Poonam Bansal

Robust Features for Noisy Speech Recognition using MFCC Computation from Magnitude Spectrum of Higher Order Autocorrelation Coefficients

robustness is one of the most challenging problem in automatic speech recognition. The goal of robust feature extraction is to improve the performance of speech recognition in adverse conditions. The mel-scaled frequency cepstral coefficients (MFCCs) derived from Fourier transform and filter bank analysis are perhaps the most widely used front-ends in state-of-the-art speech recognition systems. One of the major issues with the MFCCs is that they are very sensitive to additive noise. To improve the robustness of speech front-ends we introduce, in this paper, a new set of MFCC vector which is estimated through three steps. First, the relative higher order autocorrelation coefficients are extracted. Then magnitude spectrum of the resultant speech signal is estimated through the fast Fourier transform (FFT) and it is differentiated with respect to frequency. Finally, the differentiated magnitude spectrum is transformed into MFCC-like coefficients. These are called MFCCs extracted from Differentiated Relative Higher Order Autocorrelation Sequence Specrum (DRHOASS). Speech recognition experiments for various tasks indicate that the new feature vector is more robust than traditional mel-scaled frequency cepstral coefficients (MFCCs) in additive noise conditions.

Optimum HMM combined with vector quantization for hindi speech word recognition

Abstract This paper proposes an optimum speaker-independent, isolated word Hidden Markov Model (HMM) recognizer for the Hindi language. The recognition system is based on the combination of the vector quantization (VQ) technique at the acoustical level and the Markovian modeling at the recognition level. The recognizer consists of three modules – feature extraction, vector quantizer and HMM training and testing modules. The scheme proposed here firstly computes the acoustic features in terms of the Linear Predictive Cepstral LPC coefficients, Mel-Frequency Cepstral coefficients and delta MFCC along with noise and silence detection. Then, codebooks are created using VQ, and finally in the recognition phase, an optimum set of parameters are derived from different phases for getting the highest recognition score. The training and testing database consists of a set of 35 utterances of nine Indian cities/states and 35 utterances of nine digits spoken in Hindi by male and female speakers. The recognition rate was observed to be 98.61%.

Performance analysis of least mean square algorithm for different step size parameters with different filter order and iterations

This paper presents the performance analysis of Least Mean Square (LMS) algorithm for adaptive noise cancellation by varying its step size parameter μ for different filter order and no of iteration. The presented work has been simulated in MATLAB and verified that the step size parameter plays a vital role for implementation of Least Mean Square (LMS) algorithm. Increasing the step size parameter μ leads to fast convergence rate and instability of the least mean square algorithm. On the other side if the step size parameter μ is small then the error reduced to great amount but algorithm converges slowly and becomes stable. On the basis of obtained results we can conclude that step size parameter μ is directly proportional to convergence rate and error reduction and inversely proportional to stability. The work presented here also shown the comparison of actual weights and the estimated weights.

An approach to implement LMS and NLMS adaptive noise cancellation algorithm in frequency domain

This paper presents the implementation of adaptive algorithms like Least Mean Square (LMS) and Normalized Least Mean Square (NLMS) in the frequency domain and their comparison to that implemented in the time domain. Adaptive filtering using adaptive algorithm in frequency domain can be done by taking Fourier Transform of input signal and independent weigh coefficient. By frequency domain approach significant reduction in mathematical computation has been achieved. An expression for updating the weights is implemented in the frequency domain and statistical analysis has been performed. The SNR (Signal to Noise Ratio) is a parameter used to evaluate the performance with different step size. The signal power and noise power has also been calculated using MATLAB. The SNR of output signal rises about 8-9 times in frequency domain than the time domain.

Robust Feature Vector Set Using Higher Order Autocorrelation Coefficients

In this paper, a feature extraction method that is robust to additive background noise is proposed for automatic speech recognition. Since the background noise corrupts the autocorrelation coefficients of the speech signal mostly at the lower orders, while the higher-order autocorrelation coefficients are least affected, this method discards the lower order autocorrelation coefficients and uses only the higher-order autocorrelation coefficients for spectral estimation. The magnitude spectrum of the windowed higher-order autocorrelation sequence is used here as an estimate of the power spectrum of the speech signal. This power spectral estimate is processed further by the Mel filter bank; a log operation and the discrete cosine transform to get the cepstral coefficients. These cepstral coefficients are referred to as the Differentiated Relative Higher Order Autocorrelation Coefficient Sequence Spectrum DRHOASS. The authors evaluate the speech recognition performance of the DRHOASS features and show that they perform as well as the MFCC features for clean speech and their recognition performance is better than the MFCC features for noisy speech.

Noise Reduction from ECG Signal Using Error Normalized Step Size Least Mean Square Algorithm (ENSS) with Wavelet Transform

This paper presents the reduction of baseline wander noise found in ECG signals. The reduction has been done using wavelet transform inspired error normalized step size least mean square (ENSS-LMS) algorithm. We are presenting a wavelet decomposition-based filtering technique to minimize the computational complexity along with the good quality of output signal. The MATLAB simulation results validate the good noise rejection in output signal by analyzing parameters, excess mean square error (EMSE) and misadjustment.

The State of the Art of Feature Extraction Techniques in Speech Recognition

This paper surveys feature extraction techniques applied in automatic speech recognition. After so many researches and improvement, the accuracy is a key issue in speech recognition systems. Speech recognition process converts the speech signal into its corresponding written text by the computer system. In this paper, we brief few well-known techniques of feature extraction like LPC, MFCC, RASTA, PCA, LDA, PLP.

Noise robust acoustic signal processing using a Hybrid approach for speech recognition

The paper proposes English digit recognition using a Hybrid approach for both speaker dependent and independent mode in clean and noisy situations. This work uses Discrete Hidden Markov Model for digit recognition where DHMM is modeled with quantized vectors. Mel Frequency Cepstral coefficients is used at feature extraction stage to extract feature vectors of all digits and it also uses filters that operates on cepstrum obtained from MFCC to smoothens signal. This work uses Empirical Mode Decomposition (EMD) process for the noises separation from speech signals. The database containing digits from Zero to Nine is used for the work. This works propose that recognition rate of English digits in noisy environment can be improved by applying filters to the cepstrum obtained from MFCC and by using EMD along with Genetic Algorithm for noise removal.

An approach to implement VSS-LMS algorithm in frequency domain for adaptive noise cancellation

This paper presents the Variable Step Size Least Mean Square algorithm formulated in frequency domain by taking the (Fast Fourier Transform) FFT of signal obtained from filter. This way the algorithms performed better than its implementation in time domain in terms of Signal to Noise Ratio (SNR). The algorithms implemented in MATLAB with different colored noise surroundings. To evaluate the performance of the algorithm its comparison has been done with time domain. The algorithm has given 5-44% increased SNR compared to that implemented in time domain with different type of colored noises. The algorithm has also been tested in frequency domain for different step sizes.

Speech Synthesis - Automatic Segmentation

In this paper, after an a review of the previous work done in this field, the most frequently used approach using Hidden Markov Model (HMM) is used for implementation for phonetic segmentation. A baseline HMM phonetic segmentation tool is used for segmentation and analysis of speech at phonetic level. The results are approximately same as obtained using manual segmentation. General Terms Speech Synthesis, Automatic Segmentation