Unseen Speakers Research Articles

Exploring the Impact of Fine-Tuning the Wav2vec2 Model in Database-Independent Detection of Dysarthric Speech.

Many acoustic features and machine learning models have been studied to build automatic detection systems to distinguish dysarthric speech from healthy speech. These systems can help to improve the reliability of diagnosis. However, speech recorded for diagnosis in real-life clinical conditions can differ from the training data of the detection system in terms of, for example, recording conditions, speaker identity, and language. These mismatches may lead to a reduction in detection performance in practical applications. In this study, we investigate the use of the wav2vec2 model as a feature extractor together with a support vector machine (SVM) classifier to build automatic detection systems for dysarthric speech. The performance of the wav2vec2 features is evaluated in two cross-database scenarios, language-dependent and language-independent, to study their generalizability to unseen speakers, recording conditions, and languages before and after fine-tuning the wav2vec2 model. The results revealed that the fine-tuned wav2vec2 features showed better generalization in both scenarios and gave an absolute accuracy improvement of 1.46%-8.65% compared to the non-fine-tuned wav2vec2 features.

Using deep learning to improve the intelligibility of a target speaker in noisy multi-talker environments for people with normal hearing and hearing loss.

Understanding speech in noisy environments is a challenging task, especially in communication situations with several competing speakers. Despite their ongoing improvement, assistive listening devices and speech processing approaches still do not perform well enough in noisy multi-talker environments, as they may fail to restore the intelligibility of a speaker of interest among competing sound sources. In this study, a quasi-causal deep learning algorithm was developed that can extract the voice of a target speaker, as indicated by a short enrollment utterance, from a mixture of multiple concurrent speakers in background noise. Objective evaluation with computational metrics demonstrated that the speaker-informed algorithm successfully extracts the target speaker from noisy multi-talker mixtures. This was achieved using a single algorithm that generalized to unseen speakers, different numbers of speakers and relative speaker levels, and different speech corpora. Double-blind sentence recognition tests on mixtures of one, two, and three speakers in restaurant noise were conducted with listeners with normal hearing and listeners with hearing loss. Results indicated significant intelligibility improvements with the speaker-informed algorithm of 17% and 31% for people without and with hearing loss, respectively. In conclusion, it was demonstrated that deep learning-based speaker extraction can enhance speech intelligibility in noisy multi-talker environments where uninformed speech enhancement methods fail.

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

SIG: Speaker Identification in Literature via Prompt-Based Generation

Identifying speakers of quotations in narratives is an important task in literary analysis, with challenging scenarios including the out-of-domain inference for unseen speakers, and non-explicit cases where there are no speaker mentions in surrounding context. In this work, we propose a simple and effective approach SIG, a generation-based method that verbalizes the task and quotation input based on designed prompt templates, which also enables easy integration of other auxiliary tasks that further bolster the speaker identification performance. The prediction can either come from direct generation by the model, or be determined by the highest generation probability of each speaker candidate. Based on our approach design, SIG supports out-of-domain evaluation, and achieves open-world classification paradigm that is able to accept any forms of candidate input. We perform both cross-domain evaluation and in-domain evaluation on PDNC, the largest dataset of this task, where empirical results suggest that SIG outperforms previous baselines of complicated designs, as well as the zero-shot ChatGPT, especially excelling at those hard non-explicit scenarios by up to 17% improvement. Additional experiments on another dataset WP further corroborate the efficacy of SIG.

Phoneme Hallucinator: One-Shot Voice Conversion via Set Expansion

Voice conversion (VC) aims at altering a person's voice to make it sound similar to the voice of another person while preserving linguistic content. Existing methods suffer from a dilemma between content intelligibility and speaker similarity; i.e., methods with higher intelligibility usually have a lower speaker similarity, while methods with higher speaker similarity usually require plenty of target speaker voice data to achieve high intelligibility. In this work, we propose a novel method Phoneme Hallucinator that achieves the best of both worlds. Phoneme Hallucinator is a one-shot VC model; it adopts a novel model to hallucinate diversified and high-fidelity target speaker phonemes based just on a short target speaker voice (e.g. 3 seconds). The hallucinated phonemes are then exploited to perform neighbor-based voice conversion. Our model is a text-free, any-to-any VC model that requires no text annotations and supports conversion to any unseen speaker. Quantitative and qualitative evaluations show that Phoneme Hallucinator outperforms existing VC methods for both intelligibility and speaker similarity.

Generalizing sentence-level lipreading to unseen speakers: a two-stream end-to-end approach

Generalizing sentence-level lipreading to unseen speakers: a two-stream end-to-end approach

Prompt Tuning of Deep Neural Networks for Speaker-adaptive Visual Speech Recognition.

Visual Speech Recognition (VSR) aims to infer speech into text depending on lip movements alone. As it focuses on visual information to model the speech, its performance is inherently sensitive to personal lip appearances and movements, and this makes the VSR models show degraded performance when they are applied to unseen speakers. In this paper, to remedy the performance degradation of the VSR model on unseen speakers, we propose prompt tuning methods of Deep Neural Networks (DNNs) for speaker-adaptive VSR. Specifically, motivated by recent advances in Natural Language Processing (NLP), we finetune prompts on adaptation data of target speakers instead of modifying the pre-trained model parameters. Different from the previous prompt tuning methods mainly limited to Transformer variant architecture, we explore different types of prompts, the addition, the padding, and the concatenation form prompts that can be applied to the VSR model which is composed of CNN and Transformer in general. With the proposed prompt tuning, we show that the performance of the pre-trained VSR model on unseen speakers can be largely improved by using a small amount of adaptation data (e.g., less than 5 minutes), even if the pre-trained model is already developed with large speaker variations. Moreover, by analyzing the performance and parameters of different types of prompts, we investigate when the prompt tuning is preferred over the finetuning methods. The effectiveness of the proposed method is evaluated on both word- and sentence-level VSR databases, LRW-ID and GRID.

Emotion selectable end-to-end text-based speech editing

Text-based speech editing is a convenient way for users to edit speech by intuitively cutting, copying, and pasting text. Previous work introduced CampNet, a context-aware mask prediction network that significantly improved the quality of edited speech. However, this paper proposes a new task: adding emotional effects to the edited speech during text-based speech editing to enhance the expressiveness and controllability of the edited speech. To achieve this, we introduce Emo-CampNet, which allows users to select emotional attributes for the generated speech and has the ability to edit the speech of unseen speakers. Firstly, the proposed end-to-end model controls the generated speech's emotion by introducing additional emotion attributes based on the context-aware mask prediction network. Secondly, to prevent emotional interference from the original speech, a neutral content generator is proposed to remove the emotional components, which is optimized using the generative adversarial framework. Thirdly, two data augmentation methods are proposed to enrich the emotional and pronunciation information in the training set. Experimental results1 show that Emo-CampNet effectively controls the generated speech's emotion and can edit the speech of unseen speakers. Ablation experiments further validate the effectiveness of emotional selectivity and data augmentation methods.

Comparing accuracy in voice-based assessments of biological speaker traits across speech types

Nonverbal acoustic parameters of the human voice provide cues to a vocaliser’s sex, age, and body size that are relevant in human social and sexual communication, and also increasingly so for computer-based voice recognition and synthesis technologies. While studies have shown some capacity in human listeners to gauge these biological traits from unseen speakers, it remains unknown whether speech complexity improves accuracy. Here, in over 200 vocalisers and 1500 listeners of both sexes, we test whether voice-based assessments of sex, age, height and weight vary from isolated vowels and words, to sequences of vowels and words, to full sentences or paragraphs. We show that while listeners judge sex and especially age more accurately as speech complexity increases, accuracy remains high across speech types, even for a single vowel sound. In contrast, the actual heights and weights of vocalisers explain comparatively less variance in listener’s assessments of body size, which do not vary systematically by speech type. Our results thus show that while more complex speech can improve listeners’ biological assessments, the gain is ecologically small, as listeners already show an impressive capacity to gauge speaker traits from extremely short bouts of standardised speech, likely owing to within-speaker stability in underlying nonverbal vocal parameters such as voice pitch. We discuss the methodological, technological, and social implications of these results.

Speak, Read and Prompt: High-Fidelity Text-to-Speech with Minimal Supervision

Abstract We introduce SPEAR-TTS, a multi-speaker text-to-speech (TTS) system that can be trained with minimal supervision. By combining two types of discrete speech representations, we cast TTS as a composition of two sequence-to-sequence tasks: from text to high-level semantic tokens (akin to “reading”) and from semantic tokens to low-level acoustic tokens (“speaking”). Decoupling these two tasks enables training of the “speaking” module using abundant audio-only data, and unlocks the highly efficient combination of pretraining and backtranslation to reduce the need for parallel data when training the “reading” component. To control the speaker identity, we adopt example prompting, which allows SPEAR-TTS to generalize to unseen speakers using only a short sample of 3 seconds, without any explicit speaker representation or speaker labels. Our experiments demonstrate that SPEAR-TTS achieves a character error rate that is competitive with state-of-the-art methods using only 15 minutes of parallel data, while matching ground-truth speech in naturalness and acoustic quality.

Effective Zero-Shot Multi-Speaker Text-to-Speech Technique Using Information Perturbation and a Speaker Encoder

Effective Zero-Shot Multi-Speaker Text-to-Speech Technique Using Information Perturbation and a Speaker Encoder

LipFormer: Learning to Lipread Unseen Speakers Based on Visual-Landmark Transformers

Lipreading refers to understanding and further translating the speech of a video speaker into textual outputs. State-of-the-art lipreading methods excel in interpreting overlap speakers, i.e., speakers appear in both training and inference. However, generalizing those methods to <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">unseen</i> speakers incurs catastrophic performance degradation due to the limited number of speakers in training bank as well as the dominant visual variations caused by the shape/color of lips presented by different speakers. Therefore, merely depending on the visible changes of lips tends to overfit the model. To improve to generalise, in this paper we propose to use multi-modal features, i.e., visual and landmark, to describe the lip motion while being irrespective to speaker characteristics. The proposed sentence-level framework, dubbed <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">LipFormer</i> , is based on visual-landmark transformer architecture wherein a lip motion stream, a facial landmark stream, and a cross-modal fusion are interconnected. More specifically, the two-stream embeddings produced by self-attention are prompted into a cross-attention module to achieve the alignment across visual and landmark variations. The resulting fused features are decoded into linguistic texts by a cascaded sequence-to-sequence translation. Extensive experiments demonstrate that our method can generalise well to unseen speakers in multiple datasets.

Data augmentation for speech separation

Deep learning models have advanced the state of the art of monaural speech separation. However, the performance of a separation model considerably decreases when tested on unseen speakers and noisy conditions. Separation models trained with data augmentation generalize better to unseen conditions. In this paper, we conduct a comprehensive survey of data augmentation techniques and apply them to improve the generalization of time-domain speech separation models. The augmentation techniques include seven source-preserving approaches (Gaussian noise, Gain, Time masking, frequency masking, Short noise, Time stretch, and Pitch shift) and three non-source preserving approaches (Dynamix mixing, Mixup, and Cutmix). After hyperparameter search for each augmentation method, we test the generalization of the augmented model by cross-corpus testing on three datasets (LibriMix, TIMIT, and VCTK), and identify the best augmentation combination that enhances generalization. Experimental results indicate that a combination of several non-source preserving strategies (CutMix, Mixup, and Dynamic mixing) resulted in the best generalization performance. Finally, the augmentation combinations also improved the performance of the speech separation model even when fewer training data are available.

Deep learning-based speaker-adaptive postfiltering with limited adaptation data for embedded text-to-speech synthesis systems

End-to-end (e2e) speech synthesis systems have become popular with the recent introduction of text-to-spectrogram conversion systems, such as Tacotron, that use encoder–decoder-based neural architectures. Even though those sequence-to-sequence systems can produce mel-spectrograms from the letters without a text processing frontend, they require substantial amounts of well-manipulated, labeled audio data that have high SNR and minimum amounts of artifacts. These data requirements make it difficult to build end-to-end systems from scratch, especially for low-resource languages. Moreover, most of the e2e systems are not designed for devices with tiny memory and CPU resources. Here, we investigate using a traditional deep neural network (DNN) for acoustic modeling together with a postfilter that improves the speech features produced by the network. The proposed architectures were trained with the relatively noisy, multi-speaker, Wall Street Journal (WSJ) database and tested with unseen speakers. The thin postfilter layer was adapted with minimal data to the target speaker for testing. We investigated several postfilter architectures and compared them with both objective and subjective tests. Fully-connected and transformer-based architectures performed the best in subjective tests. The novel adversarial transformer-based architecture with adaptive discriminator loss performed the best in the objective tests. Moreover, it was faster than the other architectures both in training and inference. Thus, our proposed lightweight transformer-based postfilter architecture significantly improved speech quality and efficiently adapted to new speakers with few shots of data and a hundred training iterations, making it computationally efficient and suitable for scalability.

Audio-Driven Dubbing for User Generated Contents via Style-Aware Semi-Parametric Synthesis

Existing automated dubbing methods are usually designed for Professionally Generated Content (PGC) production, which requires massive training data and training time to learn a person-specific audio-video mapping. In this paper, we investigate an audio-driven dubbing method that is more feasible for User Generated Content (UGC) production. There are two unique challenges to design a method for UGC: 1) the appearances of speakers are diverse and arbitrary as the method needs to generalize across users; 2) the available video data of one speaker are very limited. In order to tackle the above challenges, we first introduce a new Style Translation Network to integrate the speaking style of the target and the speaking content of the source via a cross-modal AdaIN module. It enables our model to quickly adapt to a new speaker. Then, we further develop a semi-parametric video renderer, which takes full advantage of the limited training data of the unseen speaker via a video-level retrieve-warp-refine pipeline. Finally, we propose a temporal regularization for the semi-parametric renderer, generating more continuous videos. Extensive experiments show that our method generates videos that accurately preserve various speaking styles, yet with considerably lower amount of training data and training time in comparison to existing methods. Besides, our method achieves a faster testing speed than most recent methods.

Speaker-Independent Emotional Voice Conversion via Disentangled Representations

Emotional Voice Conversion (EVC) technology aims to transfer emotional state in speech while keeping the linguistic information and speaker identity unchanged. Prior studies on EVC have been limited to perform the conversion for a specific speaker or a predefined set of multiple speakers seen in the training stage. When encountering arbitrary speakers that may be unseen during training (outside the set of speakers used in training), existing EVC methods have limited conversion capabilities. However, converting the emotion of arbitrary speakers, even those unseen during the training procedure, in one model is much more challenging and much more attractive in real-world scenarios. To address this problem, in this study, we propose SIEVC, a novel speaker-independent emotional voice conversion framework for arbitrary speakers via disentangled representation learning. The proposed method employs the autoencoder framework to disentangle the emotion information and emotion-independent information of each input speech into separated representation spaces. To achieve better disentanglement, we incorporate mutual information minimization into the training process. In addition, adversarial training is applied to enhance the quality of the generated audio signals. Finally, speaker-independent EVC for arbitrary speakers could be achieved by only replacing the emotion representations of source speech with the target ones. The experimental results demonstrate that the proposed EVC model outperforms the baseline models in terms of objective and subjective evaluation for both seen and unseen speakers.

Hierarchical Multi-Class Classification of Voice Disorders Using Self-Supervised Models and Glottal Features

Previous studies on the automatic classification of voice disorders have mostly investigated the binary classification task, which aims to distinguish pathological voice from healthy voice. Using multi-class classifiers, however, more fine-grained identification of voice disorders can be achieved, which is more helpful for clinical practitioners. Unfortunately, there is little publicly available training data for many voice disorders, which lowers the classification performance on data from unseen speakers. Earlier studies have shown that the usage of glottal source features can reduce data redundancy in detection of laryngeal voice disorders. Another approach to tackle the problems caused by scarcity of training data is to utilize deep learning models, such as wav2vec 2.0 and HuBERT, that have been pre-trained on larger databases. Since the aforementioned approaches have not been thoroughly studied in the multi-class classification of voice disorders, they will be jointly studied in the present work. In addition, we study a hierarchical classifier, which enables task-wise feature optimization and more efficient utilization of data. In this work, the aforementioned three approaches are compared with traditional mel frequency cepstral coefficient (MFCC) features and one-vs-rest and one-vs-one SVM classifiers. The results in a 3-class classification problem between healthy voice and two laryngeal disorders (hyperfunctional dysphonia and vocal fold paresis) indicate that all the studied methods outperform the baselines. The best performance was achieved by using features from wav2vec 2.0 LARGE together with hierarchical classification. The balanced classification accuracy of the system was 62.77% for male speakers, and 55.36% for female speakers, which outperformed the baseline systems by an absolute improvement of 15.76% and 6.95% for male and female speakers, respectively.

AudioLM: A Language Modeling Approach to Audio Generation

We introduce AudioLM, a framework for high-quality audio generation with long-term consistency. AudioLM maps the input audio to a sequence of discrete tokens and casts audio generation as a language modeling task in this representation space. We show how existing audio tokenizers provide different trade-offs between reconstruction quality and long-term structure, and we propose a hybrid tokenization scheme to achieve both objectives. Namely, we leverage the discretized activations of a masked language model pre-trained on audio to capture long-term structure and the discrete codes produced by a neural audio codec to achieve high-quality synthesis. By training on large corpora of raw audio waveforms, AudioLM learns to generate natural and coherent continuations given short prompts. When trained on speech, and without any transcript or annotation, AudioLM generates syntactically and semantically plausible speech continuations while also maintaining speaker identity and prosody for unseen speakers. Furthermore, we demonstrate how our approach extends beyond speech by generating coherent piano music continuations, despite being trained without any symbolic representation of music.

Strong Generalized Speech Emotion Recognition Based on Effective Data Augmentation

The absence of labeled samples limits the development of speech emotion recognition (SER). Data augmentation is an effective way to address sample sparsity. However, there is a lack of research on data augmentation algorithms in the field of SER. In this paper, the effectiveness of classical acoustic data augmentation methods in SER is analyzed, based on which a strong generalized speech emotion recognition model based on effective data augmentation is proposed. The model uses a multi-channel feature extractor consisting of multiple sub-networks to extract emotional representations. Different kinds of augmented data that can effectively improve SER performance are fed into the sub-networks, and the emotional representations are obtained by the weighted fusion of the output feature maps of each sub-network. And in order to make the model robust to unseen speakers, we employ adversarial training to generalize emotion representations. A discriminator is used to estimate the Wasserstein distance between the feature distributions of different speakers and to force the feature extractor to learn the speaker-invariant emotional representations by adversarial training. The simulation experimental results on the IEMOCAP corpus show that the performance of the proposed method is 2-9% ahead of the related SER algorithm, which proves the effectiveness of the proposed method.

Modelling of a Speech-to-Text Recognition System for Air Traffic Control and NATO Air Command

&lt;p&gt;Accent invariance in speech recognition is a chal- lenging problem especially in the are of aviation. In this paper a speech recognition system is developed to transcribe accented speech between pilots and air traffic controllers. The system allows handling of accents in continuous speech by modelling phonemes using Hidden Markov Models (HMMs) with Gaussian mixture model (GMM) probability density functions for each state. These phonemes are used to build word models of the NATO phonetic alphabet as well as the numerals 0 to 9 with transcriptions obtained from the Carnegie Mellon University (CMU) pronouncing dictionary. Mel-Frequency Cepstral Co-efficients (MFCC) with delta and delta-delta coefficients are used for the feature extraction process. Amplitude normalisation and covariance scaling is implemented to improve recognition accuracy. A word error rate (WER) of 2% for seen speakers and 22% for unseen speakers is obtained.&lt;/p&gt; &lt;p&gt;&amp;nbsp;&lt;/p&gt;