Feature Extraction For Speech Recognition Research Articles

Feature extraction in speech recognition using linear predictive coding: An overview

Over the past years, advancements in speech processing have mostly been driven by DSP approaches. The speech interface was designed to convert speech input into a parametric form for further processing (Speech-to-Text) and the resulting text output to speech synthesis (Text-to-Speech). Feature extraction is done by changing the speech waveform into a parametric representation at a relatively low data rate so that it can be processed and analyzed later. There are numerous feature extraction techniques available. This paper presents the overview of Linear Predictive Coding (LPC).

Dereverberation of autoregressive envelopes for far-field speech recognition

The task of speech recognition in far-field environments is adversely affected by the reverberant artifacts that elicit as the temporal smearing of the sub-band envelopes. In this paper, we develop a neural model for speech dereverberation using the long-term sub-band envelopes of speech. The sub-band envelopes are derived using frequency domain linear prediction (FDLP) which performs an autoregressive estimation of the Hilbert envelopes. The neural dereverberation model estimates the envelope gain which when applied to reverberant signals suppresses the late reflection components in the far-field signal. The dereverberated envelopes are used for feature extraction in speech recognition. Further, the sequence of steps involved in envelope dereverberation, feature extraction and acoustic modeling for ASR can be implemented as a single neural processing pipeline which allows the joint learning of the dereverberation network and the acoustic model. Several experiments are performed on the REVERB challenge dataset, CHiME-3 dataset and VOiCES dataset. In these experiments, the joint learning of envelope dereverberation and acoustic model yields significant performance improvements over the baseline ASR system based on log-mel spectrogram as well as other past approaches for dereverberation (average relative improvements of 10–24% over the baseline system). A detailed analysis on the choice of hyper-parameters and the cost function involved in envelope dereverberation is also provided.

Automatic Speech Recognition Features Extraction Techniques: A Multi-criteria Comparison

Features extraction is an important step in Automatic Speech Recognition, which consists of determining the audio signal components that are useful for identifying linguistic content while removing background noise and irrelevant information. The main objective of features extraction is to identify the discriminative and robust features in the acoustic data. The derived feature vector should possess the characteristics of low dimensionality, long-time stability, non-sensitivity to noise, and no correlation with other features, which makes the application of a robust feature extraction technique a significant challenge for Automatic Speech Recognition. Many comparative studies have been carried out to compare different speech recognition feature extraction techniques, but none of them have evaluated the criteria to be considered when applying a feature extraction technique. The objective of this work is to answer some of the questions that may arise when considering which feature extraction techniques to apply, through a multi-criteria comparison of different features extraction techniques using the Weighted Scoring Method.

Denoising of Speech Signal using Empirical Mode Decomposition and Kalman Filter

Speech denoising is the process of removing the noise from the noise corrupted speech. The applications of speech denoising are used in speech enhancement, speech recognition and many more. In this work, a new approach is proposed to de-noise the speech which is corrupted from different noises, Empirical mode decomposition and the Kalman filter (EMD-KF) is used for speech denoising in the proposed work. The clean speech is corrupted by the noise with the different SNR’s, and further Empirical mode decomposition (EMD) is applied to the noise corrupted speech later the obtained resultant speech is passed through the Kalman filter (KF) which gives the denoised speech. The result shows that the mean squared error (MSE) values of EMD-KF are extremely less when compared to other methods like discrete wavelet transform (wavelet families like Daubechies and Symlet), empirical mode decomposition (EMD) and moving average filter followed by empirical mode decomposition (MA-EMD). As an application the proposed algorithm is used in the feature extraction for speech recognition. Mel frequency cepstral coefficient (MFCC) is performed on both the original speech and the denoised speech and found majority of the denoised speech features are similar to the original speech features and few denoised speech features are nearby to the original speech features.

Design and Development of Automatic Speech Recognition of Marathi Numerals

Speech is the way of communication of human being in this world. It is need to develop a system which can interface with computer. For speech recognition feature extraction technique and the feature matching technique are used. For the feature extraction, we used MFCC and for feature matching and the classification we used the Support Vector Machine (SVM) technique. We used PCA for dimension reduction. Database collected is in the form of ‘.wav’ file. The data collected is of Marathi numbers starting from Shunya (One) to Nau (Nine). Data is divided into male and female speakers. It is recorded in noisy environment. Feature are extracted and then classified with SVM. This paper is describes the speech recognition of Marathi digits.

On the distribution of Mel-filtered log-spectrum of speech in additive noise

This paper investigates the distribution of Mel-filtered log-spectrum (MFLS) of speech signals in noisy environments. Without employing any prior assumption and using a non-parametric method, we estimate the Joint Probability Density Functions (JPDFs) of MFLS components of clean and noisy speech, and observe that noise components exhibit a normal distribution. Furthermore, clean and noisy speech exhibit a mixture of two normal distributions. The first lobe of this mixture fits to noise and corresponds to the intervals where speech is inactive; while the second one is formed by speech components. This lobe is almost equal (highly correlated) to the PDF of clean speech. Therefore, each MFLS component is accurately modeled by a normal PDF with time-varying parameters in a two-state Markov process. This shows where noise basically affects the distribution of clean speech, contrary to previous works. As an application example, we also present a novel noise-robust improvement to feature extraction for speech recognition by separating the non-speech and speech intervals for each MFLS component. The enhanced features are extracted from the minimum-mean-square-error (MMSE) estimates of MFLS coefficients. Our proposed method, when evaluated on Aurora 2 task for recognition purposes, not only outperforms both the MFCC (Mel Frequency Cepstral Coefficient), as the baseline, and MFCC+CMVN (cepstral mean and variance normalization features), but also helps improve recognition performance when used in conjunction with ETSI-AFE, in different noisy conditions.

Speech Recognition Performance Improvement using Gamma-tone Feature Extraction Acoustic Model

Improve the recognition performance of speech recognition systems as a method for recognizing human listening skills were incorporated into the system. In noisy environments by separating the speech signal and noise, select the desired speech signal. but In terms of practical performance of speech recognition systems are factors. According to recognized environmental changes due to noise speech detection is not accurate and learning model does not match. In this paper, to improve the speech recognition feature extraction using gamma tone and learning model using acoustic model was proposed. The proposed method the feature extraction using auditory scene analysis for human auditory perception was reflected In the process of learning models for recognition. For performance evaluation in noisy environments, -10dB, -5dB noise in the signal was performed to remove 3.12dB, 2.04dB SNR improvement in performance was confirmed.

Discriminative feature extraction for speech recognition using continuous output codes

Feature transformation techniques have been widely investigated to reduce feature redundancy and to introduce additional discriminative information with the aim to improve the performance of automatic speech recognition (ASR). In this paper, we propose a novel method to obtain discriminative feature transformation based on output coding technique for speech recognition. The output coding transformation projects the speech features from their original space to a new one where each dimension of the features captures information to distinguish different phones. Using polynomial expansion, the short-time spectral features are first expanded to a high-dimensional space where the generalized linear discriminant sequence kernel is applied on the sequences of input feature vectors. Then, the output coding transformation formulated via a set of linear SVMs projects the sequences of high dimensional vectors into a tractable low-dimensional feature space where the resultant features are well-separated continuous output codes for the subsequent multi-class classification problem. Our experimental results on the TIMIT corpus show that the proposed features achieve 10.5% ASR error rate reduction over the conventional spectral features.

Modulation Spectrum Equalization for Improved Robust Speech Recognition

We propose novel approaches for equalizing the modulation spectrum for robust feature extraction in speech recognition. Common to all approaches in that the temporal trajectories of the feature parameters are first transformed into the magnitude modulation spectrum. In spectral histogram equalization (SHE) and two-band spectral histogram equalization (2B-SHE), we equalize the histogram of the modulation spectrum for each utterance to a reference histogram obtained from clean training data, or perform the equalization with two sub-bands on the modulation spectrum. In magnitude ratio equalization (MRE), we define the magnitude ratio of lower to higher modulation frequency components for each utterance, and equalize this to a reference value obtained from clean training data. These approaches can be viewed as temporal filters that are adapted to each testing utterance. Experiments performed on the Aurora 2 and 4 corpora for small and large vocabulary tasks indicate that significant performance improvements are achievable for all noise conditions. We also show that additional improvements can be obtained when these approaches are integrated with cepstral mean and variance normalization (CMVN), histogram equalization (HEQ), higher order cepstral moment normalization (HOCMN), or the advanced front-end (AFE). We analyze and discuss the reasons for these improvements from different viewpoints with different sets of data, including adaptive temporal filtering, noise behavior on the modulation spectrum, phoneme types, and modulation spectrum distance measures.

Vowels recognition using mellin transform and plp-based feature extraction

Feature extraction for speech recognition is a subject of majer interest today. Different feature have been investigated in speech recognition systems. The scale transform is a particular restriction of the Mellin transform. The key property of the scale transform is the scale invariance.The mel frequency(MFCC) and perceptual linear predicate (PLP) have usually reported to have yielded good performance. In this paper, a new method is presented which combines feature extracted from mellin transform with the plp features. To evaluate the performance of the proposed features the comparative tests with the lpc features, MFCC and PLP were performed on the vowels based speech recognition system. Preliminary experiment show that this approach posses promissing result.

MVDR based feature extraction for speech recognition

There is provided a method for extracting feature vectors from a digitized utterance. Spectral envelope estimates are computed from overlapping frames in the digitized utterance based on a Minimum Variance Distortionless Response (MVDR) method. Cepstral feature vectors are generated from the spectral envelope estimates. There is provided a method for generating spectral envelope estimates from a digitized utterance. The spectral envelope estimates are generated from overlapping frames in the digitized utterance based on a harmonic mean of at least two low- to-high resolution spectrum estimates. There is provided a method for reducing variance of a feature stream in a pattern recognition system. The feature stream is temporally or spatially averaged to reduce the variance of the feature stream.

Applying Sparse KPCA for Feature Extraction in Speech Recognition

This paper presents an analysis of the applicability of Sparse Kernel Principal Component Analysis (SKPCA) for feature extraction in speech recognition, as well as, a proposed approach to make the SKPCA technique realizable for a large amount of training data, which is an usual context in speech recognition systems. Although the KPCA (Kernel Principal Component Analysis) has proved to be an efficient technique for being applied to speech recognition, it has the disadvantage of requiring training data reduction, when its amount is excessively large. This data reduction is important to avoid computational unfeasibility and/or an extremely high computational burden related to the feature representation step of the training and the test data evaluations. The standard approach to perform this data reduction is to randomly choose frames from the original data set, which does not necessarily provide a good statistical representation of the original data set. In order to solve this problem a likelihood related re-estimation procedure was applied to the KPCA framework, thus creating the SKPCA, which nevertheless is not realizable for large training databases. The proposed approach consists in clustering the training data and applying to these clusters a SKPCA like data reduction technique generating the reduced data clusters. These reduced data clusters are merged and reduced in a recursive procedure until just one cluster is obtained, making the SKPCA approach realizable for a large amount of training data. The experimental results show the efficiency of SKPCA technique with the proposed approach over the KPCA with the standard sparse solution using randomly chosen frames and the standard feature extraction techniques.

Wavelet-based denoising for robust feature extraction for speech recognition

A new pre-processing stage based on wavelet denoising is proposed to extract robust features in the presence of additive white Gaussian noise. Recognition performance is compared with the commonly used Mel frequency cepstral coefficients with and without this pre-processing stage. The word recognition accuracy is found to improve using the proposed technique by 2 to 28% for signal-to-noise ratio in the range of 20 to 0 dB.

Optimizing feature extraction for speech recognition

We propose a method to minimize the loss of information during the feature extraction stage in speech recognition by optimizing the parameters of the mel-cepstrum transformation, a transform which is widely used in speech recognition. Typically, the mel-cepstrum is obtained by critical band filters whose characteristics play an important role in converting a speech signal into a sequence of vectors. First, we analyze the performance of the mel-cepstrum by changing the parameters of the filters such as shape, center frequency, and bandwidth. Then we propose an algorithm to optimize the parameters of the filters using the simplex method. Experiments with Korean digit words show that the recognition rate improved by about 4-7%.

Zero-crossing-based linear prediction for speech recognition

The satisfactory estimation of speech autocorrelation by means of generalised zero-crossings indicates that they can be used for efficient feature extraction in speech recognition. In addition, high consistency between the Itakura-Saito distances, calculated before and after clipping, allowed for only a mode-rate degradation of the related recognition performance, which was compensated by including the excitation distortion into the distance measure.