DT-SV: A Transformer-based Time-domain Approach for Speaker Verification

The extraction of speaker-related features through speaker representations or utterance embeddings has been extensively studied for years. Convolutional Neural Networks (CNNs) are the most commonly used backbone network for feature extraction, and deep Residual Networks (ResNets), which incorporate residual connectivity into CNNs, achieve effective frame-level feature extraction inside the same input samples. ResNets have demonstrated impressive performance as speaker embedding extractors. With emerging this backbone and its advanced extensions, such as Res2Net, as dominant backbone networks in speaker recognition, and despite its widespread adoption, a significant gap in current speaker embeddings is the unmodified use of these successful backbones from the visual domain, without considering the specific adaptation requirements for the speech field. Tailoring the backbone to account for the unique characteristics of speech representations ensures effective learning in Automatic Speaker Verification (ASV) systems. This thesis concentrates on the backbone network, the fundamental and most critical component of Deep Neural Networks (DNNs), responsible for extracting features from input data and mapping them into distinct feature representations. One of the key questions addressed in the thesis is the following: Which backbone is most robust and yields the most generalized speaker representation, with the potential to be applied to many existing ASV systems?. Through a strategic evaluation, the study analyzed the performance of integrating different residual-based backbones into the most popular ResNet-based fundamental baselines, ResNet-34 and ECAPA-TDNN based models, with highly different model complexity and CNN stem representations. To this end, the proposed architectures were evaluated on the publicly available VoxCeleb dataset. This thesis introduces an innovative approach to speaker modeling by incorporating scale and cardinality dimensions for the first time. Experimental findings reveal that utilizing multi-scale and multi-branch aggregated residual transformations within the Res2NeXt backbone yields considerable performance improvements. Notably, 6s8g and 8s8g variants of Res2NeXt, prove particularly beneficial in improving the representational capability of learned features for speaker embedding. The consistent superiority, excellent compatibility, and robustness of Res2NeXt compared to the current popular ResNet and Res2Net backbones across various architectures and conditions demonstrate its broad applicability and effectiveness as a backbone in speaker modeling. To further evaluate the Robustness and Generalization (R&G) capabilities of the proposed speaker embeddings, the thesis expanded its scope to cross-domain applications, specifically Speech Emotion Recognition (SER), one of the most challenging downstream tasks. A comparative analysis of the proposed speaker embeddings was conducted using two benchmark datasets, IEMOCAP and CREMA-D. The overarching goal of the thesis is to contribute to the development of general-purpose and universal speech representations that generalize effectively across multiple speech-related tasks. By examining the alignment and divergence of embeddings for ASV and SER, the study seeks to answer a fundamental question: Can a single representation framework cater to both speaker verification and emotion recognition or are task-specific embeddings inevitable?. The results confirm that speaker embeddings derived from the Res2NeXt backbone are also highly effective for the SER task, maintaining their performance across diverse evaluation protocols and data variations. This demonstrates the generalizability of Res2NeXt embeddings in the cross-domain task. The study attributes their success to the ability of Res2NeXt-based models to capture rich speech features, including tone, pitch, and phonation patterns, during speaker recognition training—features that are also vital for emotion recognition.

Phonetic information is one of the most essential components of a speech signal, playing an important role for many speech processing tasks. However, it is difficult to integrate phonetic information into speaker verification systems since it occurs primarily at the frame level while speaker characteristics typically reside at the segment level. In deep neural network-based speaker verification, existing methods only apply phonetic information to the frame-wise trained speaker embeddings. To improve this weakness, this paper proposes phonetic adaptation and hybrid multi-task learning and further combines these into c-vector and simplified c-vector architectures. Experiments on National Institute of Standards and Technology (NIST) speaker recognition evaluation (SRE) 2010 show that the four proposed speaker embeddings achieve better performance than the baseline. The c-vector system performs the best, providing over 30% and 15% relative improvements in equal error rate (EER) for the core-extended and 10 s–10 s conditions, respectively. On the NIST SRE 2016, 2018, and VoxCeleb datasets, the proposed c-vector approach improves the performance even when there is a language mismatch within the training sets or between the training and evaluation sets. Extensive experimental results demonstrate the effectiveness and robustness of the proposed methods.

Multi-task learning (MTL) and attention mechanism have been proven to effectively extract robust acoustic features for various speech-related tasks in noisy environments. In this study, we propose an attention-based MTL (ATM) approach that integrates MTL and the attention-weighting mechanism to simultaneously realize a multi-model learning structure that performs speech enhancement (SE) and speaker identification (SI). The proposed ATM system consists of three parts: SE, SI, and attention-Net (AttNet). The SE part is composed of a long-short-term memory (LSTM) model, and a deep neural network (DNN) model is used to develop the SI and AttNet parts. The overall ATM system first extracts the representative features and then enhances the speech signals in LSTM-SE and specifies speaker identity in DNN-SI. The AttNet computes weights based on DNN-SI to prepare better representative features for LSTM-SE. We tested the proposed ATM system on Taiwan Mandarin hearing in noise test sentences. The evaluation results confirmed that the proposed system can effectively enhance speech quality and intelligibility of a given noisy input. Moreover, the accuracy of the SI can also be notably improved by using the proposed ATM system.

This paper proposes a human-in-the-loop speaker-adaptation method for multi-speaker text-to-speech. With a conventional speaker-adaptation method, a target speaker's embedding vector is extracted from his/her reference speech using a speaker encoder trained on a speaker-discriminative task. However, this method cannot obtain an embedding vector for the target speaker when the reference speech is unavailable. Our method is based on a human-in-the-loop optimization framework, which incorporates a user to explore the speaker-embedding space to find the target speaker's embedding. The proposed method uses a sequential line search algorithm that repeatedly asks a user to select a point on a line segment in the embedding space. To efficiently choose the best speech sample from multiple stimuli, we also developed a system in which a user can switch between multiple speakers' voices for each phoneme while looping an utterance. Experimental results indicate that the proposed method can achieve comparable performance to the conventional one in objective and subjective evaluations even if reference speech is not used as the input of a speaker encoder directly.

Abstract— Audio data carries rich information in the form of speech, music, and environmental sounds, but its raw waveform is often complex and high-dimensional, making direct analysis difficult. Neural network architectures have emerged as powerful tools for extracting meaningful representations from audio signals, enabling efficient analysis and interpretation. Recurrent Neural Networks (RNNs), and Transformer-based models— for learning robust and discriminative features from audio. By automatically capturing temporal, spectral, and contextual patterns, these architectures significantly improve performance in tasks such as speech recognition, speaker identification, music classification, and environmental sound detection. The findings highlight the potential of neural networks to replace traditional handcrafted features, thereby advancing the development of scalable, accurate, and realtime audio processing applications. The rapid growth of audio data across domains such as speech, music, healthcare, and environmental monitoring has created a strong need for effective methods to extract meaningful representations from complex audio signals. Traditional approaches rely on handcrafted features like MFCCs and spectrogram descriptors, which often fail to capture the full temporal and spectral dynamics present in raw audio. Keywords: Neural Networks, Deep Learning, Audio Representation Learning, Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Transformer Models, Feature Extraction, Speech Recognition, Speaker Identification, Sound Event Detection, Spectrogram Analysis, Audio Signal Processing

Speaker Identification systems have become area of intense research in recent years due to its wide range of applications in voice matching, biometric identification, mobile access security, health care management and transportation. The major challenge in speaker identification is how to provide robust and computationally efficient features that accurately captures speaker's unique identity for higher generalized identification. This paper presents hierarchical classification approach for speaker identification using robust time domain features. In hierarchical classification approach, the first level classifier identifies the gender voice (i.e. male or female voice). In addition, the second level classifier identifies the specific speaker voice. The experiments were conducted using the speech data collected from ten subjects including male and female to ensure the generalized model evaluations. Several time domain features were extracted from the collected dataset that were proven highly discriminative for gender identification and specific speaker identification. The proposed method helps to reduce computational time with increase speaker recognition rate. The experimental results obtained highest accuracy of 96.9% using random forest classifier for gender identification. Moreover, for specific speaker identification the highest accuracy of 78% was observed in male speaker identification and 88.7% accuracy was obtained in female speaker identification using random forest classifier. The encouraging results of our experiments show the impact of hierarchical speaker identification using carefully selected time domain features.

LSTM-based speaker verification usually uses a fixed-length local segment randomly truncated from an utterance to learn the utterance-level speaker embedding, while using the average embedding of all segments of a test utterance to verify the speaker, which results in a critical mismatch between testing and training. This mismatch degrades the performance of speaker verification, especially when the durations of training and testing utterances are very different. To alleviate this issue, we propose the deep segment attentive embedding method to learn the unified speaker embeddings for utterances of variable duration. Each utterance is segmented by a sliding window and LSTM is used to extract the embedding of each segment. Instead of only using one local segment, we use the whole utterance to learn the utterance-level embedding by applying an attentive pooling to the embeddings of all segments. Moreover, the similarity loss of segment-level embeddings is introduced to guide the segment attention to focus on the segments with more speaker discriminations, and jointly optimized with the similarity loss of utterance-level embeddings. Systematic experiments on Tongdun and VoxCeleb show that the proposed method significantly improves robustness of duration variant and achieves the relative Equal Error Rate reduction of 50% and 11.54% , respectively.

This paper proposes a serialized multi-layer multi-head attention for neural speaker embedding in text-independent speaker verification. In prior works, frame-level features from one layer are aggregated to form an utterance-level representation. Inspired by the Transformer network, our proposed method utilizes the hierarchical architecture of stacked self-attention mechanisms to derive refined features that are more correlated with speakers. Serialized attention mechanism contains a stack of self-attention modules to create fixed-dimensional representations of speakers. Instead of utilizing multi-head attention in parallel, the proposed serialized multi-layer multi-head attention is designed to aggregate and propagate attentive statistics from one layer to the next in a serialized manner. In addition, we employ an input-aware query for each utterance with the statistics pooling. With more layers stacked, the neural network can learn more discriminative speaker embeddings. Experiment results on VoxCeleb1 dataset and SITW dataset show that our proposed method outperforms other baseline methods, including x-vectors and other x-vectors + conventional attentive pooling approaches by 9.7% in EER and 8.1% in DCF0.01.

Speaker and language recognition in noisy and degraded channel conditions continue to be a challenging problem mainly due to the mismatch between clean training and noisy test conditions. In the presence of noise, the most reliable portions of the signal are the high energy regions which can be used for robust feature extraction. In this paper, we propose a front end processing scheme based on autoregressive (AR) models that represent the high energy regions with good accuracy followed by a modulation filtering process. The AR model of the spectrogram is derived using two separable time and frequency AR transforms. The first AR model (temporal AR model) of the sub-band Hilbert envelopes is derived using frequency domain linear prediction (FDLP). This is followed by a spectral AR model applied on the FDLP envelopes. The output 2-D AR model represents a low-pass modulation filtered spectrogram of the speech signal. The band-pass modulation filtered spectrograms can further be derived by dividing two AR models with different model orders (cut-off frequencies). The modulation filtered spectrograms are converted to cepstral coefficients and are used for a speaker recognition task in noisy and reverberant conditions. Various speaker recognition experiments are performed with clean and noisy versions of the NIST-2010 speaker recognition evaluation (SRE) database using the state-of-the-art speaker recognition system. In these experiments, the proposed front-end analysis provides substantial improvements (relative improvements of up to 25%) compared to baseline techniques. Furthermore, we also illustrate the generalizability of the proposed methods using language identification (LID) experiments on highly degraded high-frequency (HF) radio channels and speech recognition experiments on noisy data.

The paper presents an unsupervised binaural scene analyzer that is capable of performing simultaneous operations, such as localization, detection and recognition of target speaker amidst in a reverberant and noise inferential environment. The proposed technique includes three main stages: sound source localization, speaker recognition and auditory saliency-based indexing system. The stage one involves the localization of target speaker by incorporating the binaural cues into Azimuth dependant GMM-EM classifier. During the second stage, the study proposes a Gabor-Hilbert Envelope Coefficient (GHEC) based spectro-temporal extractor as an efficient speaker recognition method which shows better robustness with minimum computational complexity. The Hilbert Envelope provides relevant acoustic information and also it improvises the performance of speaker identification process in different reverberant environments and SNR values. Later in the third stage, the auditory saliency based diarization is proposed as a process of indexing the speech contents based on the image identity of the target speaker. The proposed system may be used to catalogue the entire speech content with the corresponding image of the target speaker that finds wide range of applications, including teleconference, hands-free communication and meeting hall content localization and fast audio retrieval in a repository.

Speaker identification is the process of automatically determining who is speaking from the known speakers by the model. It is crucial in voice-based authentication, forensic investigations, security and surveillance. In recent studies, the combination of convolutional neural network (CNN) and recurrent neural network (RNN) variants performed better than separate models of both. However, only limited studies are conducted in speaker identification using a combination of CNN and RNN variants. In this study, we proposed speaker identification using hybrid two-dimensional CNN (2DCNN) and bidirectional gated recurrent unit (BiGRU) to improve performance. The proposed model integrates the advantage of 2DCNN and BiGRU layers to improve the performance of the model. 2DCNN layers have the advantage of extracting short-term spatial features from input data and it has a limited number of parameters for computation. BiGRU layers have an advantage in extracting long-term temporal dependency between the features in both directions (i.e. backward and forward) and it is efficient in achieving convergence during training. Spectrograms of the speech were used as input in our proposed model because of the rich acoustic features of the speaker. To compare the performance of the proposed model, additional experiments are conducted using the models 2DCNN, CNN-LSTM, CNN-BiLSTM and CNN-GRU. The experiments were conducted on the VoxCeleb1 audio dataset, which consists of 153,516 utterances collected from the 1251 speakers. The accuracy, precision, recall and f1 score of the proposed model are 98.28%, 99.08%, 98.92% and 98.97% respectively. The proposed model was compared with the existing works to show the effectiveness of the proposed model. The experiment results and the comparison with the existing works show that the proposed model has higher performance than both existing works and other models experimented in this study.

Significance of analytic phase of speech signals in speaker verification

High-fidelity speech can be synthesized by end-to-end text-to-speech models in recent years. However, accessing and controlling speech attributes such as speaker identity, prosody, and emotion in a text-to-speech system remains a challenge. This paper presents a system involving feedback constraint for multispeaker speech synthesis. We manage to enhance the knowledge transfer from the speaker verification to the speech synthesis by engaging the speaker verification network. The constraint is taken by an added loss related to the speaker identity, which is centralized to improve the speaker similarity between the synthesized speech and its natural reference audio. The model is trained and evaluated on publicly available datasets. Experimental results, including visualization on speaker embedding space, show significant improvement in terms of speaker identity cloning in the spectrogram level. Synthesized samples are available online for listening. (this https URL)

A human speaker recognition expert often observes the speech spectrogram in multiple different scales for speaker recognition, especially under the short utterance condition. Inspired by this action, this paper proposes a novel multi-resolution time frequency feature (MRTF) extraction method, which is obtained by performing a 2-Dimensional discrete cosine transform (DCT) in multi-scale on the time frequency spectrogram matrix and then selecting and combining to the final multi-scaled transformed elements. Compared to the traditional Mel-Frequency Cepstral Coefficient (MFCC) feature extraction, the proposed method can make better use of multi-resolution temporal-frequency information. Beyond this, we also proposed three complementary combination strategies of MFCC and MRTF: in feature level, in i-vector level and in score level. Comparing their performance. We found the best results are obtained by combination in i-vector level. In the three NIST 2008 Speaker Recognition Evaluation datasets, the proposed method is the most effective for improving the performance under short utterance than under long utterance. And after the combination, we can achieve an EER of 11.32 % and MinDCF of 0.054 in the 10sec-10sec trials on the male dataset, which is an absolute 3 % improvement of EER than the best reported result in this field.

In this paper, we present a novel system for joint speaker identification and speech separation. For speaker identification a single-channel speaker identification algorithm is proposed which provides an estimate of signal-to-signal ratio (SSR) as a by-product. For speech separation, we propose a sinusoidal model-based algorithm. The speech separation algorithm consists of a double-talk/single-talk detector followed by a minimum mean square error estimator of sinusoidal parameters for finding optimal codevectors from pre-trained speaker codebooks. In evaluating the proposed system, we start from a situation where we have prior information of codebook indices, speaker identities and SSR-level, and then, by relaxing these assumptions one by one, we demonstrate the efficiency of the proposed fully blind system. In contrast to previous studies that mostly focus on automatic speech recognition (ASR) accuracy, here, we report the objective and subjective results as well. The results show that the proposed system performs as well as the best of the state-of-the-art in terms of perceived quality while its performance in terms of speaker identification and automatic speech recognition results are generally lower. It outperforms the state-of-the-art in terms of intelligibility showing that the ASR results are not conclusive. The proposed method achieves on average, 52.3% ASR accuracy, 41.2 points in MUSHRA and 85.9% in speech intelligibility.