Mel Frequency Research Articles

Time-frequency analysis of speech signal using Chirplet transform for automatic diagnosis of Parkinson's disease.

Parkinson's disease (PD) is the second most prevalent neurodegenerative disorder in the world after Alzheimer's disease. Early diagnosing PD is challenging as it evolved slowly, and its symptoms eventuate gradually. Recent studies have demonstrated that changes in speech may be utilized as an excellent biomarker for the early diagnosis of PD. In this study, we have proposed a Chirplet transform (CT) based novel approach for diagnosing PD using speech signals. We employed CT to get the time-frequency matrix (TFM) of each speech recording, and we extracted time-frequency based entropy (TFE) features from the TFM. The statistical analysis demonstrates that the TFE features reflect the changes in speech that occurs in the speech due to PD, hence can be used for classifying the PD and healthy control (HC) individuals. The effectiveness of the proposed framework is validated using the vowels and words from the PC-GITA database. The genetic algorithm is utilized to select the optimum features subset, while a support vector machine (SVM), decision tree (DT), K-Nearest Neighbor (KNN), and Naïve Bayes (NB) classifiers are employed for classification. The TFE features outperform the breathiness and Mel frequency cepstral coefficients (MFCC) features. The SVM classifier is most effective compared to other machine-learning classifiers. The highest classification accuracy rates of 98% and 99% are achieved using the vowel /a/ and word /atleta/, respectively. The results reveal that the proposed CT-based entropy features effectively diagnose PD using the speech of a person.

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.

Multichannel CNN-BLSTM Architecture for Speech Emotion Recognition System by Fusion of Magnitude and Phase Spectral Features Using DCCA for Consumer Applications

Conventional Speech Emotion Recognition (SER) approaches put more emphasis on extracting magnitude spectrum-based features, such as Mel Frequency Cepstral Coefficients (MFCCs), and Mel spectrogram. However, phase information is ignored due to signal processing difficulties such as the phase wrapping issue. This work develops a multichannel Convolution Neural Network-Bidirectional Long Short Term Memory (CNN-BLSTM) architectures with an attention mechanism for speaker-independent SER by considering phase and magnitude spectrum-based features. The phase-based features are extracted using the Modified Group Delay Function (MODGD). The obtained phase features are combined with MFCC features. The CNN-BLSTM network extract learned representation from magnitude and phase features. The learned representation from MFCCs and MODGD are combined and given as an input to the Support Vector Machine (SVM) for classification. The Deep Canonical Correlation Analysis (DCCA) is used to maximize the correlation between magnitude and phase information to improve the conventional SER system’s performance. Here the IEMOCAP database is used for performance analysis. The experimental results show improvement over MFCC features and existing approaches for unimodal SER. In this work, we also developed real-time web server application for the proposed architecture.

Raga classification using enhanced spatial bound whale optimization algorithm

A raga is a unique set of notes with certain rules that carefully followed, retain and protect its purity and produce amazing musical effects. An automated raga transcription and identification is important for computational musicology, which is an important step for musicology for indexing, classifying, and recommending tunes. In the present research, the audio features such as mel frequency cepstrum coefficients (MFCCs), spectral flux, short time energy, audio feature extractor, and spectral centroid features are used for the prediction of a raga. The model showed more complexity which means it required lots of training data. The proposed enhanced spatial bound whale optimization algorithm (ESBWOA) is used that overcome the feature selection problem of high dimensional features. In addition to this, a weighted salp swarm algorithm (SSA) is used for selecting the tone-based features from the ragas based on amplitude or each raga sample. The features were fed for bidirectional long short-term memory (Bi-LSTM) network, which enhanced the success rate for raga identification and classification. The present research uses CompMusic dataset in the research work where 9 classes for Carnatic music and 7 classes in Hindustani music are considered for the classification of ragas.

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%.

Song Classification using Machine Learning

Abstract: The classification of music by genre is crucial in the modern world since the number of music tracks, both online and offline, is growing quickly. We must appropriately index them in order to have greater access to them. To retrieve music from a vast collection, automatic music genre classification is crucial. The majority of the current methods for categorising music genres rely on machine learning. We give a music dataset with ten distinct genres in this article. The system is trained and classified using a Deep Learning technique. Convolution neural networks are employed in this instance for training and classification. For audio analysis, feature extraction is the most important step. For sound samples, the Mel Frequency Cepstral Coefficient (MFCC) is employed as a feature vector. The suggested technique uses feature vector extraction to categorise music into different genres. Our findings indicate that our system's accuracy level is approximately 76%, which will significantly increase and facilitate the automatic classification of musical genres.

Interfacial debonding detection of steel beams reinforced by CFRP plates based on percussion method

Steel structure reinforced engineering makes extensive use of carbon fiber reinforced plastic (CFRP) plates. Although it happens frequently, the interfacial debonding of CFRP plate-reinforced steel beams has been the main cause of failure in flexural strengthened steel beams, which results in the brittle failure of steel structures. This paper proposes a novel percussion method to locate and quantify the debonding area based on the Welch power spectrum density (PSD) estimate, mel frequency cepstral coefficient (MFCC), and convolutional neural network (CNN). According to our knowledge, this is the first attempt to use the percussion method to identify and measure the debonding defects of steel beams strengthened by CFRP plates. Additionally, this experiment takes into account how percussion instruments affect localization and quantification. To begin with, the debonding location is qualitatively detected using the Welch PSD estimate method. The presented percussion method can successfully and accurately detect the interfacial debonding of CFRP plate-reinforced steel beams, according to repeated results. Subsequently, a novel CNN model is constructed to classify the various debonding areas based on the MFCC feature. In the end, the experimental findings show that the localization and quantification of the hammer types vary only slightly. Overall, the percussion-based approach can offer a fresh line of inquiry into the debonding problem on steel beams reinforced with CFRP plates.

Comparative analysis of various feature extraction techniques for classification of speech disfluencies

Speech plays a vital role in communication, from expressing oneself, to utilizing speech-based platforms, speech is a necessity. Any disruption in speech is referred to as disfluency, and can impact one's quality of life. This paper presents an experimental study on various techniques for the detection and classification of speech disfluencies. Six different types of disfluencies are examined in this paper, namely Interjection, Sound Repetition, Word Repetition, Phrase Repetition, Revision and Prolongation (6 classes). However, this paper also goes a step further by including the clean speech signals as an added class alongside the six disfluencies, thereby making this work more robust with 7 classes. Various machine learning approaches have been investigated on the University College London Archive of Stuttered Speech (UCLASS) dataset; a standard disfluency dataset generated by University College London (UCL). Five different feature extraction techniques viz. Mel Frequency Cepstral Coefficients (MFCC), Linear Predictive Cepstral Coefficients (LPCC), Gammatone Frequency Cepstral Coefficients (GFCC), Mel-filterbank energy features, and Spectrograms have been used. Comparative analysis of various classifiers shows that MFCC, GFCC, and Spectrograms achieved greater than 90% accuracy on both 6 and 7 classes with the kNN classifier. As a future scope to this study, the authors aim to focus on tackling the challenges of detecting multiple disfluencies present simultaneously in a speech sample.

A hierarchical birdsong feature extraction architecture combining static and dynamic modeling

To conserve bird biodiversity and monitor the distribution of species in the region, it is of tremendous necessity to identify birds by their songs and explore the rich ecological information birdsong contains. The audios recorded in the monitoring area generally have complex background noise, the characteristics of the song are not prominent and the biological spectrum information is not comprehensive, which brings some challenges to the identification of birds. This study proposes a hierarchical birdsong feature extraction architecture combining dynamic and static modeling to cope with complex environments as a modeling context. Firstly, six common speech features were extracted for the characteristics of birdsong. The Pearson correlation coefficient is then used to analyze the correlations between birdsong and human speech, examining the correlations between each feature in the presence and absence of environmental noise interference. Combined with the scatter plot matrix analysis, we conclude that Mel Frequency Cepstral Coefficient (MFCC) is more suitable comparing with other features when dealing with birdsong and can superiorly cope with a complex background noise. Secondly, a feature extraction architecture is built, which integrates static and dynamic modeling to fully explore the contextual relationship, to solve the problem of ignoring the internal structure information of the patch and losing some spatial information in the Transformer-type model. Finally, a hierarchical refinement module is designed to help extract more detailed features, as well as to optimize the computational cost of the Transformer-type model that requires many training data and has high complexity. The performance of the model can be detected with 93.67 % accuracy on the self-built birdsong dataset, 95.19 % accuracy on the public birdsong dataset Birdsdata and 97.02 % accuracy on the public environmental dataset UrbanSound8k.

Coal-gangue recognition via multi-branch convolutional neural network based on MFCC in noisy environment

Traditional coal-gangue recognition methods usually do not consider the impact of equipment noise, which severely limits its adaptability and recognition accuracy. This paper mainly studies the more accurate recognition of coal-gangue in the noise site environment with the operation of shearer, conveyor, transfer machine and other device in the process of top coal caving. Mel Frequency Cepstrum Coefficients (MFCC) smoothing method was introduced to express the intrinsic feature of sound pressure more clearly in the coal-gangue recognition site. Then, a multi-branch convolution neural network (MBCNN) model with three branches was developed, and the smoothed MFCC feature was incorporated into this model to realize the recognition of falling coal and gangue in noisy environment. The sound pressure signal datasets under the operation of different device were constructed through a great deal of laboratory and site data acquisition. Comparative experiments were carried out on noiseless dataset, single noise dataset and simulated site dataset, and the results show that our method can provide higher correct recognition accuracy and better robustness. The proposed coal-gangue recognition approach based on MBCNN and MFCC smoothing can not only recognize the state of falling coal or gangue, but also recognize the operational state of site device.

Detecting somatisation disorder via speech: introducing the Shenzhen Somatisation Speech Corpus

ObjectiveSpeech recognition technology is widely used as a mature technical approach in many fields. In the study of depression recognition, speech signals are commonly used due to their convenience and ease of acquisition. Though speech recognition is popular in the research field of depression recognition, it has been little studied in somatisation disorder recognition. The reason for this is the lack of a publicly accessible database of relevant speech and benchmark studies. To this end, we introduced our somatisation disorder speech database and gave benchmark results. MethodsBy collecting speech samples of somatisation disorder patients, in cooperation with the Shenzhen University General Hospital, we introduced our somatisation disorder speech database, the Shenzhen Somatisation Speech Corpus (SSSC). Moreover, a benchmark for SSSC using classic acoustic features and a machine learning model was proposed in our work. ResultsTo obtain a more scientific benchmark, we compared and analysed the performance of different acoustic features, i. e., the full ComPare feature set, or only Mel frequency cepstral coefficients (MFCCs), fundamental frequency (F0), and frequency and bandwidth of the formants (F1-F3). By comparison, the best result of our benchmark was the 76.0% unweighted average recall achieved by a support vector machine with formants F1–F3. ConclusionThe proposal of SSSC may bridge a research gap in somatisation disorder, providing researchers with a publicly accessible speech database. In addition, the results of the benchmark could show the scientific validity and feasibility of computer audition for speech recognition in somatization disorders.

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.

Speech Emotion Recognition through Hybrid Features and Convolutional Neural Network

Speech emotion recognition (SER) is the process of predicting human emotions from audio signals using artificial intelligence (AI) techniques. SER technologies have a wide range of applications in areas such as psychology, medicine, education, and entertainment. Extracting relevant features from audio signals is a crucial task in the SER process to correctly identify emotions. Several studies on SER have employed short-time features such as Mel frequency cepstral coefficients (MFCCs), due to their efficiency in capturing the periodic nature of audio signals. However, these features are limited in their ability to correctly identify emotion representations. To solve this issue, this research combined MFCCs and time-domain features (MFCCT) to enhance the performance of SER systems. The proposed hybrid features were given to a convolutional neural network (CNN) to build the SER model. The hybrid MFCCT features together with CNN outperformed both MFCCs and time-domain (t-domain) features on the Emo-DB, SAVEE, and RAVDESS datasets by achieving an accuracy of 97%, 93%, and 92% respectively. Additionally, CNN achieved better performance compared to the machine learning (ML) classifiers that were recently used in SER. The proposed features have the potential to be widely utilized to several types of SER datasets for identifying emotions.

Intra- and inter-speaker variation in eight Russian fricativesa).

Acoustic variation is central to the study of speaker characterization. In this respect, specific phonemic classes such as vowels have been particularly studied, compared to fricatives. Fricatives exhibit important aperiodic energy, which can extend over a high-frequency range beyond that conventionally considered in phonetic analyses, often limited up to 12 kHz. We adopt here an extended frequency range up to 20.05 kHz to study a corpus of 15 812 fricatives produced by 59 speakers in Russian, a language offering a rich inventory of fricatives. We extracted two sets of parameters: the first is composed of 11 parameters derived from the frequency spectrum and duration (acoustic set) while the second is composed of 13 mel frequency cepstral coefficients (MFCCs). As a first step, we implemented machine learning methods to evaluate the potential of each set to predict gender and speaker identity. We show that gender can be predicted with a good performance by the acoustic set and even more so by MFCCs (accuracy of 0.72 and 0.88, respectively). MFCCs also predict individuals to some extent (accuracy = 0.64) unlike the acoustic set. In a second step, we provide a detailed analysis of the observed intra- and inter-speaker acoustic variation.

Wavelet-based Parametric Feature Subset Selection for Speaker and Accent Recognition using Genetic Algorithm

Research on speaker and accent recognition studies using the Malay language in the field of Automatic Speech Recognition (ASR) is limited, with most studies focusing on speech recognition. This study proposes to increase the performance Malaysian speakers and accent recognition using wavelets transform, namely Wavelet Packet Transform (WPT) and Dual-Tree Complex Wavelet Packet Transform (DT-CWPT). A variety of feature extraction combinations, including conventional Mel Frequency Cepstral Coefficients (MFCC), Linear Predictive Coding (LPC) and wavelets transform, were implemented to compare the effectiveness of the proposed method. Although the proposed approach resulted in improved detection rate, it faced challenges in terms of high feature dimensionality and increased computation time. To address these issues, the Genetic Algorithm (GA) approach has been adopted to reduce the number of irrelevant features, accelerate the learning system and achieve better performance. The extracted features were trained using various classifiers, including k-Nearest Neighbors (k-NN), Support Vector Machine (SVM) and Extreme Learning Machine (ELM). The experimental results showed that the best speaker recognition accuracy was 97.33% for English numbers using SVM classifier and 96.02% for Malay words using the ELM classifier with a combination of wavelets, LPC and MFCC features. For accent recognition, the ELM classifier yielded the best performance, achieving 95.28% accuracy for English numbers with a combination of wavelets and MFCC features and 96.72% for Malay words using combined feature extraction of wavelets, LPC and MFCC feature extraction. It can be concluded that Malay words yielded better recognition rates than English numbers. Furthermore, use of GA effectively reduced the overall number of features while maintaining high accuracy level.

Extreme-Low-Speed Heavy Load Bearing Fault Diagnosis by Using Improved RepVGG and Acoustic Emission Signals.

With the worldwide carbon neutralization boom, low-speed heavy load bearings have been widely used in the field of wind power. Bearing failure generates impulses when the rolling element passes the cracked surface of the bearing. Over the past decade, acoustic emission (AE) techniques have been used to detect failure signals. However, the high sampling rates of AE signals make it difficult to design and extract fault features; thus, deep neural network-based approaches have been proposed. In this paper, we proposed an improved RepVGG bearing fault diagnosis technique. The normalized and noise-reduced bearing signals were first converted into Mel frequency cepstrum coefficients (MFCCs) and then inputted into the model. In addition, the exponential moving average method was used to optimize the model and improve its accuracy. Data were extracted from the test bench and wind turbine main shaft bearing. Four damage classes were studied experimentally. The experimental results demonstrated that the improved RepVGG model could be employed for classifying low-speed heavy load bearing states by using MFCCs. Furthermore, the effectiveness of the proposed model was assessed by performing comparisons with existing models.

Enhancement in Speaker Identification through Feature Fusion using Advanced Dilated Convolution Neural Network

There are various challenges in identifying the speakers accurately. The Extraction of discriminative features is a vital task for accurate identification in the speaker identification task. Nowadays, speaker identification is widely investigated using deep learning. The complex and noisy speech data affects the performance of Mel Frequency Cepstral Coefficients (MFCC); hence, MFCC fails to represent the speaker characteristics accurately. In this proposed work, a novel text-independent speaker identification system is developed to enhance the performance by fusion of Log-MelSpectrum and excitation features. The excitation information is obtained due to the vibration of vocal folds, and it is represented using Linear Prediction (LP) residual. The various types of features extracted from the excitation are residual phase, sharpness, Energy of Excitation (EoE), and Strength of Excitation (SoE). The extracted features were processed with the dilated convolution neural network (dilated CNN) to fulfill the identification task. The extensive evaluation showed that the fusion of excitation features gives better results than the existing methods. The accuracy reaches 94.12% for 11 complex classes and 91.34% for 80 speakers, and Equal Error Rate (EER) is reduced to 1.16% for the proposed model. The proposed model is tested with the Librispeech corpus using Matlab 2021b tool, outperforming the existing baseline models. The proposed model achieves an accuracy improvement of 1.34% compared to the baseline system.

Machine learning approach for detecting Covid-19 from speech signal using Mel frequency magnitude coefficient.

The Covid-19 pandemic is one of the most significant global health concerns that have emerged in this decade. Intelligent healthcare technology and techniques based on speech signal and artificial intelligence make it feasible to provide a faster and more efficient timely detection of Covid-19. The main objective of our study is to design speech signal-based noninvasive, low-cost, remote diagnosis of Covid-19. In this study, we have developed system to detect Covid-19 from speech signal using Mel frequency magnitude coefficients (MFMC) and machine learning techniques. In order to capture higher-order spectral features, the spectrum is divided into a larger number of subbands with narrower bandwidths as MFMC, which leads to better frequency resolution and less overall noise. As a consequence of an improvement in frequency resolution as well as a decrease in the quantity of noise that is included with the extraction of MFMC, the higher-order MFMCs are able to identify Covid-19 from speech signals with an increased level of accuracy. The procedures for machine learning are often less complicated than those for deep learning, and they may commonly be carried out on regular computers. However, deep learning systems need extensive computing power and data storage. Twelve, twenty-four, thirty, and forty spectral coefficients are obtained using MFMC in our study, and from these coefficients, performance is accessed using machine learning classifiers, such as random forests and K-nearest neighbor (KNN); however, KNN has performed better than the other model with having AUC score of 0.80.

Fault diagnosis of transformer winding short circuit based on WKPCA-WM and IPOA-CNN

When the winding of the power transformer is short-circuited, the winding will experience constant vibration, which will cause axial instability of the winding, and then lead to winding looseness, deformation, bulge, etc., therefore, a diagnosis method based on the Improved Pelican Optimization Algorithm and Convolutional Neural Network (IPOA-CNN) for short-circuit voiceprint signal of transformer windings is proposed. At the same time, considering the input parameter dimension of deep learning cannot be too high, a new feature parameter selection method is constructed for this model. Firstly, the frequency characteristics of winding acoustic vibration signals are analyzed, and then the characteristic parameters of transformer acoustic signals are extracted by Wavelet Packet Energy Spectrum (WPES) and Mel Frequency Cepstrum Coefficient (MFCC), respectively. Then, the two methods are combined to construct the WM feature extraction algorithm, and the Weighted Kernel Principal Component Analysis (WKPCA) is used to reduce the dimension of the feature to obtain the feature parameters with accurate feature information and low redundancy; Finally, combined with Sobol sequence to optimize the initial population of Pelican Optimization Algorithm (POA), the convolution kernel of Convolutional neural network (CNN) was optimized by IPOA, and the optimal convolution kernel was obtained. The transformer winding short-circuits voiceprint diagnosis models of WKPCA-WM and IPOA-CNN were constructed, which realized the accurate diagnosis of winding short-circuit voiceprint. The validity and feasibility of the method are verified by the acoustic signal data collected in the laboratory.

Ensemble learning with speaker embeddings in multiple speech task stimuli for depression detection.

As a biomarker of depression, speech signal has attracted the interest of many researchers due to its characteristics of easy collection and non-invasive. However, subjects' speech variation under different scenes and emotional stimuli, the insufficient amount of depression speech data for deep learning, and the variable length of speech frame-level features have an impact on the recognition performance. The above problems, this study proposes a multi-task ensemble learning method based on speaker embeddings for depression classification. First, we extract the Mel Frequency Cepstral Coefficients (MFCC), the Perceptual Linear Predictive Coefficients (PLP), and the Filter Bank (FBANK) from the out-domain dataset (CN-Celeb) and train the Resnet x-vector extractor, Time delay neural network (TDNN) x-vector extractor, and i-vector extractor. Then, we extract the corresponding speaker embeddings of fixed length from the depression speech database of the Gansu Provincial Key Laboratory of Wearable Computing. Support Vector Machine (SVM) and Random Forest (RF) are used to obtain the classification results of speaker embeddings in nine speech tasks. To make full use of the information of speech tasks with different scenes and emotions, we aggregate the classification results of nine tasks into new features and then obtain the final classification results by using Multilayer Perceptron (MLP). In order to take advantage of the complementary effects of different features, Resnet x-vectors based on different acoustic features are fused in the ensemble learning method. Experimental results demonstrate that (1) MFCC-based Resnet x-vectors perform best among the nine speaker embeddings for depression detection; (2) interview speech is better than picture descriptions speech, and neutral stimulus is the best among the three emotional valences in the depression recognition task; (3) our multi-task ensemble learning method with MFCC-based Resnet x-vectors can effectively identify depressed patients; (4) in all cases, the combination of MFCC-based Resnet x-vectors and PLP-based Resnet x-vectors in our ensemble learning method achieves the best results, outperforming other literature studies using the depression speech database. Our multi-task ensemble learning method with MFCC-based Resnet x-vectors can fuse the depression related information of different stimuli effectively, which provides a new approach for depression detection. The limitation of this method is that speaker embeddings extractors were pre-trained on the out-domain dataset. We will consider using the augmented in-domain dataset for pre-training to improve the depression recognition performance further.