Voice Signal Research Articles

Voice Signal Quality Assessment Based on Signal Quality Standards and Analysis

Voice Signal Quality Assessment Based on Signal Quality Standards and Analysis

Degree of Breathiness in a Synthesized Voice Signal as it Differentiates Masculine versus Feminine Voices

Most studies determining speakers' perceived gender as binarily female or male are reliant on F0 perception, although other vocal parameters may also contribute to the perception of gender. The current study focused on the impact of breathiness on the perception of speakers' gender as a biological variable (feminine or masculine). n=31 normal hearing, native English speakers, 18 female, 13 male, mean age 23 (SD=3.54), were auditorily and visually trained in and then took part in a categorical perception task. A continuum of nine samples of the word "hello", was created in an airway modulation model of speech and voice production. Resting vocal fold length, resting vocal fold thickness, F0, and vocal tract length were fixed. Glottal width at the vocal process, posterior glottal gap, and bronchial pressure were continually modified for all stimuli. Each stimulus was randomly presented 30 times within each of the fiveblocks (150 presentations in total). Participants rated stimuli as binarily female or male. Showed a sigmoidal shift in breathiness along the continuum between perceived feminine or masculine voicing. This shift was evident at stimuli four and five, indicating a nonlinear, discrete perception of breathiness among participants. Response times were also significantly slower in these two stimuli, suggesting a categorical perception of breathiness among participants. Breathiness created by the change in glottal width of at least 0.21cm may influence the perception of a speaker's perceived gender.

Deep time-recurrence features

In this study, the notion of recurrence in chaos and nonlinear dynamics is formulated as a new method for deep feature extraction from complex data. This feature type can enhance the learning power of a state-of-the-art approach in artificial intelligence using time series or sequential data. The deep time-recurrence features are derived using the fuzzy recurrence algorithm and the iterative procedure of convolution, non-linear transformation, and down-sampling. Experimental results obtained from using a public database of healthy and pathological voice signals show that the proposed features can significantly improve the classification performance of long short-term memory networks.

An Experimental Analysis on Multicepstral Projection Representation Strategies for Dysphonia Detection.

Biometrics-based authentication has become the most well-established form of user recognition in systems that demand a certain level of security. For example, the most commonplace social activities stand out, such as access to the work environment or to one's own bank account. Among all biometrics, voice receives special attention due to factors such as ease of collection, the low cost of reading devices, and the high quantity of literature and software packages available for use. However, these biometrics may have the ability to represent the individual impaired by the phenomenon known as dysphonia, which consists of a change in the sound signal due to some disease that acts on the vocal apparatus. As a consequence, for example, a user with the flu may not be properly authenticated by the recognition system. Therefore, it is important that automatic voice dysphonia detection techniques be developed. In this work, we propose a new framework based on the representation of the voice signal by the multiple projection of cepstral coefficients to promote the detection of dysphonic alterations in the voice through machine learning techniques. Most of the best-known cepstral coefficient extraction techniques in the literature are mapped and analyzed separately and together with measures related to the fundamental frequency of the voice signal, and its representation capacity is evaluated on three classifiers. Finally, the experiments on a subset of the Saarbruecken Voice Database prove the effectiveness of the proposed material in detecting the presence of dysphonia in the voice.

Performance Improvement of Speech Emotion Recognition Systems by Combining 1D CNN and LSTM with Data Augmentation

In recent years, the increasing popularity of smart mobile devices has made the interaction between devices and users, particularly through voice interaction, more crucial. By enabling smart devices to better understand users’ emotional states through voice data, it becomes possible to provide more personalized services. This paper proposes a novel machine learning model for speech emotion recognition called CLDNN, which combines convolutional neural networks (CNN), long short-term memory neural networks (LSTM), and deep neural networks (DNN). To design a system that closely resembles the human auditory system in recognizing audio signals, this article uses the Mel-frequency cepstral coefficients (MFCCs) of audio data as the input of the machine learning model. First, the MFCCs of the voice signal are extracted as the input of the model. Local feature learning blocks (LFLBs) composed of one-dimensional CNNs are employed to calculate the feature values of the data. As audio signals are time-series data, the resulting feature values from LFLBs are then fed into the LSTM layer to enhance learning on the time-series level. Finally, fully connected layers are used for classification and prediction. The experimental evaluation of the proposed model utilizes three databases: RAVDESS, EMO-DB, and IEMOCAP. The results demonstrate that the LSTM model effectively models the features extracted from the 1D CNN due to the time-series characteristics of speech signals. Additionally, the data augmentation method applied in this paper proves beneficial in improving the recognition accuracy and stability of the systems for different databases. Furthermore, according to the experimental results, the proposed system achieves superior recognition rates compared to related research in speech emotion recognition.

Generation of Voice Signal Tone Sandhi and Melody Based on Convolutional Neural Network

There is a need to prevent the use of modulated voice signals to conduct criminal activities. Voice signal change detection based on convolutional neural networks is proposed. We use three commonly used voice processing software (Audacity, CoolEdit, and RTISI) to change tones in voice libraries. The research further raises each voice by five semitones and are recorded at different levels (+4, +5, +6, +7, and +8, respectively). Simultaneously, every voice is lowered by five halftones, represented as –4, –5, –6, –7, and –8, respectively. The convolution neural network corresponding to network b-3 is determined as the final classifier in this article through experiments. The average accuracy A1 of its three categories has reached more than 97%, the detection accuracy A2 of electronic tone sandhi speech has reached more than 97%, and the false alarm rate of the original speech is less than 1.9%. The outcomes obtained shows that the detection algorithm in this article is effective, and it has good generalization ability.

A pragmatic approach of Parkinson disease detection using hybrid case-based reasoning neuro-fuzzy classification system over Mobile edge computing

Parkinson’s disease is neurological degenerative disorder cause by deficient dopamine production which in turn harms the motor functionality and speech. With latest IoT advancement in the health care era, we propose intelligent and smart Parkinson’s disease detection system based on voice signal analysis. Addition to PDs detection, we propose remote health monitoring feature that keep on monitoring and diagnosing PD person activity. To perform all tasks efficiently we divide our propose model in three phases: monitoring, diagnosing and analysis. During monitoring phase, PDs person voice signal is monitored and captured via IoT sensor enabled Smartphone device. This voices signal is further processed for PD detection over MEC server during diagnosing phase. We use Tunable Q factor wavelet transform (TQWT) for extracting feature from voice sample, these extracted feature are reduced FRS methods. For feature reduction PCA and LDA are used. Theses processed feature are then applied to hybrid case-based reasoning neuro-fuzzy (ANFIS) classification system to detect Parkinson’s disease. On the detection of PDs abnormality, the proposed healthcare monitoring system immediately generates notification to the patient simultaneously send detection report to centralized healthcare cloud system. This PDs detection report is further analyzed and stored at cloud server during analysis phase where report is analyzed by professional health expert and send the appropriate treatment and medication to PD infected person or care taker. For experimentation and performance evaluation benchmark baseline UCI dataset of PDs are used. We analyzed our proposed hybrid ANFIS-CBR classifier with existing classifiers over the accuracy, sensitivity and specificity parameter. Based on the result analysis, it is observed that proposed hybrid classifier maximum accuracy, sensitivity, and specificity of 98.23%, 99.1%, and 95.3% in comparison to other classifier.

Voice disorder classification using convolutional neural network based on deep transfer learning

Voice disorders are very common in the global population. Many researchers have conducted research on the identification and classification of voice disorders based on machine learning. As a data-driven algorithm, machine learning requires a large number of samples for training. However, due to the sensitivity and particularity of medical data, it is difficult to obtain sufficient samples for model learning. To address this challenge, this paper proposes a pretrained OpenL3-SVM transfer learning framework for the automatic recognition of multi-class voice disorders. The framework combines a pre-trained convolutional neural network, OpenL3, and a support vector machine (SVM) classifier. The Mel spectrum of the given voice signal is first extracted and then input into the OpenL3 network to obtain high-level feature embedding. Considering the effects of redundant and negative high-dimensional features, model overfitting easily occurs. Therefore, linear local tangent space alignment (LLTSA) is used for feature dimension reduction. Finally, the obtained dimensionality reduction features are used to train the SVM for voice disorder classification. Fivefold cross-validation is used to verify the classification performance of the OpenL3-SVM. The experimental results show that OpenL3-SVM can effectively classify voice disorders automatically, and its performance exceeds that of the existing methods. With continuous improvements in research, it is expected to be considered as auxiliary diagnostic tool for physicians in the future.

Voice pathology identification system using a deep learning approach based on unique feature selection sets

AbstractVoice pathology diagnosis requires extracting significant features from voice signals, and classical machine learning models can overfit to the training data, which can cause difficult issues and pose challenges. The study aimed to develop a reliable and efficient system for identifying voice pathologies utilizing the long short‐term memory (LSTM) method. The study combined unique feature sets such as the mel frequency cepstral coefficients (MFCCs), zero crossing rate (ZCR), and mel spectrograms, which have not been used together in previous works. Voice pathology identification improved the accuracy rate using the LSTM approach on the Saarbruecken voice database (SVD) samples. The best results achieved by the proposed system showed an accuracy rate of 99.3% for /u/ vowel samples in neutral pitch, 99.2% for /a/ vowel samples in high pitch, 99% for /i/ vowel samples in neutral pitch, and 99.2% for sentence samples. The experimental results were evaluated utilizing accuracy, precision, specificity, sensitivity, and F1 measures. Additionally, the study compared the performance of LSTM with that of artificial neural networks (ANNs) and found that LSTM achieved better outcomes.

PD-ADSV: An automated diagnosing system using voice signals and hard voting ensemble method for Parkinson’s disease

PD-ADSV: An automated diagnosing system using voice signals and hard voting ensemble method for Parkinson’s disease

Vowel classification with combining pitch detection and one‐dimensional convolutional neural network based classifier for gender identification

AbstractHuman speech signals may contain specific information regarding a speaker's characteristics, and these signals can be very useful in applications involving interactive voice response (IVR) and automatic speech recognition (ASR). For IVR and ASR applications, speaker classification into different ages and gender groups can be applied in human–machine interaction or computer‐based interaction systems for customised advertisement, translation (text generation), machine dialog systems, or self‐service applications. Hence, an IVR‐based system dictates that ASR should function through users' voices (specific voice‐frequency bands) to identify customers' age and gender and interact with a host system. In the present study, we intended to combine a pitch detection (PD)‐based extractor and a voice classifier for gender identification. The Yet Another Algorithm for Pitch Tracking (YAAPT)‐based PD method was designed to extract the voice fundamental frequency (F0) from non‐stationary speaker's voice signals, allowing us to achieve gender identification, by distinguishing differences in F0 between adult females and males, and classify voices into adult and children groups. Then, in vowel voice signal classification, a one‐dimensional (1D) convolutional neural network (CNN), consisted of a multi‐round 1D kernel convolutional layer, a 1D pooling process, and a vowel classifier that could preliminary divide feature patterns into three level ranges of F0, including adult and children groups. Consequently, a classifier was used in the classification layer to identify the speakers' gender. The proposed PD‐based extractor and voice classifier could reduce complexity and improve classification efficiency. Acoustic datasets were selected from the Hillenbrand database for experimental tests on 12 vowels classifications, and K‐fold cross‐validations were performed. The experimental results demonstrated that our approach is a very promising method to quantify the proposed classifier's performance in terms of recall (%), precision (%), accuracy (%), and F1 score.

Classification of Speaking and Singing Voices Using Bioimpedance Measurements and Deep Learning

The acts of speaking and singing are different phenomena displaying distinct characteristics. The classification and distinction of these voice acts is vastly approached utilizing voice audio recordings and microphones. The use of audio recordings, however, can become challenging and computationally expensive due to the complexity of the voice signal. The research presented in this paper seeks to address this issue by implementing a deep learning classifier of speaking and singing voices based on bioimpedance measurement in replacement of audio recordings. In addition, the proposed research aims to develop a real-time voice act classification for the integration with voice-to-MIDI conversion. For such purposes, a system was designed, implemented, and tested using electroglottographic signals, Mel Frequency Cepstral Coefficients, and a deep neural network. The lack of datasets for the training of the model was tackled by creating a dedicated dataset 7200 bioimpedance measurement of both singing and speaking. The use of bioimpedance measurements allows to deliver high classification accuracy whilst keeping low computational needs for both preprocessing and classification. These characteristics, in turn, allows a fast deployment of the system for near-real-time applications. After the training, the system was broadly tested achieving a testing accuracy of 92% to 94%.

Hybrid optimization enabled deep learning model for Parkinson's disease classification

ABSTRACT The analysis of Parkinson's disease (PD) is an inspiring task that necessitates the analysis of numerous motor and non-motor indications. During analysis, some abnormalities are considered as important symptoms to analyze the disease. Hence, this research introduced the proposed chronological smart sunflower optimization Algorithm (CSSFOA) for classifying the PD from voice data and voice signal samples. For voice signal, the input signals are pre-processed by the Gaussian filter, and then the significant features are extracted from it. The selection of optimal features is done by the chronological smart flower optimization Algorithm (CSFOA). The proposed CSFOA-based feature selection method considered the features by the Bray–Curtis distance. The PD classification is done by the ZF-Net which is trained by proposed CSSFOA to increase the performance of classification. The experimental result reveals that the proposed CSSFOA_ZF-Net algorithm got a better testing accuracy of 0.945, a sensitivity of 0.919, and a specificity of 0.957.

Paralinguistic Privacy Protection at the Edge

Voice user interfaces and digital assistants are rapidly entering our lives and becoming singular touch points spanning our devices. These always-on services capture and transmit our audio data to powerful cloud services for further processing and subsequent actions. Our voices and raw audio signals collected through these devices contain a host of sensitive paralinguistic information that is transmitted to service providers regardless of deliberate or false triggers. As our emotional patterns and sensitive attributes like our identity, gender, and well-being are easily inferred using deep acoustic models, we encounter a new generation of privacy risks by using these services. One approach to mitigate the risk of paralinguistic-based privacy breaches is to exploit a combination of cloud-based processing with privacy-preserving, on-device paralinguistic information learning and filtering before transmitting voice data. In this article we introduce EDGY , a configurable, lightweight, disentangled representation learning framework that transforms and filters high-dimensional voice data to identify and contain sensitive attributes at the edge prior to offloading to the cloud. We evaluate EDGY’s on-device performance and explore optimization techniques, including model quantization and knowledge distillation, to enable private, accurate, and efficient representation learning on resource-constrained devices. Our results show that EDGY runs in tens of milliseconds with 0.2% relative improvement in “zero-shot” ABX score or minimal performance penalties of approximately 5.95% word error rate (WER) in learning linguistic representations from raw voice signals, using a CPU and a single-core ARM processor without specialized hardware.

Speech to Text Conversion System for Myanmar Alphabet

This paper is aimed to implement the speech to text conversion system for Myanmar alphabet. The Myanmar alphabet consists of 33 characters from ?ka? to ?ah?. The proposed system is software architecture which allows the user to speak against the computer in Myanmar language and the corresponding character is printed on the screen in the Microsoft Office Word Document Format. The system is emphasized on Speaker Independent Isolated Word Recognition System. The proposed system directly acquires and converts speech to text. This system contains two main modules: feature extraction and feature matching. Mel Frequency Cepstrum Coefficients (MFCC) is applied for feature extraction which extracts a small amount of data from the voice signal that can later be used to represent each character. Feature matching involves the actual procedure to identify the unknown character by comparing extracted features from the voice inputs of a set of known characters. In this system, Vector Quantization (VQ) approach using Linde, Buzo and Gray (LBG) clustering algorithm, which reduces the amount of data and complexity, is applied for feature matching. To implement this system MATLAB programming language is used.

Visual Deprivation Alters Functional Connectivity of Neural Networks for Voice Recognition: A Resting-State fMRI Study

Humans recognize one another by identifying their voices and faces. For sighted people, the integration of voice and face signals in corresponding brain networks plays an important role in facilitating the process. However, individuals with vision loss primarily resort to voice cues to recognize a person's identity. It remains unclear how the neural systems for voice recognition reorganize in the blind. In the present study, we collected behavioral and resting-state fMRI data from 20 early blind (5 females; mean age = 22.6 years) and 22 sighted control (7 females; mean age = 23.7 years) individuals. We aimed to investigate the alterations in the resting-state functional connectivity (FC) among the voice- and face-sensitive areas in blind subjects in comparison with controls. We found that the intranetwork connections among voice-sensitive areas, including amygdala-posterior "temporal voice areas" (TVAp), amygdala-anterior "temporal voice areas" (TVAa), and amygdala-inferior frontal gyrus (IFG) were enhanced in the early blind. The blind group also showed increased FCs of "fusiform face area" (FFA)-IFG and "occipital face area" (OFA)-IFG but decreased FCs between the face-sensitive areas (i.e., FFA and OFA) and TVAa. Moreover, the voice-recognition accuracy was positively related to the strength of TVAp-FFA in the sighted, and the strength of amygdala-FFA in the blind. These findings indicate that visual deprivation shapes functional connectivity by increasing the intranetwork connections among voice-sensitive areas while decreasing the internetwork connections between the voice- and face-sensitive areas. Moreover, the face-sensitive areas are still involved in the voice-recognition process in blind individuals through pathways such as the subcortical-occipital or occipitofrontal connections, which may benefit the visually impaired greatly during voice processing.

Pathological voice classification based on the features of an asymmetric fluid–structure interaction vocal cord model

Currently, the acoustic features used for classifying pathological voices do not take the physiological structure and vibration characteristics of the vocal cords into account. We proposed a feature extraction method based on the asymmetric fluid–structure interaction vocal cord model. The airflow viscous force was introduced into this model. The left and right asymmetric coefficients of the upper mass and the lower mass were designed to characterize the vibration mechanism of the diseased vocal cord. Then a hybrid simulated annealing genetic algorithm (HSAGA) optimized the vocal cord model inversion parameters to match the target voice source. The vocal cord model parameters corresponding to the actual voice signal were extracted and combined with the phonation pressure threshold and the phonation flow threshold to classify pathological voices. The experimental results (deposited in the Massachusetts Eye and Ear Infirmary (MEEI) database) showed that the weighted average relative error of the objective optimization function classifying the vibration characteristics of the vocal cords with lesions calculated by the HSAGA was 4.0%. When classifying pathological voices, the accuracy of the vocal cord model parameters was 77.18%, which is 33.18% higher than classifications using acoustic features, and effectively distinguishes different types of pathological voice.

SPEECH TO TEXT BAHASA SASAK MENGGUNAKAN EXTRAKSI FITUR MEL-FREQUENCY CEPSTRAL COEFFICIENTS DAN KLASIFIKASI CONVOLUTIONAL NEURAL NETWORKS

Artificial intelligence technology allows digital signals to be processed by computers. Currently speech to text is available only in Indonesian and English versions. Speech to text is a system that performs commands from human voice input and is then translated into words. The development of speech to text in regional languages ​​is needed because it can be a bridge between culture and technological progress. From 5 research literature, it was found that the mel-frequency cepstral coefficients (MFCC) and convolutional neural networks (CNN) methods are a combination of the commonly used voice signal analysis methods and get accuracy between 70.00% to 99.00%. This study uses the CNN and MFCC methods in the speech to text field to recognize the Sasak language and convert it into text. The result of this research is a real time conversion system from voice to text in Sasak language. The analysis carried out includes determining the best amount of training data, testing the training data on the number of votes based on accuracy, the sensitivity of the algorithm to words that have similar prefixes using the MFCC method as feature extraction and CNN as a classifier for the voice dataset. This study aims to obtain the accuracy of the dataset used and the sensitivity of the algorithm to sentences that have similarities. In this study got 2 results. The first result is the result of training with the accuracy of CNN training is 90% and loss is 0.5%. The second result is the result of an experiment using 3 voice samples for each word on the dateset with 43 correct words, 6 correct words 2, 1 correct word 1 and none of the words incorrect. So it has a percentage of 86% all correct, 12% correct 2, and 2% correct 1, and 0% all wrong.

Small-size spectral features for machine learning in voice signal analysis and classification tasks

Objectives. The problem of developing a method for calculating small-sized spectral features that increases the efficiency of existing machine learning systems for analyzing and classifying voice signals is being solved.Methods. Spectral features are extracted using a generative approach, which involves calculating a discrete Fourier spectrum for a sequence of samples generated using an autoregressive model of input voice signal. The generated sequence processed by the discrete Fourier transform considers the periodicity of the transform and thereby increase the accuracy of spectral estimation of analyzed signal.Results. A generative method for calculating spectral features intended for use in machine learning systems for the analysis and classification of voice signals is proposed and described. An experimental analysis of the accuracy and stability of the spectrum representation of a test signal with a known spectral composition has been carried out using the envelopes. The envelopes were calculated using proposed generative method and using discrete Fourier transform with different analysis windows (rectangular window and Hanna window). The analysis showed that spectral envelopes obtained using the proposed method more accurately represent the spectrum of test signal according to the criterion of minimum square error. A comparison of the effectiveness of voice signal classification with proposed features and the features based on the mel-frequency kepstral coefficients is carried out. A diagnostic system for amyotrophic lateral sclerosis was used as a basic test system to evaluate the effectiveness of proposed approach in practice. Conclusion. The obtained experimental results showed a significant increase of classification accuracy when using proposed approach for calculating features compared with the features based on the mel-frequency kepstral coefficients.

Deep Learning-Based Speech Enhancement of an Extrinsic Fabry-Perot Interferometric Fiber Acoustic Sensor System.

To achieve high-quality voice communication technology without noise interference in flammable, explosive and strong electromagnetic environments, the speech enhancement technology of a fiber-optic external Fabry-Perot interferometric (EFPI) acoustic sensor based on deep learning is studied in this paper. The combination of a complex-valued convolutional neural network and a long short-term memory (CV-CNN-LSTM) model is proposed for speech enhancement in the EFPI acoustic sensing system. Moreover, the 3 × 3 coupler algorithm is used to demodulate voice signals. Then, the short-time Fourier transform (STFT) spectrogram features of voice signals are divided into a training set and a test set. The training set is input into the established CV-CNN-LSTM model for model training, and the test set is input into the trained model for testing. The experimental findings reveal that the proposed CV-CNN-LSTM model demonstrates exceptional speech enhancement performance, boasting an average Perceptual Evaluation of Speech Quality (PESQ) score of 3.148. In comparison to the CV-CNN and CV-LSTM models, this innovative model achieves a remarkable PESQ score improvement of 9.7% and 11.4%, respectively. Furthermore, the average Short-Time Objective Intelligibility (STOI) score witnesses significant enhancements of 4.04 and 2.83 when contrasted with the CV-CNN and CV-LSTM models, respectively.