Speaker Embedding Research Articles

Speech Emotion Recognition Using Transfer Learning: Integration of Advanced Speaker Embeddings and Image Recognition Models

Automatic Speech Emotion Recognition (SER) plays a vital role in making human–computer interactions more natural and effective. A significant challenge in SER development is the limited availability of diverse emotional speech datasets, which hinders the application of advanced deep learning models. Transfer learning is a machine learning technique that helps address this issue by utilizing knowledge from pre-trained models to improve performance on a new task in a target domain, even with limited data. This study investigates the use of transfer learning from various pre-trained networks, including speaker embedding models such as d-vector, x-vector, and r-vector, and image classification models like AlexNet, GoogLeNet, SqueezeNet, ResNet-18, and ResNet-50. We also propose enhanced versions of the x-vector and r-vector models incorporating Multi-Head Attention Pooling and Angular Margin Softmax, alongside other architectural improvements. Additionally, reverberation from the Room Impulse Response datasets was added to the speech utterances to diversify and augment the available data. Notably, the enhanced r-vector model achieved classification accuracies of 74.05% Unweighted Accuracy (UA) and 73.68% Weighted Accuracy (WA) on the IEMOCAP dataset, and 80.25% UA and 79.81% WA on the CREMA-D dataset, outperforming the existing state-of-the-art methods. This study shows that using cross-domain transfer learning is beneficial for low-resource emotion recognition. The enhanced models developed in other domains (for non-emotional tasks) can further improve the accuracy of SER.

Which to select?: Analysis of speaker representation with graph attention networks.

Although the recent state-of-the-art systems show almost perfect performance, analysis of speaker embeddings has been lacking thus far. An in-depth analysis of speaker representation will be performed by looking into which features are selected. To this end, various intermediate representations of the trained model are observed using graph attentive feature aggregation, which includes a graph attention layer and graph pooling layer followed by a readout operation. To do so, the TIMIT dataset, which has comparably restricted conditions (e.g., the region and phoneme) is used after pre-training the model on the VoxCeleb dataset and then freezing the weight parameters. Through extensive experiments, there is a consistent trend in speaker representation in that the models learn to exploit sequence and phoneme information despite no supervision in that direction. The results shed light to help understand speaker embedding, which is yet considered to be a black box.

Zero-shot voice conversion based on feature disentanglement

Voice conversion (VC) aims to convert the voice from a source speaker to a target speaker without modifying the linguistic content. Zero-shot voice conversion has attracted significant attention in the task of VC because it can achieve conversion for speakers who did not appear during the training stage. Despite the significant progress made by previous methods in zero-shot VC, there is still room for improvement in separating speaker information and content information. In this paper, we propose a zero-shot VC method based on feature disentanglement. The proposed model uses a speaker encoder for extracting speaker embeddings, introduces mixed speaker layer normalization to eliminate residual speaker information in content encoding, and employs adaptive attention weight normalization for conversion. Furthermore, dynamic convolution is introduced to improve speech content modeling while requiring a small number of parameters. The experiments demonstrate that performance of the proposed model is superior to several state-of-the-art models, achieving both high similarity with the target speaker and intelligibility. In addition, the decoding speed of our model is much higher than the existing state-of-the-art models.

CLESSR-VC: Contrastive learning enhanced self-supervised representations for one-shot voice conversion

One-shot voice conversion (VC) has attracted more and more attention due to its broad prospects for practical application. In this task, the representation ability of speech features and the model’s generalization are the focus of attention. This paper proposes a model called CLESSR-VC, which enhances pre-trained self-supervised learning (SSL) representations through contrastive learning for one-shot VC. First, SSL features from the 23rd and 9th layers of the pre-trained WavLM are adopted to extract content embedding and SSL speaker embedding, respectively, to ensure the model’s generalization. Then, the conventional acoustic feature mel-spectrograms and contrastive learning are introduced to enhance the representation ability of speech features. Specifically, contrastive learning combined with the pitch-shift augmentation method is applied to disentangle content information from SSL features accurately. Mel-spectrograms are adopted to extract mel speaker embedding. The AM-Softmax and cross-architecture contrastive learning are applied between SSL and mel speaker embeddings to obtain the fused speaker embedding that helps improve speech quality and speaker similarity. Both objective and subjective evaluation results on the VCTK corpus confirm that the proposed VC model has outstanding performance and few trainable parameters.

Speaker-Attributed Training for Multi-Speaker Speech Recognition Using Multi-Stage Encoders and Attention-Weighted Speaker Embedding

Automatic speech recognition (ASR) aims at understanding naturally spoken human speech to be used as text inputs to machines. In multi-speaker environments, where multiple speakers are talking simultaneously with a large amount of overlap, a significant performance degradation may occur with conventional ASR systems if they are trained by recordings of single talkers. This paper proposes a multi-speaker ASR method that incorporates speaker embedding information as an additional input. The embedding information for each of the speakers in the training set was extracted as numeric vectors, and all of the embedding vectors were stacked to construct a total speaker profile matrix. The speaker profile matrix from the training dataset enables finding embedding vectors that are close to the speakers of the input recordings in the test conditions, and it helps to recognize the individual speakers’ voices mixed in the input. Furthermore, the proposed method efficiently reuses the decoder from the existing speaker-independent ASR model, eliminating the need for retraining the entire system. Various speaker embedding methods such as i-vector, d-vector, and x-vector were adopted, and the experimental results show 0.33% and 0.95% absolute (3.9% and 11.5% relative) improvements without and with the speaker profile in the word error rate (WER).

Multisensory Fusion for Unsupervised Spatiotemporal Speaker Diarization.

Speaker diarization consists of answering the question of "who spoke when" in audio recordings. In meeting scenarios, the task of labeling audio with the corresponding speaker identities can be further assisted by the exploitation of spatial features. This work proposes a framework designed to assess the effectiveness of combining speaker embeddings with Time Difference of Arrival (TDOA) values from available microphone sensor arrays in meetings. We extract speaker embeddings using two popular and robust pre-trained models, ECAPA-TDNN and X-vectors, and calculate the TDOA values via the Generalized Cross-Correlation (GCC) method with Phase Transform (PHAT) weighting. Although ECAPA-TDNN outperforms the Xvectors model, we utilize both speaker embedding models to explore the potential of employing a computationally lighter model when spatial information is exploited. Various techniques for combining the spatial-temporal information are examined in order to determine the best clustering method. The proposed framework is evaluated on two multichannel datasets: the AVLab Speaker Localization dataset and a multichannel dataset (SpeaD-M3C) enriched in the context of the present work with supplementary information from smartphone recordings. Our results strongly indicate that the integration of spatial information can significantly improve the performance of state-of-the-art deep learning diarization models, presenting a 2-3% reduction in DER compared to the baseline approach on the evaluated datasets.

Automatic cross‐ and multi‐lingual recognition of dysphonia by ensemble classification using deep speaker embedding models

Automatic cross‐ and multi‐lingual recognition of dysphonia by ensemble classification using deep speaker embedding models

X-TF-GridNet: A time–frequency domain target speaker extraction network with adaptive speaker embedding fusion

Target speaker extraction (TSE) which has the capability to directly extract desired speech given enrollment utterances of the target speaker has attracted more and more attention for its potential applications in solving the cocktail-party problem. Despite the considerable progress made by existing time-domain methods, which have become the dominant approach for TSE, these methods often significantly degrade their performance under more realistic conditions. This paper proposes an innovative approach in the time–frequency (T–F) domain, namely X-TF-GridNet, which uses complex spectrum mapping to extract the real and imaginary (RI) components of the target speech. Specifically, the TF-GridNet block was designed to serve as the primary speaker extractor module. Our proposed method boasts two key extensions: first, a U2-Net style network adeptly extracts robust fixed speaker embeddings, which could efficiently capture and represent target speaker information. Second, an adaptive embedding fusion (AEA) mechanism ensures the effective utilization of target speaker information, which makes the backbone extractor focus on the speech of interest. Additionally, we also introduced a multi-task learning framework, comprising two distinct loss functions, to explicitly enhance both the discriminative speaker embeddings for the reference speech and the overall quality of the target speech. We conducted extensive ablation studies and quantitative comparisons against previous TSE methods on both the WSJ0-2mix and its noisy and reverberant counterparts. Our proposed method achieved a commendable SI-SDR of 19.7 dB with a moderate model size on the WSJ0-2mix dataset, and the SI-SDR can be improved to 20.7 dB with a larger model. Experimental results demonstrated that, compared with existing time-domain approaches, our proposed method not only achieved competitive performance across multiple objective metrics but also mitigated speaker confusion errors under more challenging conditions, including various interferences such as noises and reverberation.

Effect of identical twins on deep speaker embeddings based forensic voice comparison

Deep learning has gained widespread adoption in forensic voice comparison in recent years. It is mainly used to learn speaker representations, known as embedding features or vectors. In this work, the effect of identical twins on two state-of-the-art deep speaker embedding methods was investigated with special focus on metrics of forensic voice comparison. The speaker verification performance has been assessed using the likelihood-ratio framework by likelihood ratio cost and equal error rate. The AVTD twin speech dataset was applied. The results show a significant reduction in speaker verification performance when twin samples are present. Neither the adaptation of LR score calculation to twin samples, nor fine-tuning the pre-trained speaker embedding models seemed to be able to leverage this limitation. It was found that the recognition of same or different speakers was possible even in the case of identical twins but the performance dropped greatly. The lowest EER of the best performing model was 3.4% in the case of non-twin; at the same time, EER was 25.3% when twins were present. This doesn’t mean that the presented methods are useless in case of identical twins, but it must be taken into consideration that in case of a higher likelihood-ratio score (which indicates same speakers on the tested samples), the possibility of twins must also be considered in a real casework.

Sensing to Hear through Memory

Speech enhancement on mobile devices is a very challenging task due to the complex environmental noises. Recent works using lip-induced ultrasound signals for speech enhancement open up new possibilities to solve such a problem. However, these multi-modal methods cannot be used in many scenarios where ultrasound-based lip sensing is unreliable or completely absent. In this paper, we propose a novel paradigm that can exploit the prior learned ultrasound knowledge for multi-modal speech enhancement only with the audio input and an additional pre-enrollment speaker embedding. We design a memory network to store the ultrasound memory and learn the interrelationship between the audio and ultrasound modality. During inference, the memory network is able to recall the ultrasound representations from audio input to achieve multi-modal speech enhancement without needing real ultrasound signals. Moreover, we introduce a speaker embedding module to further boost the enhancement performance as well as avoid the degradation of the recalling when the noise level is high. We adopt an end-to-end multi-task manner to train the proposed framework and perform extensive evaluations on the collected dataset. The results show that our method yields comparable performance with audio-ultrasound methods and significantly outperforms the audio-only methods.

MRMI-TTS: Multi-Reference Audios and Mutual Information Driven Zero-Shot Voice Cloning

Voice cloning in text-to-speech (TTS) is the process of replicating the voice of a target speaker with limited data. Among various voice cloning techniques, this article focuses on zero-shot voice cloning. Although existing TTS models can generate high-quality speech for seen speakers, cloning the voice of an unseen speaker remains a challenging task. The key aspect of zero-shot voice cloning is to obtain a speaker embedding from the target speaker. Previous works have used a speaker encoder to obtain a fixed-size speaker embedding from a single reference audio unsupervised, but they suffer from insufficient speaker information and content information leakage in speaker embedding. To address these issues, this article proposes MRMI-TTS, a FastSpeech2-based framework that uses speaker embedding as a conditioning variable to provide speaker information. The MRMI-TTS extracts speaker embedding and content embedding from multi-reference audios using a speaker encoder and a content encoder. To obtain sufficient speaker information, multi-reference audios are selected based on sentence similarity. The proposed model applies mutual information minimization on the two embeddings to remove entangled information within each embedding. Experiments on the public English dataset VCTK show that our method can improve synthesized speech in terms of both similarity and naturalness, even for unseen speakers. Compared to state-of-the-art reference embedding learned methods, our method achieves the best performance on the zero-shot voice cloning task. Furthermore, we demonstrate that the proposed method has a better capability of maintaining the speaker embedding in different languages. Sample outputs are available on the demo page. 1

SAMI: an M-Health application to telemonitor intelligibility and speech disorder severity in head and neck cancers.

Perceptual measures, such as intelligibility and speech disorder severity, are widely used in the clinical assessment of speech disorders in patients treated for oral or oropharyngeal cancer. Despite their widespread usage, these measures are known to be subjective and hard to reproduce. Therefore, an M-Health assessment based on an automatic prediction has been seen as a more robust and reliable alternative. Despite recent progress, these automatic approaches still remain somewhat theoretical, and a need to implement them in real clinical practice rises. Hence, in the present work we introduce SAMI, a clinical mobile application used to predict speech intelligibility and disorder severity as well as to monitor patient progress on these measures over time. The first part of this work illustrates the design and development of the systems supported by SAMI. Here, we show how deep neural speaker embeddings are used to automatically regress speech disorder measurements (intelligibility and severity), as well as the training and validation of the system on a French corpus of head and neck cancer. Furthermore, we also test our model on a secondary corpus recorded in real clinical conditions. The second part details the results obtained from the deployment of our system in a real clinical environment, over the course of several weeks. In this section, the results obtained with SAMI are compared to an a posteriori perceptual evaluation, conducted by a set of experts on the new recorded data. The comparison suggests a high correlation and a low error between the perceptual and automatic evaluations, validating the clinical usage of the proposed application.

Improved Convolutional Neural Network–Time-Delay Neural Network Structure with Repeated Feature Fusions for Speaker Verification

The development of deep learning greatly promotes the progress of speaker verification (SV). Studies show that both convolutional neural networks (CNNs) and dilated time-delay neural networks (TDNNs) achieve advanced performance in text-independent SV, due to their ability to sufficiently extract the local feature and the temporal contextual information, respectively. Also, the combination of the above two has achieved better results. However, we found a serious gridding effect when we apply the 1D-Res2Net-based dilated TDNN proposed in ECAPA-TDNN for SV, which indicates discontinuity and local information losses of frame-level features. To achieve high-resolution process for speaker embedding, we improve the CNN–TDNN structure with proposed repeated multi-scale feature fusions. Through the proposed structure, we can effectively improve the channel utilization of TDNN and achieve higher performance under the same TDNN channel. And, unlike previous studies that have all converted CNN features to TDNN features directly, we also studied the latent space transformation between CNN and TDNN to achieve efficient conversion. Our best method obtains 0.72 EER and 0.0672 MinDCF on VoxCeleb-O test set, and the proposed method performs better in cross-domain SV without additional parameters and computational complexity.

Real-time multilingual speech recognition and speaker diarization system based on Whisper segmentation.

This research presents the development of a cutting-edge real-time multilingual speech recognition and speaker diarization system that leverages OpenAI's Whisper model. The system specifically addresses the challenges of automatic speech recognition (ASR) and speaker diarization (SD) in dynamic, multispeaker environments, with a focus on accurately processing Mandarin speech with Taiwanese accents and managing frequent speaker switches. Traditional speech recognition systems often fall short in such complex multilingual and multispeaker contexts, particularly in SD. This study, therefore, integrates advanced speech recognition with speaker diarization techniques optimized for real-time applications. These optimizations include handling model outputs efficiently and incorporating speaker embedding technology. The system was evaluated using data from Taiwanese talk shows and political commentary programs, featuring 46 diverse speakers. The results showed a promising word diarization error rate (WDER) of 2.68% in two-speaker scenarios and 11.65% in three-speaker scenarios, with an overall WDER of 6.96%. This performance is comparable to that of non-real-time baseline models, highlighting the system's ability to adapt to various complex conversational dynamics, a significant advancement in the field of real-time multilingual speech processing.

Self-Supervised Disentangled Representation Learning for Robust Target Speech Extraction

Speech signals are inherently complex as they encompass both global acoustic characteristics and local semantic information. However, in the task of target speech extraction, certain elements of global and local semantic information in the reference speech, which are irrelevant to speaker identity, can lead to speaker confusion within the speech extraction network. To overcome this challenge, we propose a self-supervised disentangled representation learning method. Our approach tackles this issue through a two-phase process, utilizing a reference speech encoding network and a global information disentanglement network to gradually disentangle the speaker identity information from other irrelevant factors. We exclusively employ the disentangled speaker identity information to guide the speech extraction network. Moreover, we introduce the adaptive modulation Transformer to ensure that the acoustic representation of the mixed signal remains undisturbed by the speaker embeddings. This component incorporates speaker embeddings as conditional information, facilitating natural and efficient guidance for the speech extraction network. Experimental results substantiate the effectiveness of our meticulously crafted approach, showcasing a substantial reduction in the likelihood of speaker confusion.

Learning Deep Embedding with Acoustic and Phoneme Features for Speaker Recognition in FM Broadcasting

Text-independent speaker verification (TI-SV) is a crucial task in speaker recognition, as it involves verifying an individual’s claimed identity from speech of arbitrary content without any human intervention. The target for TI-SV is to design a discriminative network to learn deep speaker embedding for speaker idiosyncrasy. In this paper, we propose a deep speaker embedding learning approach of a hybrid deep neural network (DNN) for TI-SV in FM broadcasting. Not only acoustic features are utilized, but also phoneme features are introduced as prior knowledge to collectively learn deep speaker embedding. The hybrid DNN consists of a convolutional neural network architecture for generating acoustic features and a multilayer perceptron architecture for extracting phoneme features sequentially, which represent significant pronunciation attributes. The extracted acoustic and phoneme features are concatenated to form deep embedding descriptors for speaker identity. The hybrid DNN demonstrates not only the complementarity between acoustic and phoneme features but also the temporality of phoneme features in a sequence. Our experiments show that the hybrid DNN outperforms existing methods and delivers a remarkable performance in FM broadcasting TI-SV.

TER-CA-WGNN: Trimodel Emotion Recognition Using Cumulative Attribute-Weighted Graph Neural Network

Affective computing is a multidisciplinary field encompassing artificial intelligence, natural language processing, linguistics, computer science, and social sciences. This field aims to deepen our comprehension and capabilities by deploying inventive algorithms. This article presents a groundbreaking approach, the Cumulative Attribute-Weighted Graph Neural Network, which is innovatively designed to integrate trimodal textual, audio, and visual data from the two multimodal datasets. This method exemplifies its effectiveness in performing comprehensive multimodal sentiment analysis. Our methodology employs vocal inputs to generate speaker embeddings trimodal analysis. Using a weighted graph structure, our model facilitates the efficient integration of these diverse modalities. This approach underscores the interrelated aspects of various emotional indicators. The paper’s significant contribution is underscored by its experimental results. Our novel algorithm achieved impressive performance metrics on the CMU-MOSI dataset, with an accuracy of 94% and precision, recall, and F1-scores above 92% for Negative, Neutral, and Positive emotion categories. Similarly, on the IEMOCAP dataset, the algorithm demonstrated its robustness with an overall accuracy of 93%, where exceptionally high precision and recall were noted in the Neutral and Positive categories. These results mark a notable advancement over existing state-of-the-art models, illustrating the potential of our approach in enhancing Sentiment Recognition through the synergistic use of trimodal data. This study’s comprehensive analysis and significant results demonstrate the proposed algorithm’s effectiveness in nuanced emotional state recognition and pave the way for future advancements in affective computing, emphasizing the value of integrating multimodal data for improved accuracy and robustness.

Comparison of Modern Deep Learning Models for Speaker Verification

This research presents an extensive comparative analysis of a selection of popular deep speaker embedding models, namely WavLM, TitaNet, ECAPA, and PyAnnote, applied in speaker verification tasks. The study employs a specially curated dataset, specifically designed to mirror the real-world operating conditions of voice models as accurately as possible. This dataset includes short, non-English statements gathered from interviews on a popular online video platform. The dataset features a wide range of speakers, with 33 males and 17 females, making a total of 50 unique voices. These speakers vary in age from 20 to 70 years old. This variety helps in thoroughly testing speaker verification models. This dataset is especially useful for research on speaker verification with short recordings. It consists of 10 clips for each person, each clip being no longer than 10 s, adding up to 500 recordings in total. The total length of all recordings is about 1 h and 30 min, which averages to roughly 100 s for each speaker. This dataset is a valuable tool for research in speaker verification, particularly for studies involving short audio clips. The performance of these models is evaluated using common biometric metrics such as false acceptance rate (FAR), false rejection rate (FRR), equal error rate (EER) and detection cost function (DCF). The results reveal that the TitaNet and ECAPA models stand out by presenting the lowest EER (1.91% and 1.71%, respectively) and thus exhibiting higher discriminative features, ensuring, on the one hand, a reduction in intra-class distance (the same speaker), and, on the other hand, maximizing the distance between different speaker embeddings. This analysis also highlights the ECAPA model’s advantageous balance of performance and efficiency, achieving an inference time of 69.43 milliseconds, slightly longer than the PyAnnote models. This study not only compares the performance of models but also provides a comparative analysis of respective model embeddings, offering insights into their strengths and weaknesses. The presented findings serve as a foundation for guiding future research in speaker verification, especially in the context of short audio samples or limited data. This may be particularly relevant for applications requiring quick and accurate speaker identification from short voice clips.

An analytic study on clustering driven self-supervised speaker verification

One of the most widely used self-supervised speaker verification system training methods is to optimize the speaker embedding network in a discriminative fashion using clustering algorithm-driven Pseudo-Labels. Although the pseudo-labels-based self-supervised training scheme showed impressive performance, recent studies have shown that label noise can significantly impact performance. In this paper, we have explored various pseudo-labels driven by different clustering algorithms and conducted a fine-grained analysis of the relationship between the quality of the pseudo-labels and the speaker verification performance. Experimentally, we shed light on several previously overlooked aspects of the pseudo-labels that can impact speaker verification performance. Moreover, we could observe that the self-supervised speaker verification performance is heavily dependent on multiple qualitative aspects of the clustering algorithms used to generate the pseudo-labels. Furthermore, we show that speaker verification performance can be severely degraded from overfitting the noisy pseudo-labels and that the mixup strategy can mitigate the memorization effects of label noise.

Joint speaker encoder and neural back-end model for fully end-to-end automatic speaker verification with multiple enrollment utterances

Conventional automatic speaker verification systems can usually be decomposed into a front-end model such as time delay neural network (TDNN) for extracting speaker embeddings and a back-end model such as statistics-based probabilistic linear discriminant analysis (PLDA) or neural network-based neural PLDA (NPLDA) for similarity scoring. However, the sequential optimization of the front-end and back-end models may lead to a local minimum, which theoretically prevents the whole system from achieving the best optimization. Although some methods have been proposed for jointly optimizing the two models, such as the generalized end-to-end (GE2E) model and NPLDA E2E model, most of these methods have not fully investigated how to model the intra-relationship between multiple enrollment utterances. In this paper, we propose a new E2E joint method for speaker verification especially designed for the practical scenario of multiple enrollment utterances. To leverage the intra-relationship among multiple enrollment utterances, our model comes equipped with frame-level and utterance-level attention mechanisms. Additionally, focal loss is utilized to balance the importance of positive and negative samples within a mini-batch and focus on the difficult samples during the training process. We also utilize several data augmentation techniques, including conventional noise augmentation using MUSAN and RIRs datasets and a unique speaker embedding-level mixup strategy for better optimization.