Hierarchical subband linear predictive cepstral (HSLPC) features for HMM-based speech recognition

Singer identification is one of the important aspects of music information retrieval (MIR). In this work, traditional feature-based and trending convolutional neural network (CNN) based approaches are considered and compared for identifying singers. Two different datasets, namely artist20 and the Indian popular singers’ database with 20 singers are used in this work to evaluate proposed approaches. Cepstral features such as Mel-frequency cepstral coefficients (MFCCs) and linear prediction cepstral coefficients (LPCCs) are considered to represent timbre information. Shifted delta cepstral (SDC) features are also computed beside the cepstral coefficients to capture temporal information. In addition, chroma features are computed from 12 semitones of a musical octave, overall forming a 46-dimensional feature vector. Experiments are conducted with different feature combinations, and suitable features are selected using the genetic algorithm-based feature selection (GAFS) approach. Two different classification techniques, namely artificial neural networks (ANNs) and random forest (RF), are considered on the features mentioned above. Further, spectrograms and chromagrams of audio clips are directly fed to CNN for classification. The singer identification results obtained using CNNs seem to be better than the traditional isolated and ensemble classifiers. Average accuracy of around 75% is observed with CNN in the case of Indian popular singers database. Whereas, on artist20 dataset, the proposed configuration of feature-based approach and CNN could not give better than 60% accuracy.

Because of existence of pseudo-periodic excitation impulses, the conventional linear prediction (LP) analysis is often not accurate in estimating the vocal tract parameters of voiced speech. In order to overcome this, we propose a scheme for high quality linear prediction analysis of speech, it contains two main stages. First, the glottal closure instant (GCI) is detected using a wavelet transform of the speech signals, the interval of adjacent two GCIs is a pitch period, the LP analysis can be executed pitch-synchronously on one or several pitch periods. Second, the excitation impulses are much stronger around the GCI than any other time in a pitch period. We select those samples which are far from the GCI and LP analysis is executed on these selected samples rather than all samples in an analysis frame. The prediction errors of the selected samples become very small. Thus, the accuracy of estimating the vocal tract parameters is improved significantly. The validity of the proposed method is confirmed by comparing the results obtained from the proposed and conventional LP analysis methods.

Onset detection is an important first step in music analysis. We propose a pre-processing scheme for improved onset detection of complex strokes of Indian percussion instruments with resonance characteristics. In this work, we explore the onset detection of Tabla (Indian percussion instrument) strokes. The resonance characteristics of tabla strokes poses challenges to onset detection. In such cases, the energy-based and spectral flux-based onset detectors are often inaccurate on the raw signal. We propose an onset detection algorithm addressing these challenges using Linear Prediction (LP) analysis and Hilbert envelope (HE) in tandem. Tabla signal is modeled using LP, and its residual highlights the onset time instances very well. Unipolar nature of HE on top of LP residual further enhances the onset instances. Onset detection is performed using energy based, spectral flux based and the state of the art Machine Learning based onset detectors on the Hilbert envelope of LP residual (HELP). Experiments were performed on tabla solo played at various tempi and the results show that the HELP based approach provides 12% relative improvement in F-measures compared to the performance on raw tabla signal.

Many features can be used to describe a human face but few have been used in combination. Extending the feature vector using orthogonal sets of measurements can reduce the variance of a matching measure, to improve discrimination capability. This paper investigates how different features can be used for discrimination, alone or when integrated into an extended feature vector. This study concentrates on improving feature definition and extraction from a frontal view image, incorporating and extending established measurements. These form an extended feature vector based on four feature sets: geometric (distance) measurements, the eye region, the outline contour, and the profile. The profile, contour, and eye region are described by the Walsh power spectrum, normalized Fourier descriptors, and normalized moments, respectively. Although there is some correlation between the geometrical measures and the other sets, their bases (distance, shape description, sequency, and statistics) are orthogonal and hence appropriate for this research. A database of face images was analyzed using two matching measures which were developed to control differently the contributions of elements of the feature sets. The match was evaluated for both measures for the separate feature sets and for the extended feature vector. Results demonstrated that no feature set alone was sufficient for recognition whereas the extended feature vector could discriminate between subjects successfully.

Acoustic classification of Australian frogs based on enhanced features and machine learning algorithms

Speaker recognition is an Artificial Intelligent (AI) technology that lets the machine to process, interpret and respond to human language. In this work, the recorded speech developed from a collection of audio speeches is used as a database. Mel-frequency cepstral coefficients (MFCC) and gammatone frequency cepstral coefficients (GFCC) are two different cepstral features used in this work. These extracted features are then used to train, validate and test the classifier. Support Vector Machine (SVM) is the classifier used in developing the speaker identification system. This classifier is trained to classify the input speech into one of the ethnicity classes: Malay, Chinese, Indian or Bumiputera. The results are based on the two different usages of cepstral features from the same speech utterances by speakers. Finally, the comparative analysis of the speaker identification system is made concerning features and classifier. The results revealed that a combination of GFCC and pitch as the feature vectors (Model 4) produced the highest accuracy rate of 86.1%.

In this work, a Recurrent Neural Network (RNN) is trained using cepstral features and a set of difference cepstral feature (DCF) vectors on a frame to frame basis. The DCF vector is formulated to capture the temporal patterns of fricative sounds or phonemes of Assamese language. A hybrid algorithm is developed to recognize these fricative phonemes from certain words containing them. To preserve the temporal information of the speech segment, we here consider a frame-based hybrid approach to recognize fricatives from Assamese speech. A hybrid feature set is developed where simple frame-based feature is combined with differential frame-based feature. Investigation of feature extraction techniques like Linear Predictive Cepstral Coefficient (LPCC) and Mel-Frequency Cepstral Coefficient (MFCC) have been carried out and their performances have been evaluated in comparison to that obtained from the DCF set for Assamese fricative phoneme recognition. Here, speech segment is divided into 20 millisecond frames with overlap of 10 millisecond to extract features. Also, difference of a current frame with its preceding and succeeding frame is considered for forming a more accurate dynamic approach for fricative recognition. The differential processing enables to reduce correlation and retain only the most relevant portion of the input. After obtaining the feature vectors, Self Organizing Map (SOM) has been used to categorize the related features into different classes and remove repeating data. Thus the features obtained from the phoneme signal has been reduced into different sized cluster centres provided by SOM. The reduced feature vector is next applied to the RNN based hybrid classifier for learning the pattern and recognizing any unknown fricative segment.

Spectral representation of speech is complete when both the Fourier transform magnitude and phase spectra are specified. In conventional speech recognition systems, features are generally derived from the short-time magnitude spectrum. Although the importance of Fourier transform phase in speech perception has been realized, few attempts have been made to extract features from it. This is primarily because the resonances of the speech signal which manifest as transitions in the phase spectrum are completely masked by the wrapping of the phase spectrum. Hence, an alternative to processing the Fourier transform phase, for extracting speech features, is to process the group delay function which can be directly computed from the speech signal. The group delay function has been used in earlier efforts, to extract pitch and formant information from the speech signal. In all these efforts, no attempt was made to extract features from the speech signal and use them for speech recognition applications. This is primarily because the group delay function fails to capture the short-time spectral structure of speech owing to zeros that are close to the unit circle in the z-plane and also due to pitch periodicity effects. In this paper, the group delay function is modified to overcome these effects. Cepstral features are extracted from the modified group delay function and are called the modified group delay feature (MODGDF). The MODGDF is used for three speech recognition tasks namely, speaker, language, and continuous-speech recognition. Based on the results of feature and performance evaluation, the significance of the MODGDF as a new feature for speech recognition is discussed

Multi-resolution sub-band features and models for HMM-based phonetic modelling

Multimodal biometric template protection (BTP) is gaining increasing attention as it overcomes the drawback of unimodal BTP. In this work, a token-less cancellable biometric scheme, namely Extended Feature Vector (EFV) Hashing, is applied to fuse the multimodal binary face and fingerprint template at a feature-level fusion. In advance to the unimodal EFV hashing, this work reorganizes the feature extension mechanism in EFV transformation and introduces the random sampling mechanism to increase the difficulty for reverse processing of the cancellable template. In addition, two fusion options are developed for producing the cancellable template. Experiments are conducted on the well-known Fingerprint FVC2004 and Face LFW datasets. The results demonstrate that the propose approach achieves a satisfactory verification rate with EER ~~ 0.1%.

Effective and efficient access to multiparty meeting recordings requires techniques for meeting analysis and indexing. Since meeting participants are generally stationary, speaker location information may be used to identify meeting events e.g., detect speaker changes. Time-delay estimation (TDE) utilizing cross-correlation of multichannel speech recordings is a common approach for deriving speech source location information. Research improved TDE by calculating TDE from linear prediction (LP) residual signals obtained from LP analysis on each individual speech channel. This paper investigates the use of LP residuals for speech TDE, where the residuals are obtained from jointly modeling the multiple speech channels. Experiments conducted with a simulated reverberant room and real room recordings show that jointly modeled LP better predicts the LP coefficients, compared to LP applied to individual channels. Both the individually and jointly modeled LP exhibit similar TDE performance, and outperform TDE on the speech alone, especially with the real recordings.

A Neurocomputing Approach for Monitoring Plinian Volcanic Eruptions Using Infrasound

The accuracy of spectrographic techniques and of linear prediction analysis in measuring formant frequencies was compared. The first three-formant frequencies of ninety synthetic speech tokens were measured by three experienced spectrographic readers and by linear prediction analysis. The synthetic speech tokens were made by parallel synthesis and were chosen to represent a wide range of formant bandwidths, frequencies, and fundamental frequencies. For fundamental frequencies between 100 and 350 Hz, both methods are accurate to within approximately ±60 Hz for both first and second formants; the third formant can be measured with the same degree of accuracy by linear prediction, but only to within ±120 Hz by spectrographic means. For linear prediction analysis, accuracy does not decrease with increasing fundamental frequency, although the accuracy of both methods decreases greatly when fundamental frequency is 400 Hz or greater. These limits of measurement appear to be within the vicinity of difference limens for formant frequencies. [Supported by NINCDS Grant NS03856.]

We propose a new feature vector consisting of mel-frequency discrete wavelet coefficients (MFDWC). The MFDWC are obtained by applying the discrete wavelet transform (DWT) to the mel-scaled log filterbank energies of a speech frame. The purpose of using the DWT is to benefit from its localization property in the time and frequency domains. MFDWC are similar to subband-based (SUB) features and multi-resolution (MULT) features in that both attempt to achieve good time and frequency localization. However, MFDWC have better time/frequency localization than SUB features and MULT features. We evaluated the performance of new features for clean speech and noisy speech and compared the performance of MFDWC with mel-frequency cepstral coefficients (MFCC), SUB features and MULT features. Experimental results on a phoneme recognition task showed that a MFDWC-based recognizer gave better results than recognizers based on MFCC, SUB features, and MULT features for white Gaussian noise, band-limited white Gaussian noise and clean speech cases.

Presents a method to evaluate the dynamic behaviour of the vocal tract shape. Solving the linear prediction model by using Levinson Durbin's algorithm the dynamic behaviour of the vocal tract shape can be obtained only frame by frame. A more realistic approach can be possible by estimating the shape of the vocal tract point by point using an adaptive filter. In the present work the authors have proposed a variable step adaptive linear prediction scheme for reconstruction of the vocal tract. This results in faster convergence. This method can be used for real time estimation of the vocal tract shape for diagnostic and therapeutic purposes.