Unvoiced Frames Research Articles

Alaryngeal Speech Enhancement for Noisy Environments Using a Pareto Denoising Gated LSTM

Loss of the larynx significantly alters natural voice production, requiring alternative communication modalities and rehabilitation methods to restore speech intelligibility and improve the quality of life of affected individuals. This paper explores advances in alaryngeal speech enhancement to improve signal quality and reduce background noise, focusing on individuals who have undergone laryngectomy. In this study, speech samples were obtained from 23 Lithuanian males who had undergone laryngectomy with secondary implantation of the tracheoesophageal prosthesis (TEP). Pareto optimized gated LSTM was trained on tracheoesophageal speech data to recognize complex temporal connections and contextual information in speech signals. The system was able to distinguish between actual speech and various forms of noise and artifacts, resulting in a 25% drop in the mean signal-to-noise ratio (SNR) compared to other approaches. According to acoustic analysis, the system significantly decreased the number of unvoiced frames (PVF) from 40% to 10% while maintaining stable proportions of voiced frames (PVS) and average voicing evidence (AVE), indicating the accuracy of the approach in selectively attenuating noise and undesired speech artifacts while preserving important speech information.

"Lombard Effect" and Voice Changes in Adductor Laryngeal Dystonia: A Pilot Study.

The aim was to describe the acoustic, auditory-perceptive, and subjective voice changes under the Lombard effect (LE) in adductor laryngeal dystonia (AdLD) patients. Subjective perception of vocal effort (OMNI Vocal Effort Scale OMNI-VES), Maximum Phonation Time (MPT), and the perceptual severity of dysphonia (GRBAS scale) were assessed in condition of stillness and under LE in 10 AdLD patients and in 10 patients with typical voice. Speakers were asked to produce the sustained vowel /a/ and to read a phonetically balanced text aloud. Using the PRAAT software, the following acoustic parameters were analyzed: Mean Pitch (Hz), Minimum and Maximum Intensity (dB), the Fraction of Locally Unvoiced Frames, the Number of Voice Breaks, the Degree of Voice Breaks (%), the Cepstral Peak Prominence-Smoothed (CPPS) (dB). Under LE, the AdLD group showed a decrease of both G and S parameters of GRBAS and subjective effort, mean MPT increased significantly; in the controls there were no significant changes. In both groups under LE, pitch and intensity of the sustained vowel /a/ significantly increased consistently with LE. In the AdLD group the mean gain of OMNI-VES score and the mean gain of each parameter of the speech analysis were significantly greater than the controls' ones. Auditory feedback deprivation obtained under LE improves subjective, perceptual-auditory, and acoustics parameters of AdLD patients. These findings encourage further research to provide new knowledge into the role of the auditory system in the pathogenesis of AdLD and to develop new therapeutic strategies. 4 Laryngoscope, 134:3754-3760, 2024.

Investigating temporal and prosodic markers in clinical high-risk for psychosis participants using automated acoustic analysis.

Language disturbances are a candidate biomarker for the early detection of psychosis. Temporal and prosodic abnormalities have been observed in schizophrenia patients, while there is conflicting evidence whether such deficits are present in participants meeting clinical high-risk for psychosis (CHR-P) criteria. Clinical interviews from CHR-P participants (n=50) were examined for temporal and prosodic metrics and compared against a group of healthy controls (n=17) and participants with affective disorders and substance abuse (n=23). There were no deficits in acoustic variables in the CHR-P group, while participants with affective disorders/substance abuse were characterized by slower speech rate, longer pauses and higher unvoiced frames percentage. Our finding suggests that temporal and prosodic aspects of speech are not impaired in early-stage psychosis. Further studies are required to clarify whether such abnormalities are present in sub-groups of CHR-P participants with elevated psychosis-risk.

Voice Activity Detection: Fusion of Time and Frequency Domain Features with A SVM Classifier

Voice activity detection (VAD) discriminates between segments of an audio signal that has speech content from the ones with either noise or silence. It is deployed as the front-end of some speech processing applications such as voice recognition, and speaker recognition to improve their performance in terms of accuracy and efficiency. It is also used in the communication system to bring about efficient utilization of transmission bandwidth by ensuring only segments of the audio signal with voice activity are encoded and transmitted. In this work, the VAD algorithm was implemented using a features-fusion strategy. In the pre-processing stage, contents outside the human auditory frequency range were removed with the aid of a digital Butterworth bandpass filter. The signal was then fragmented into frames from where time-domain features (zero-crossing, standard deviation, normalized envelope, kurtosis, skewness, and root-mean-square energy.) and frequency-domain features (13MFCCs) were extracted and then combined to form a feature representation of each frame. Recursive feature elimination was applied to the dataset to reduce the features to seven (7) which was used to train a Support Vector Machine (SVM) to be able to distinguish between voiced and unvoiced frames. A State-of-art performance was recorded by this simple SVM-based VAD system with an accuracy of 100%, recall of 100%, precision of 100% and F1 score of 100% which is at par with similar implementations which utilizes a complex architecture of deep neural network with high computational cost and training time. Keywords : Voice activity detection, fusion strategy, support vector machine, frequency domain features, time domain features DOI: 10.7176/CEIS/13-3-03 Publication date: May 31 st 2022

More on Sibilant Devoicing in Spanish Diachrony: An Initial Phonetic Approach

The devoicing of sibilants took place in Early Modern Spanish, a phenomenon which has been considered problematic to account for due to its occurrence context (medial intervocalic position). Traditional explanations invoked Basque influence or a structural reorganization in search for a more balanced system. However, phonetically based reasons were proposed by some scholars. This research is a preliminary attempt to support these proposals with experimental data from a comparative grammar perspective. The Catalan sibilant system, which is very similar to the Medieval Spanish one, is acoustically and perceptively studied in order to investigate the acoustic cues of voicing and to determine if devoicing is possible. Results indicate that (a) voicing relies mainly in the proportion of unvoiced frames of the segments, on its duration, and, to a lesser extent, on its intensity; (b) sibilant devoicing occurs in all voiced categories; (c) auditorily, confusion between voiced and voiceless segments can be attested for every sibilant pair, and (d) the misparsings are more common in affricate and in palatal sibilants, [d͡ʒ] being the most prone to be labelled as unvoiced. These findings prove that the historical process in Spanish could have a phonetic basis.

​Association of Bio-acoustic Features of Vocal Signals with Age and Semen Quality in Sahiwal Bulls

Background: Bio-acoustic features of animal’s voice can provide meaningful information about their biological and physical characteristics. The present study was conducted to get indicators of age from voice analysis and explore the relationship between voice features and seminal parameters in Sahiwal bulls. Methods: Voice samples were collected from healthy bulls (n=20), maintained at ICAR-NDRI, Karnal. Bulls were classified into two groups i.e. young bulls and adult bulls. The voice signals were analyzed by Adobe Premium software and acoustic features were extracted by using PRAAT software. Result: The mean of acoustic features viz. call duration (sec), mean intensity (dB), total energy (P2S), amplitude (P), pitch (Hz), unvoiced frame (%), jitter (%), bandwidth (Hz) mean N/H ratio (%) have been found significantly different while mean H/N ratio (dB), shimmer (%) and pulses were not found statistically (P greater than 0.05) different between adult and young bulls. The seminal parameters viz mass activity (0-5 scale), individual progressive motility (%), live sperm count (%) and total sperm abnormality (%) were found significantly different between adult and young bulls. No significant association between voice features and semen quality of bulls was observed. Hence, voice signals of male might provide some clues about their age but for semen quality, there is further need to explore the interesting relationship between voice features and fertility of breeding bulls.

Extraction of Voiced Regions of Speech from Emotional Speech Signals Using Wavelet-Pitch Method

Extraction of voiced regions of speech is one of the latest topics in speech domain for various speech applications. Emotional speech signals contain most of the information in voiced regions of speech. In this particular work, voiced regions of speech are extracted from emotional speech signals using wavelet-pitch method. Daubechies wavelet (Db4) is applied on the speech frames after downsampling the speech signals. Autocorrelation function is performed on the extracted approximation coefficients of each speech frame and corresponding pitch values are obtained. A local threshold is defined on obtained pitch values to extract voiced regions. The threshold values are different for male and female speakers, as male pitch values are low compared to the female pitch values in general. The obtained pitch values are scaled down and are compared with the thresholds to extract the voiced frames. The transition frames between the voiced and unvoiced frames are also extracted if the previous frame is voiced frame, to preserve the emotional content in extracted frames. The extracted frames are reshaped to have desired emotional speech signal. Signal to Noise Ratio (SNR), Normalized Root Mean Square Error (NRMSE) and statistical parameters are used as evaluation metrics. This particular work provides better SNR and Normalized Root Mean Square Error values compared to the zero crossing-energy and residual signal based methods in voiced region extraction. Db4 wavelet provides better results compared to Haar and Db2 wavelets in extracting voiced regions using wavelet-pitch method from emotional speech signals.

Mel Scale-Based Linear Prediction Approach to Reduce the Prediction Filter Order in CELP Paradigm

This paper proposes a novel method to reduce the order of prediction filter from 10 to 7 in Code Excited Linear Prediction (CELP) coding framework by the inclusion of psychoacoustic Mel scale into Linear Predictive Coding (Mel-LPC). Efficient quantization methods using 2-split Vector Quantization (VQ) for Mel-LPC obtained a reduction of 4 bits/frame and resulted in a total bit gain of 200 bps. A weighting scheme for the Euclidean distance measure gave a reduction of 6 bits/frame that adds up to a total bit gain of 300 bps. A lower Mel-LPC order of 3 has been employed for unvoiced frames by using the perceptual quality as selection criteria and an efficient VQ method using 5 bits is developed which brought down the average bit requirement to 11.5 bits/frame. To incorporate this into Mel-LPC-based CELP encoding scheme, a neural network-based voiced-unvoiced classification algorithm using 5 derived features as input has been constructed and this selection of filter order based on signal statistics provides the benefit of bit reduction by 625 and 325 bps, respectively, for 10th order LPC and 7th order Mel-LPC. In addition to all, the incorporation of Mel-LPC gives a better performance in the estimation of formants.

Stimulus Parameters Underlying Sound-Symbolic Mapping of Auditory Pseudowords to Visual Shapes.

Sound symbolism refers to non-arbitrary mappings between the sounds of words and their meanings and is often studied by pairing auditory pseudowords such as "maluma" and "takete" with rounded and pointed visual shapes, respectively. However, it is unclear what auditory properties of pseudowords contribute to their perception as rounded or pointed. Here, we compared perceptual ratings of the roundedness/pointedness of large sets of pseudowords and shapes to their acoustic and visual properties using a novel application of representational similarity analysis (RSA). Representational dissimilarity matrices (RDMs) of the auditory and visual ratings of roundedness/pointedness were significantly correlated crossmodally. The auditory perceptual RDM correlated significantly with RDMs of spectral tilt, the temporal fast Fourier transform (FFT), and the speech envelope. Conventional correlational analyses showed that ratings of pseudowords transitioned from rounded to pointed as vocal roughness (as measured by the harmonics-to-noise ratio, pulse number, fraction of unvoiced frames, mean autocorrelation, shimmer, and jitter) increased. The visual perceptual RDM correlated significantly with RDMs of global indices of visual shape (the simple matching coefficient, image silhouette, image outlines, and Jaccard distance). Crossmodally, the RDMs of the auditory spectral parameters correlated weakly but significantly with those of the global indices of visual shape. Our work establishes the utility of RSA for analysis of large stimulus sets and offers novel insights into the stimulus parameters underlying sound symbolism, showing that sound-to-shape mapping is driven by acoustic properties of pseudowords and suggesting audiovisual cross-modal correspondence as a basis for language users' sensitivity to this type of sound symbolism.

Design of MELPe-Based Variable-Bit-Rate Speech Coding with Mel Scale Approach Using Low-Order Linear Prediction Filter and Representing Excitation Signal Using Glottal Closure Instants

In this paper, we propose a variable-bit-rate speech codec-based on mixed excitation linear prediction enhanced (MELPe) with an average bit rate of 2 kbps and with a better representation of excitation signal. The order of the prediction filter in MELPe coding architecture is reduced from 10 to 7 without affecting the perceptual quality of the decoded speech by using psychoacoustic Mel scale. An efficient two-split vector quantization is developed with weighted Euclidean distance measure for Mel scale-based linear predictive coding (Mel-LPC), and it requires only 18 bits/frame. The instantaneous pitch or epoch that is vital for many speech processing applications is preserved in this codec by including it in the excitation signal used for reconstructing the voiced speech. The quantization scheme developed for glottal closure instants (GCIs) causes an increase in the bit requirement for voiced frames by 4–25 bits depending on the position of GCIs. To compensate for that, the Mel-LPC order for both silence and unvoiced frames has been brought down to 4 without compromising the perceptual quality of reconstructed speech. The lowered bit budget for unvoiced frame is 41 bits/frame, and for silence, it is 31 bits/frame. Further reduction of 10 bits for silence frame is obtained by reducing the number of transmitted parameters and by tuning the quantization bit requirement for each. For categorizing the speech frames at the entry of the encoder, a neural network-based voiced/unvoiced/silence classification algorithm using five-dimensional feature set is created. The experimental results show that the proposed coding scheme operates at an average bit rate of 2 kbps, which is less than the bit rate of MELPe (2.4 kbps), but with a better perceptual score. In addition to all these, the incorporation of Mel-LPC gives a better performance in the estimation of formants and GCIs.

Performance evaluation of psycho-acoustically motivated front-end compensator for TIMIT phone recognition

Wavelet-based front-end processing technique has gained popularity for its noise removing capability. In this paper, a robust automatic speech recognition system is proposed by utilizing the advantages of psycho-acoustically motivated wavelet-based front-end compensator. In the front-end compensator block, voiced speech probability-based voice activity detector system is designed to separate voiced and unvoiced frames and to update noise statistics. The wavelet packet decomposition tree is designed according to equal rectangular bandwidth (ERB) scale. Wavelet decomposition based on ERB scale is utilized here as the central frequency of the ERB distribution resembles frequency response of human cochlea. Voiced and unvoiced frames are separately decomposed into 24 sub-bands to estimate average sub-band energy (ASE) of each frame. ASE is then used to calculate threshold value. Lastly, Wiener filtering is employed for reducing the residual noise before final reconstruction stage. The proposed system is evaluated on TIMIT database under various noise conditions. The phoneme recognition accuracy of the proposed system is compared with different baseline and robust features as well as with existing front-end compensation techniques. Additionally, the proposed front-end compensator is evaluated in terms of phoneme classification accuracy. Performance improvement is observed in all above experiments.

Recognition of isolated digits using DNN–HMM and harmonic noise model

Speech recognition is an area that is constantly developing. In this study, the authors present a new system of speech recognition applied to the Arabic language. The system proposed here is based on the harmonic plus noise model (HNM). This model is rather used in speech synthesis tasks and is known for providing excellent speech production quality. Thus, their contribution lies in replacing the conventional mel-frequency cepstrum coefficients (MFCC) parameters with a set of acoustic parameters, extracted through the HNM estimation process. The HNM model allows development of a more adapted processing by distinguishing voiced and unvoiced speech frames and by characterising the harmonic property of speech. As common, their system consists of both training and recognition phases. Deep neural networks and hidden Markov models (DNN–HMM) are used for modelling the voiced frames corresponding to the harmonic part. The DNN model is estimated with static and dynamic parameters. Moreover, the unvoiced frames, which represent the noise part of the HNM, are clustered with an HMM model. The spoken Arabic digits are used to measure the performance of the proposed recognition system and a comparison with the MFCC-based approach is performed.

Discrimination of karan fries cow’s individuality by the mean of their vocal acoustic features

Sound is one of the most important means of conveying information over long distances as well as in close vicinity. Utterance of animals becomes unique for them when they communicate their individuality or physiological state to the other co species partner. Present study was based on hypothesis of discrimination of individual identity through vocal signal of lactating karan fries (KF) cows. For this 25 KF cows were selected for recording of their vocal signal. Vocal call was recorded after separation of animal from their living herd in the morning hours. Acoustic features of vocalization of individual cow were extracted with the help of PRAAT acoustic analysis software. Analysis of all acoustic features extracted from 250 voice samples of 25 KF cows revealed that differences for amplitude, Total energy, pitch, intensity, formants, pulse, periods, unvoiced frames, voice breaks, jitter, shimmer, mean noise/harmonic ratio and mean harmonic/noise ratio were found highly significant (p less than 0.001). Out of these only few acoustic features viz. pulse, pitch, jitter, shimmer, voice break and formants were observed to have significant (p less than 0.05) difference between each and every individual KF cow. Among these feature formants frequency of every individual cow had a unique pattern in the distribution of frequency contour in their vocal spectrogram. Study concluded that vocal signal of KF cow have some unique feature at individual level, by which cow identification could be made.

Multiple Time-Instances Features of Degraded Speech for Single Ended Quality Measurement

The use of single time-instance features, where entire speech utterance is used for feature computation, is not accurate and adequate in capturing the time localized information of short-time transient distortions and their distinction from plosive sounds of speech, particularly degraded by impulsive noise. Hence, the importance of estimating features at multiple time-instances is sought. In this, only active speech segments of degraded speech are used for features computation at multiple time-instances on per frame basis. Here, active speech means both voiced and unvoiced frames except silence. The features of different combinations of multiple contiguous active speech segments are computed and called multiple time-instances features. The joint GMM training has been done using these features along with the subjective MOS of the corresponding speech utterance to obtain the parameters of GMM. These parameters of GMM and multiple time-instances features of test speech are used to compute the objective MOS values of different combinations of multiple contiguous active speech segments. The overall objective MOS of the test speech utterance is obtained by assigning equal weight to the objective MOS values of the different combinations of multiple contiguous active speech segments. This algorithm outperforms the Recommendation ITU-T P.563 and recently published algorithms.

Acoustic features of vocalization during different phases of estrous cycle in Murrah buffaloes

The present study was conducted to investigate the changes in acoustic features of vocal signals articulated by Murrah buffaloes during different phases of estrous cycle. Data (voice) were collected from 20 Murrah buffaloes, maintained at Livestock Research Centre, NDRI, Karnal, during 4 phases of estrous cycle. Following acoustic features (call duration, formants, intensity, pitch, noise/harmonic (N/H) ratio, harmonic/noise ratio, number of periods, pulses, jitter, shimmer and unvoiced frame) were extracted from vocal signals of Murrah buffaloes for analysis purpose. The acoustic features, viz. call duration, fourth formant (F4), shimmer and intensity were significantly higher during proestrus phase as compared to other phases of estrous cycle. Whereas, F1, F5, noise to harmonic ratio, per cent unvoiced frames and pitch were significantly higher during estrus phase. All these features have significant correlation with estradiol hormone during proestrus and estrus phase. The other features namely number of periods and pulses were significantly correlated with progesterone hormone during diestrus phase. Therefore, the acoustic features, viz. F1, F5, N/H ratio and unvoiced frame can be utilized for identification of estrus phase in Murrah buffaloes.

Sparse Representations for Single Channel Speech Enhancement Based on Voiced/Unvoiced Classification

The approach presented here in relies on a new voicing decision algorithm based on the multi-scale product (MP) characteristics. The MP is based on the multiplication of Wavelet Transform Coefficients at some scales. According to the voicing decision, improved subspace decomposition is operated on the voiced segments of the noisy speech signal and a multi-scale principal component analysis is applied on the unvoiced segments of the same signal. Furthermore, the voiced frames are decomposed into three subspaces: sparse, low rank, and the remainder noise components. Then, we calculate the components as a segregation problem. In the unvoiced frames, we combine the straightforward multivariate generalization of the wavelet denoising technique with the principal component analysis method. Experiments on NOIZEUS and NTT databases show that the proposed approach obtains satisfying results for most types of noise with little speech degradation and outperforms several competitive methods.

A New Biologically Inspired Fuzzy Expert System-Based Voiced/Unvoiced Decision Algorithm for Speech Enhancement

In this paper, we propose a speech enhancement approach for a single-microphone system. The main idea is to apply a specific transformation on the speech signal depending on the voicing state of the signal. We apply a voiced/unvoiced algorithm based on the multi-scale product analysis with the use of fuzzy logic to make more cognitively inspired use of speech information. A comb filtering is applied on the voiced frames of the noisy speech signal, and a spectral subtraction is operated on the unvoiced frames of the same signal. Further, the harmonics are enhanced by performing a designed comb filtering using an adjustable bandwidth. The comb filter is tuned by an accurate fundamental frequency estimation method. The fundamental frequency estimation method is based on computing the multi-scale product analysis of the noisy speech. Experimental results show that the proposed approach is capable of reducing noise in adverse noise environments with little speech degradation and outperforms several competitive methods.

Automatic detection of Parkinson's disease in running speech spoken in three different languages

The aim of this study is the analysis of continuous speech signals of people with Parkinson's disease (PD) considering recordings in different languages (Spanish, German, and Czech). A method for the characterization of the speech signals, based on the automatic segmentation of utterances into voiced and unvoiced frames, is addressed here. The energy content of the unvoiced sounds is modeled using 12 Mel-frequency cepstral coefficients and 25 bands scaled according to the Bark scale. Four speech tasks comprising isolated words, rapid repetition of the syllables /pa/-/ta/-/ka/, sentences, and read texts are evaluated. The method proves to be more accurate than classical approaches in the automatic classification of speech of people with PD and healthy controls. The accuracies range from 85% to 99% depending on the language and the speech task. Cross-language experiments are also performed confirming the robustness and generalization capability of the method, with accuracies ranging from 60% to 99%. This work comprises a step forward for the development of computer aided tools for the automatic assessment of dysarthric speech signals in multiple languages.

Bio-acoustic: A non-invasive and effective sensing technique in monitoring of dairy buffaloes

The present study was performed to observe the significant effect of individuality of dairy buffaloes on the acoustic features of their vocal signals. The mean call duration, pitch, 1st formant, periodicity and degree of voice breaks of adult lactating Murrah buffaloes were observed to be 2.15±0.05 s, 143.48 ± 2.51 Hz, 900.11 ± 4.21 Hz, 95.31% (183.95 periodic pulses out of 192.99 pulses) and 20.78 ± 0.89 % respectively. Analysis of bioacoustics features extracted from 300 voice samples of 10 adult lactating Murrah buffaloes revealed that differences for amplitudes (minimum, maximum and mean), total energy, mean power, pitch (median, minimum, range and mean), intensities (mean, minimum and maximum), formants (F1, F2, F3, F4 and F5), bandwidths (B1, B3 and B4), number of pulse, number of period, mean period, unvoiced frames, degree of voice breaks, jitter, shimmer, mean noise to harmonic ratio (%) and mean harmonic to noise ratio (dB) were highly significant. Out of these only few acoustic features viz. formants (F1, F2, F3, F4 and F5), number of pulse, number of period, degree of voice breaks, mean noise to harmonic ratio (%) and mean harmonic to noise ratio (dB) were observed to have significant difference for each and every individual dairy buffalo, hence only these features could be selected as the best suited acoustic features for discrimination of individual Murrah buffaloes from their herd.

Use of baseband phase structure to improve the performance of current speech enhancement algorithms

In this study we propose a noise estimation technique based on spectral sparsity, detected by using the harmonic property of voiced segments of the speech. We estimate the frame to frame phase difference for clean speech, in the baseband Short Time Fourier Transform (STFT) domain, from corrupted speech. This estimated frame-to-frame phase difference is used as a means of detecting the noise and harmonic dominant frequency bins in voiced frames. In the unvoiced frames noise is estimated using a Voice Activity Detector (VAD). Using this approach gives better noise estimation for the highly non-stationary noises like babble, restaurant and subway noise. The proposed noise estimation algorithm along with the phase difference as an additional prior is used to extend the standard spectral subtraction algorithm. Estimated baseband phase difference is used to prevent over-suppression of harmonic dominant speech in the voiced frames by adjusting the over-attenuation factor of the spectral subtraction algorithm. We verify the effectiveness of the proposed noise estimation technique when used with the log-Minimum Mean Squared Error Short Time Spectral Amplitude Estimator (log-MMSE STSA) speech enhancement algorithm. Finally, spectral subtraction and log-MMSE STSA are combined to achieve better noise suppression with minimum musical noise and speech distortion. This combination results in further improvement of speech quality.