signal. At each GCI, the cross-correlation coefficients of successive glottal cycles of the LP residual, the normalized jitter and the logarithm of the peak normalized excitation strength (LPNES) are calculated. Further, the parameters of Gaussian approximation models are derived from the distributions of the excitation parameters. These model parameters are used to identify the regions of the glottal stop sounds in continuous speech. For the database used in this study 92.89% of the glottal stop regions are identified correctly, with 8.50% false indications.     

GMM based language identification system using robust features

In this work, we have proposed new feature vectors for spoken language identification (LID) system. The Mel frequency cepstral coefficients (MFCC) and formant frequencies derived using short-time window speech signal. Formant frequencies are extracted from linear prediction (LP) analysis of speech signal. Using these two kind of features of speech signal, new feature vectors are derived using cluster based computation. A GMM based classifier has been designed using these new feature vectors. The language specific apriori knowledge is applied on the recognition output. The experiments are carried out on OGI database and LID recognition performance is improved.     

A hierarchical language identification system for Indian languages

Automatic spoken Language IDentification (LID) is the task of identifying the language from a short duration of speech signal uttered by an unknown speaker. In this work, an attempt has been made to develop a two level language identification system for Indian languages using acoustic features. In the first level, the system identifies the family of the spoken language, and then it is fed to the second level which aims at identifying the particular language in the corresponding family. The performance of the system is analyzed for various acoustic features and different classifiers. The suitable acoustic feature and the pattern classification model are suggested for effective identification of Indian languages. The system has been modeled using hidden Markov model (HMM), Gaussian mixture model (GMM) and artificial neural networks (ANN). We studied the discriminative power of the system for the features mel frequency cepstral coefficients (MFCC), MFCC with delta and acceleration coefficients and shifted delta cepstral (SDC) coefficients. Then the LID performance as a function of the different training and testing set sizes has been studied. To carry out the experiments, a new database has been created for 9 Indian languages. It is shown that GMM based LID system using MFCC with delta and acceleration coefficients is performing well with 80.56% accuracy. The performance of GMM based LID system with SDC is also considerable.     

Linear prediction residual features for automatic speaker verification anti-spoofing

Automatic speaker verification (ASV) systems are highly vulnerable against spoofing attacks. Anti-spoofing, determining whether a speech signal is natural/genuine or spoofed, is very important for improving the reliability of the ASV systems. Spoofing attacks using the speech signals generated using speech synthesis and voice conversion have recently received great interest due to the 2015 edition of Automatic Speaker Verification Spoofing and Countermeasures Challenge (ASVspoof 2015). In this paper, we propose to use linear prediction (LP) residual based features for anti-spoofing. Three different features extracted from LP residual signal were compared using the ASVspoof 2015 database. Experimental results indicate that LP residual phase cepstral coefficients (LPRPC) and LP residual Hilbert envelope cepstral coefficients (LPRHEC) obtained from the analytic signal of the LP residual yield promising results for anti-spoofing. The proposed features are found to outperform standard Mel-frequency cepstral coefficients (MFCC) and Cosine Phase (CosPhase) features. LPRPC and LPRHEC features give the smallest equal error rates (EER) for eight spoofing methods out of ten spoofing attacks in comparison to MFCC and CosPhase features.     

Voice conversion using General Regression Neural Network

The objective of voice conversion system is to formulate the mapping function which can transform the source speaker characteristics to that of the target speaker. In this paper, we propose the General Regression Neural Network (GRNN) based model for voice conversion. It is a single pass learning network that makes the training procedure fast and comparatively less time consuming. The proposed system uses the shape of the vocal tract, the shape of the glottal pulse (excitation signal) and long term prosodic features to carry out the voice conversion task. In this paper, the shape of the vocal tract and the shape of source excitation of a particular speaker are represented using Line Spectral Frequencies (LSFs) and Linear Prediction (LP) residual respectively. GRNN is used to obtain the mapping function between the source and target speakers. The direct transformation of the time domain residual using Artificial Neural Network (ANN) causes phase change and generates artifacts in consecutive frames. In order to alleviate it, wavelet packet decomposed coefficients are used to characterize the excitation of the speech signal. The long term prosodic parameters namely, pitch contour (intonation) and the energy profile of the test signal are also modified in relation to that of the target (desired) speaker using the baseline method. The relative performances of the proposed model are compared to voice conversion system based on the state of the art RBF and GMM models using objective and subjective evaluation measures. The evaluation measures show that the proposed GRNN based voice conversion system performs slightly better than the state of the art models.     

Improved automatic detection of creak

This paper describes a new algorithm for automatically detecting creak in speech signals. Detection is made by utilising two new acoustic parameters which are designed to characterise creaky excitations following previous evidence in the literature combined with new insights from observations in the current work. In particular the new method focuses on features in the Linear Prediction (LP) residual signal including the presence of secondary peaks as well as prominent impulse-like excitation peaks. These parameters are used as input features to a decision tree classifier for identifying creaky regions. The algorithm was evaluated on a range of read and conversational speech databases and was shown to clearly outperform the state-of-the-art. Further experiments involving degradations of the speech signal demonstrated robustness to both white and babble noise, providing better results than the state-of-the-art down to at least 20 dB signal to noise ratio.     

A GMM-BASED HIERARCHICAL AUTOMATIC LANGUAGE IDENTIFICATION SYSTEM FOR INDIAN LANGUAGES

Automatic spoken language identification (LID) is the task of identifying a language from a short utterance of speech by an unknown speaker. This article describes a novel two-level identification system for Indian languages using acoustic features. In the first level, the system identifies the family of the spoken language; the second level aims at identifying the particular language within the corresponding family. The proposed system has been modeled using Gaussian mixture model (GMM) and utilizes the following acoustic features: mel frequency cepstral coefficients (MFCC) and shifted delta cepstrum (SDC). A new database has been created for nine Indian languages. It is shown that a GMM-based LID system using MFCC with delta and acceleration coefficients is performing well, with 80.56% accuracy. The performance accuracy of the GMM-based LID system with SDC is also considerable.     

A pitch synchronous approach to design voice conversion system using source-filter correlation

We propose a pitch synchronous approach to design the voice conversion system taking into account the correlation between the excitation signal and vocal tract system characteristics of speech production mechanism. The glottal closure instants (GCIs) also known as epochs are used as anchor points for analysis and synthesis of the speech signal. The Gaussian mixture model (GMM) is considered to be the state-of-art method for vocal tract modification in a voice conversion framework. However, the GMM based models generate overly-smooth utterances and need to be tuned according to the amount of available training data. In this paper, we propose the support vector machine multi-regressor (M-SVR) based model that requires less tuning parameters to capture a mapping function between the vocal tract characteristics of the source and the target speaker. The prosodic features are modified using epoch based method and compared with the baseline pitch synchronous overlap and add (PSOLA) based method for pitch and time scale modification. The linear prediction residual (LP residual) signal corresponding to each frame of the converted vocal tract transfer function is selected from the target residual codebook using a modified cost function. The cost function is calculated based on mapped vocal tract transfer function and its dynamics along with minimum residual phase, pitch period and energy differences with the codebook entries. The LP residual signal corresponding to the target speaker is generated by concatenating the selected frame and its previous frame so as to retain the maximum information around the GCIs. The proposed system is also tested using GMM based model for vocal tract modification. The average mean opinion score (MOS) and ABX test results are 3.95 and 85 for GMM based system and 3.98 and 86 for the M-SVR based system respectively. The subjective and objective evaluation results suggest that the proposed M-SVR based model for vocal tract modification combined with modified residual selection and epoch based model for prosody modification can provide a good quality synthesized target output. The results also suggest that the proposed integrated system performs slightly better than the GMM based baseline system designed using either epoch based or PSOLA based model for prosody modification.     

Language identification using phase information

The present work investigates the importance of phase in language identification (LID). We have proposed three phase based features for the language recognition task. In this work, auto-regressive model with scale factor error augmentation have been used for better representation of phase based features. We have developed three group delay based systems, namely, normal group delay based system, auto-regressive model group delay based system and auto-regressive group delay with scale factor augmentation based system. As mel-frequency cepstral coefficients (MFCCs) are extracted from the magnitude of the Fourier transform, we have combined this MFCC-based system with our phase-based systems to exploit the complete information contained in a speech signal. In this work, we have used IITKGP-MLILSC speech database and OGI Multi-language Telephone Speech (OGI-MLTS) corpus for our experiments. We have used Gaussian mixture models for building the language models. From the experimental results it is observed that the LID accuracy obtained from our proposed phase based features is comparable with MFCC features. We have also observed some performance improvement in the LID accuracy on combining the proposed phase-based systems with the state of the art MFCC-based system.     

Robust Speech Feature Extraction by Growth Transformation in Reproducing Kernel Hilbert Space

The performance of speech recognition systems depends on consistent quality of the speech features across variable environmental conditions encountered during training and evaluation. This paper presents a kernel-based nonlinear predictive coding procedure that yields speech features which are robust to nonstationary noise contaminating the speech signal. Features maximally insensitive to additive noise are obtained by growth transformation of regression functions that span a reproducing kernel Hilbert space (RKHS). The features are normalized by construction and extract information pertaining to higher-order statistical correlations in the speech signal. Experiments with the TI-DIGIT database demonstrate consistent robustness to noise of varying statistics, yielding significant improvements in digit recognition accuracy over identical models trained using Mel-scale cepstral features and evaluated at noise levels between 0 and 30-dB signal-to-noise ratio.     

Speaker verification using excitation source information

In this work we develop a speaker recognition system based on the excitation source information and demonstrate its significance by comparing with the vocal tract information based system. The speaker-specific excitation information is extracted by the subsegmental, segmental and suprasegmental processing of the LP residual. The speaker-specific information from each level is modeled independently using Gaussian mixture modeling—universal background model (GMM-UBM) modeling and then combined at the score level. The significance of the proposed speaker recognition system is demonstrated by conducting speaker verification experiments on the NIST-03 database. Two different tests, namely, Clean test and Noisy test are conducted. In case of Clean test, the test speech signal is used as it is for verification. In case of Noisy test, the test speech is corrupted by factory noise (9 dB) and then used for verification. Even though for Clean test case, the proposed source based speaker recognition system still provides relatively poor performance than the vocal tract information, its performance is better for Noisy test case. Finally, for both clean and noisy cases, by providing different and robust speaker-specific evidences, the proposed system helps the vocal tract system to further improve the overall performance.     

Prosody modification using instants of significant excitation

Prosody modification involves changing the pitch and duration of speech without affecting the message and naturalness. This paper proposes a method for prosody (pitch and duration) modification using the instants of significant excitation of the vocal tract system during the production of speech. The instants of significant excitation correspond to the instants of glottal closure (epochs) in the case of voiced speech, and to some random excitations like onset of burst in the case of nonvoiced speech. Instants of significant excitation are computed from the linear prediction (LP) residual of speech signals by using the property of average group-delay of minimum phase signals. The modification of pitch and duration is achieved by manipulating the LP residual with the help of the knowledge of the instants of significant excitation. The modified residual is used to excite the time-varying filter, whose parameters are derived from the original speech signal. Perceptual quality of the synthesized speech is good and is without any significant distortion. The proposed method is evaluated using waveforms, spectrograms, and listening tests. The performance of the method is compared with linear prediction pitch synchronous overlap and add (LP-PSOLA) method, which is another method for prosody manipulation based on the modification of the LP residual. The original and the synthesized speech signals obtained by the proposed method and by the LP-PSOLA method are available for listening at http://speech.cs.iitm.ernet.in/Main/result/prosody.html.     

一种基于正弦激励的线性预测模型的语音转换方法

在正弦激励模型的线性预测(LP)残差转换的基础上,提出了一种改进语音特征转换性能的语音转换方法.基于线性预测分析和综合的构架,该方法一方面通过谱包络估计声码器提取源说话人的线性预测编码(LPC)倒谱包络,并使用双线性变换函数实现倒谱包络的转换;另一方面由谐波正弦模型对线性预测残差信号建模和分解,采用基音频率变换将源说话人的残差信号转换为近似目标说话人的残差信号.最后由修正后的残差信号激励时变滤波器得到转换语音,滤波器参数通过转换得到的LPC倒谱包络实时更新.实验结果表明,该方法在主观和客观测试中都具有良好的结果,能有效地转换说话人声音特征,获得高相似度的转换语音.     

Articulatory and excitation source features for speech recognition in read,extempore and conversation modes

In our previous works, we have explored articulatory and excitation source features to improve the performance of phone recognition systems (PRSs) using read speech corpora. In this work, we have extended the use of articulatory and excitation source features for developing PRSs of extempore and conversation modes of speech, in addition to the read speech. It is well known that the overall performance of speech recognition system heavily depends on accuracy of phone recognition. Therefore, the objective of this paper is to enhance the accuracy of phone recognition systems using articulatory and excitation source features in addition to conventional spectral features. The articulatory features (AFs) are derived from the spectral features using feedforward neural networks (FFNNs). We have considered five AF groups, namely: manner, place, roundness, frontness and height. Five different AF-based tandem PRSs are developed using the combination of Mel frequency cepstral coefficients (MFCCs) and AFs derived from FFNNs. Hybrid PRSs are developed by combining the evidences from AF-based tandem PRSs using weighted combination approach. The excitation source information is derived by processing the linear prediction residual of the speech signal. The vocal tract information is captured using MFCCs. The combination of vocal tract and excitation source features is used for developing PRSs. The PRSs are developed using hidden Markov models. Bengali speech database is used for developing PRSs of read, extempore and conversation modes of speech. The results are analyzed and the performance is compared across different modes of speech. From the results, it is observed that the use of either articulatory or excitation source features along-with to MFCCs will improve the performance of PRSs in all three modes of speech. The improvement in the performance using AFs is much higher compared to the improvement obtained using excitation source features.     

Investigation on LP-residual representations for speaker identification

Feature extraction is an essential and important step for speaker recognition systems. In this paper, we propose to improve these systems by exploiting both conventional features such as mel frequency cepstral coding (MFCC), linear predictive cepstral coding (LPCC) and non-conventional ones. The method exploits information present in the linear predictive (LP) residual signal. The features extracted from the LP-residue are then combined to the MFCC or the LPCC. We investigate two approaches termed as temporal and frequential representations. The first one consists of an auto-regressive (AR) modelling of the signal followed by a cepstral transformation in a similar way to the LPC-LPCC transformation. In order to take into account the non-linear nature of the speech signals we used two estimation methods based on second and third-order statistics. They are, respectively, termed as R-SOS-LPCC (residual plus second-order statistic based estimation of the AR model plus cepstral transformation) and R-HOS-LPCC (higher order). Concerning the frequential approach, we exploit a filter bank method called the power difference of spectra in sub-band (PDSS) which measures the spectral flatness over the sub-bands. The resulting features are named R-PDSS. The analysis of these proposed schemes are done over a speaker identification problem with two different databases. The first one is the Gaudi database and contains 49 speakers. The main interest lies in the controlled acquisition conditions: mismatch between the microphones and the interval sessions. The second database is the well-known NTIMIT corpus with 630 speakers. The performances of the features are confirmed over this larger corpus. In addition, we propose to compare traditional features and residual ones by the fusion of recognizers (feature extractor + classifier). The results show that residual features carry speaker-dependent features and the combination with the LPCC or the MFCC shows global improvements in terms of robustness under different mismatches. A comparison between the residual features under the opinion fusion framework gives us useful information about the potential of both temporal and frequential representations.     

A syllable-scale framework for language identification

Whilst several examples of segment based approaches to language identification (LID) have been published, they have been typically conducted using only a small number of languages, or varying feature sets, thus making it difficult to determine how the segment length influences the accuracy of LID systems. In this study, phone-triplets are used as crude approximates for a syllable-length sub-word segmental unit. The proposed pseudo-syllabic length framework is subsequently used for both qualitative and quantitative examination of the contributions made by acoustic, phonotactic and prosodic information sources, and trialled in accordance with the NIST 1996 LID protocol. Firstly, a series of experimental comparisons are conducted which examine the utility of using segmental units for modelling short term acoustic features. These include comparisons between language specific Gaussian mixture models (GMMs), language specific GMMs for each segmental unit, and finally language specific hidden Markov models (HMM) for each segment, undertaken in an attempt to better model the temporal evolution of acoustic features. In a second tier of experiments, the contribution of both broad and fine class phonotactic information, when considered over an extended time frame, is contrasted with an implementation of the currently popular parallel phone recognition language modelling (PPRLM) technique. Results indicate that this information can be used to complement existing PPRLM systems to obtain improved performance. The pseudo-syllabic framework is also used to model prosodic dynamics and compared to an implemented version of a recently published system, achieving comparable levels of performance.     

Emotion recognition from speech using source, system, and prosodic features

In this work, source, system, and prosodic features of speech are explored for characterizing and classifying the underlying emotions. Different speech features contribute in different ways to express the emotions, due to their complementary nature. Linear prediction residual samples chosen around glottal closure regions, and glottal pulse parameters are used to represent excitation source information. Linear prediction cepstral coefficients extracted through simple block processing and pitch synchronous analysis represent the vocal tract information. Global and local prosodic features extracted from gross statistics and temporal dynamics of the sequence of duration, pitch, and energy values represent the prosodic information. Emotion recognition models are developed using above mentioned features separately, and in combination. Simulated Telugu emotion database (IITKGP-SESC) is used to evaluate the proposed features. The emotion recognition results obtained using IITKGP-SESC are compared with the results of internationally known Berlin emotion speech database (Emo-DB). Autoassociative neural networks, Gaussian mixture models, and support vector machines are used to develop emotion recognition systems with source, system, and prosodic features, respectively. Weighted combination of evidence has been used while combining the performance of systems developed using different features. From the results, it is observed that, each of the proposed speech features has contributed toward emotion recognition. The combination of features improved the emotion recognition performance, indicating the complementary nature of the features.     

Improvement of phone recognition accuracy using speech mode classification

In this work, we have developed a speech mode classification model for improving the performance of phone recognition system (PRS). In this paper, we have explored vocal tract system, excitation source and prosodic features for development of speech mode classification (SMC) model. These features are extracted from voiced regions of a speech signal. In this study, conversation, extempore, and read speech are considered as three different modes of speech. The vocal tract component of speech is extracted using Mel-frequency cepstral coefficients (MFCCs). The excitation source features are captured through Mel power differences of spectrum in sub-bands (MPDSS) and residual Mel-frequency cepstral coefficients (RMFCCs) of the speech signal. The prosody information is extracted from pitch and intensity. Speech mode classification models are developed using above described features independently, and in fusion. The experiments carried out on Bengali speech corpus to analyze the accuracy of the speech mode classification model using the artificial neural network (ANN), naive Bayes, support vector machines (SVMs) and k-nearest neighbor (KNN). We proposed four classification models which are combined using maximum voting approach for optimal performance. From the results, it is observed that speech mode classification model developed using the fusion of vocal tract system, excitation source and prosodic features of speech, yields the best performance of 98%. Finally, the proposed speech mode classifier is integrated to the PRS, and the accuracy of phone recognition system is observed to be improved by 11.08%.