Linear prediction coefficients correction method for digital speech processing systems with data compression based on the autoregressive model of a voice signal

Khasi Speech Recognition using Hidden Markov Model with Different Spectral Features: A Comparison

LPC-based formant enhancement method in Kalman filtering for speech enhancement

Background/Objectives: Speech produced by individuals with hearing loss differs notably from that of normal-hearing (NH) individuals. Although cochlear implants (CIs) provide sufficient auditory input to support speech acquisition and control, there remains considerable variability in speech intelligibility among CI users. As a result, speech produced by CI talkers often exhibits distinct acoustic characteristics compared to that of NH individuals. Methods: Speech data were obtained from eight cochlear-implant (CI) and eight normal-hearing (NH) talkers, while electroencephalogram (EEG) responses were recorded from 11 NH listeners exposed to the same speech stimuli. Support Vector Machine (SVM) classifiers employing 3-fold cross-validation were evaluated using classification accuracy as the performance metric. This study evaluated the efficacy of Support Vector Machine (SVM) algorithms using four kernel functions (Linear, Polynomial, Gaussian, and Radial Basis Function) to classify speech produced by NH and CI talkers. Six acoustic features—Log Energy, Zero-Crossing Rate (ZCR), Pitch, Linear Predictive Coefficients (LPC), Mel-Frequency Cepstral Coefficients (MFCCs), and Perceptual Linear Predictive Cepstral Coefficients (PLP-CC)—were extracted. These same features were also extracted from electroencephalogram (EEG) recordings of NH listeners who were exposed to the speech stimuli. The EEG analysis leveraged the assumption of quasi-stationarity over short time windows. Results: Classification of speech signals using SVMs yielded the highest accuracies of 100% and 94% for the Energy and MFCC features, respectively, using Gaussian and RBF kernels. EEG responses to speech achieved classification accuracies exceeding 70% for ZCR and Pitch features using the same kernels. Other features such as LPC and PLP-CC yielded moderate to low classification performance. Conclusions: The results indicate that both speech-derived and EEG-derived features can effectively differentiate between CI and NH talkers. Among the tested kernels, Gaussian and RBF provided superior performance, particularly when using Energy and MFCC features. These findings support the application of SVMs for multimodal classification in hearing research, with potential applications in improving CI speech processing and auditory rehabilitation.

ABSTRACTThe goal of this study is to present a comparative analysis between the three feature extraction techniques: Linear Predictive Coefficients (LPC), Linear Prediction Cepstral Coefficients (LPCC), and Weighted Linear Prediction Cepstral Coefficients (WLPCC) for distinguishing speech of Intellectually Disabled (ID) children from Typically Developed (TD). Speech samples of ID children were recorded from a government-owned special school in India and used for the analysis. Control samples were taken from the author’s institute. Pre-processing techniques were used for better parameterization. Two classifiers, Artificial Neural Network (ANN) and Linear Discriminant Analysis , were applied to classify between disordered and the normal speech. Cross-validation technique was used to evaluate the reliability of the classifier. This study has discussed the effect on the classification accuracy using different values of frame size, percentage of frame overlapping, α value of pre-emphasis first-order filter and the order p. The experimental calculation interpreted that all the three parameterization techniques can be used for the correct classification of the ID speech from controls and WLPCC features slightly outperform than LPCC and considerably outperforms than LPC features.

In this paper, we propose a post-filter using warped linear prediction (WLP) in the frequency domain to suppress pre-echo artifacts. Conventional post-filter in the frequency domain are derived from linear prediction (LP) coefficients. The proposed method is simply implemented by the WLP coefficients instead of the LP coefficients. Since the time domain signal could be adjusted by WLP analysis in the frequency domain, we analyzed the onset position of the transient signal and set the appropriate warping factor to emphasize the resolution of the time domain signal. A subjective listening test was performed for transient signals and it was confirmed that the proposed method offers better subjective quality.

Automatic Speech Emotion Recognition (SER) is a current research topic in the field of Human Computer Interaction (HCI) with a wide range of applications. The purpose of speech emotion recognition system is to automatically classify speaker's utterances into different emotional states such as disgust, boredom, sadness, neutral, and happiness. The speech samples in this paper are from the Berlin emotional database. Mel Frequency cepstrum coefficients (MFCC), Linear prediction coefficients (LPC), linear prediction cepstrum coefficients (LPCC), Perceptual Linear Prediction (PLP) and Relative Spectral Perceptual Linear Prediction (Rasta-PLP) features are used to characterize the emotional utterances using a combination between Gaussian mixture models (GMM) and Support Vector Machines (SVM) based on the Kullback-Leibler Divergence Kernel. In this study, the effect of feature type and its dimension are comparatively investigated. The best results are obtained with 12-coefficient MFCC. Utilizing the proposed features a recognition rate of 84% has been achieved which is close to the performance of humans on this database.

Feature extraction from speech representation is one of the processes in speech recognition. Parametric modeling is a dominant approach to model speech signals. Within a localized interval, speech representation is equivalent to a noise driven output from an all-pole system that can be estimated using linear prediction. Besides the characteristics of speech, temporal variability of speech signal model is also due to the computation of linear prediction coefficients. Thus, an alternative representation is proposed based on the Gabor coefficients. In this paper, a comparison is made with the linear prediction coefficients to show the consistency of the parameters that are generated for implementation in the speech recognition system.

In this paper, a mode decomposition analysis-based adaptive approach is proposed to provide high diagnostic performance for automated voice pathology detection systems. The aim of the study is to develop a reliable and effective system using adaptive cepstral domain features derived from the empirical mode decomposition (EMD), ensemble empirical mode decomposition (EEMD), and complete empirical mode decomposition with adaptive noise (CEEMDAN) methods. The descriptive feature sets are obtained by applying mel-frequency cepstral coefficients (MFCCs) and their derivatives, linear predictive coefficients (LPCs) and linear predictive cepstral coefficients (LPCCs) techniques to each decomposition level. The class-balanced data are generated on the VOice ICar fEDerico II database samples using the synthetic minority oversampling technique (SMOTE). The ReliefF algorithm is used to select the most effective and distinctive features. A combination of selected features and a support vector machine (SVM) classifier is used to identify pathological voices. In the pathology detection approach, the results show that the cepstral features based on EMD and SVM-cubic achieves the highest performance with 99.85\% accuracy, 99.85\% F1-score and 0.997 Matthews correlation coefficient (MCC). In pathology-type classification, the cepstral features based on EEMD and SVM-quadratic approach provided the highest performance with 96.49\% accuracy, 96.46\% F1 and 0.949 MCC values. The comprehensive results of this study reveal that mode decomposition-based approaches are more successful and effective than traditional methods for detection and classification of pathological voices.

Parkinson’s disease (PD) is one type of progressive neurodegenerative disease known as motor system syndrome, which is due to the death of dopamine-generating cells, a region of the human midbrain. PD normally affects people over 60 years of age, which at present has influenced a huge part of worldwide population. Lately, many researches have shown interest into the connection between PD and speech disorders. Researches have revealed that speech signals may be a suitable biomarker for distinguishing between people with Parkinson’s (PWP) from healthy subjects. Therefore, early diagnosis of PD through the speech signals can be considered for this aim. In this research, the speech data are acquired based on speech behaviour as the biomarker for differentiating PD severity levels (mild and moderate) from healthy subjects. Feature extraction algorithms applied are Mel Frequency Cepstral Coefficients (MFCC), Linear Predictive Coefficients (LPC), Linear Prediction Cepstral Coefficients (LPCC), and Weighted Linear Prediction Cepstral Coefficients (WLPCC). For classification, two types of classifiers are used: k-Nearest Neighbour (KNN) and Probabilistic Neural Network (PNN). The experimental results demonstrated that PNN classifier and KNN classifier achieve the best average classification performance of 92.63% and 88.56% respectively through 10-fold cross-validation measures. Favourably, the suggested techniques have the possibilities of becoming a new choice of promising tools for the PD detection with tremendous performance.

In view of the growing use of automatic speech recognition in the modern society, we study various alternative representations of the speech signal that have the potential to contribute to the improvement of the recognition performance. In this paper wavelet based features using different wavelets are used for Hindi digits recognition. The recognition performance of these features has been compared with Linear Prediction Coefficients (LPC) and Perceptual Linear Prediction (PLP) features. All features have been tested using Hidden Markov Model (HMM) based classifier for speaker independent Hindi digits recognition. The recognition performance of PLP features is11.3% better than LPC features. The recognition performance with db10 features has shown a further improvement of 12.55% over PLP features. The recognition performance with db10 is best among all wavelet based features.

A sonar trainer requires a wide variety of broadband background noises to create realistic sounds and displays. Linear prediction techniques are typically used for speech. This paper describes how linear prediction coefficient (LPC) techniques were used to simulate a wide variety of underwater background noises for trainers. First the spectral and spatial background noise models are developed. Then the application of a single channel LPC synthesizer for generating background noise is described, along with test results.

Several interpolation techniques of linear prediction coefficients (LPC) for speech signal coding were experimentally analyzed. The obtained results show that direct interpolation of LPC can be applied for coding and decoding of significantly long segments of voiced speech. The best performance in sense of least number of unstable filters, smallest spectral distortions and greatest MOS score, was achieved when for interpolation were used linear and Hermite polynomials. Higher order interpolation polynomials such us cubic or third order splines, produced significant alternations in obtained LPC that led to worst results of the quality of the decoded speech. Also higher order polynomials without additional post processing of interpolated LPC produced more unstable filters. Ill. 5, bibl. 11 (in English; summaries in English, Russian, Lithuanian).

This paper proposes a maximum likelihood sequence estimation (MLSE) for the differential space-time block code (DSTBC) in cooperation with blind linear prediction (BLP) of fast frequency-flat fading channels. This method that linearly predicts the fading complex envelope derives its linear prediction coefficients by the method of Lagrange multipliers, and does not require data of decision-feedback or information on the channel parameters such as the maximum Doppler frequency in contrast to conventional ones. Computer simulations under fast fading conditions demonstrate that the proposed method with an appropriate degree of polynomial approximation is superior in BER performance to the conventional method that estimates the coefficients by the RLS algorithm using a training sequence.

In this paper, a replacement algorithm for Linear Prediction Coefficients (LPC) along with Hamming Correction Code based Compressor (HCDC) algorithms are investigated for speech compression. We started with an CELP system with order 12 and with Discrete Cosine Transform (DCT) based residual excitation. Forty coefficients with transmission rate of 5.14 kbps were first used. For each frame of the testing signals we applied a multistage HCDC, we tested the compression performance for parities from 2 to 7, we were able to achieve compression only at parity 4. This rate reduction was made with no compromise in the original CELP signal quality since compression is lossless. The compression approach is based on constructing dynamic reflection coefficients codebook, this codebook is constructed and used simultaneously using a certain store/retrieve threshold. The initial linear prediction codec we used is excited by a discrete cosine transform (DCT) residual, the results were tested using the MOS and SSNR, we had acceptable ranges for the MOS (average 3.6), and small variations of the SSNR (±5 db).

The performance of Kalman filters (KF) in speech enhancement is highly dependent on the estimation accuracy of linear predictive coefficients (LPCs). To address the significant estimation bias of LPCs in complex noisy environments, this paper proposes a time-frequency dual-domain Conformer-based Kalman filter speech enhancement method—CAKFNet. The model adopts an encoder-decoder architecture: during the encoding phase, a Dilated DenseNet module is introduced to exponentially expand the receptive field and enhance local feature continuity, thereby achieving multi-scale feature aggregation. Subsequently, dual Conformer modules in both time and frequency domains jointly model the long-term dependencies of speech signals and local convolutional features, fully capturing spatio-temporal correlations. The decoding stage employs a decoupled dual-branch architecture to separately regress the LPC power spectrum (LPC-PS) of clean speech and noise. Building upon this, an autocorrelation matrix is generated via inverse Fourier transform. The Levinson-Durbin recursive algorithm is then employed to compute the clean speech and noise LPC coefficients required for the Augmented Kalman Filter (AKF), along with their prediction error variance, ultimately achieving speech enhancement. The proposed method undergoes systematic experimental evaluation on the publicly available VoiceBank + DEMAND dataset. Results demonstrate that CAKFNet effectively reduces spectral distortion in LPC estimation. Objective metrics, including PESQ and STOI, show significant improvements over baseline methods, while subjective listening tests confirm substantial enhancements in speech quality and intelligibility.