Audio-visual speech recognition (AVSR) incorporates auditory and visual modalities to improve recognition accuracy, particularly in noisy environments where audio-only speech systems are insufficient. While previous research has largely addressed audio disruptions, few studies have dealt with visual corruptions, e.g., lip occlusions or blurred videos, which are also detrimental. To address this real-world challenge, we propose CAV2vec, a novel self-supervised speech representation learning framework particularly designed to handle audio-visual joint corruption. CAV2vec employs a self-distillation approach with a corrupted prediction task, where the student model learns to predict clean targets, generated by the teacher model, with corrupted input frames. Specifically, we suggest a unimodal multi-task learning, which distills cross-modal knowledge and aligns the corrupted modalities, by predicting clean audio targets with corrupted videos, and clean video targets with corrupted audios. This strategy mitigates the dispersion in the representation space caused by corrupted modalities, leading to more reliable and robust audio-visual fusion. Our experiments on robust AVSR benchmarks demonstrate that the corrupted representation learning method significantly enhances recognition accuracy across generalized environments involving various types of corruption. Our code is available at https://github.com/sungnyun/cav2vec.
Self-supervised learning (SSL) has achieved great success in speech-related tasks. While Transformer and Conformer architectures have dominated SSL backbones, encoders like Zipformer, which excel in automatic speech recognition (ASR), remain unexplored in SSL. Concurrently, inefficiencies in data processing within existing SSL training frameworks, such as fairseq, pose challenges in managing the growing volumes of training data. To address these issues, we propose k2SSL, an open-source framework that offers faster, more memory-efficient, and better-performing self-supervised speech representation learning, focusing on downstream ASR tasks. The optimized HuBERT and proposed Zipformer-based SSL systems exhibit substantial reductions in both training time and memory usage during SSL training. Experiments on LibriSpeech demonstrate that Zipformer Base significantly outperforms HuBERT and WavLM, achieving up to a 34.8% relative WER reduction compared to HuBERT Base after fine-tuning, along with a 3.5x pre-training speedup in GPU hours. When scaled to 60k hours of LibriLight data, Zipformer Large exhibits remarkable efficiency, matching HuBERT Large's performance while requiring only 5/8 pre-training steps.
In this paper, we present EH-MAM (Easy-to-Hard adaptive Masked Acoustic Modeling), a novel self-supervised learning approach for speech representation learning. In contrast to the prior methods that use random masking schemes for Masked Acoustic Modeling (MAM), we introduce a novel selective and adaptive masking strategy. Specifically, during SSL training, we progressively introduce harder regions to the model for reconstruction. Our approach automatically selects hard regions and is built on the observation that the reconstruction loss of individual frames in MAM can provide natural signals to judge the difficulty of solving the MAM pre-text task for that frame. To identify these hard regions, we employ a teacher model that first predicts the frame-wise losses and then decides which frames to mask. By learning to create challenging problems, such as identifying harder frames and solving them simultaneously, the model is able to learn more effective representations and thereby acquire a more comprehensive understanding of the speech. Quantitatively, EH-MAM outperforms several state-of-the-art baselines across various low-resource speech recognition and SUPERB benchmarks by 5%-10%. Additionally, we conduct a thorough analysis to show that the regions masked by EH-MAM effectively capture useful context across speech frames.
Self-supervised speech representation learning has become essential for extracting meaningful features from untranscribed audio. Recent advances highlight the potential of deriving discrete symbols from the features correlated with linguistic units, which enables text-less training across diverse tasks. In particular, sentence-level Self-Distillation of the pretrained HuBERT (SD-HuBERT) induces syllabic structures within latent speech frame representations extracted from an intermediate Transformer layer. In SD-HuBERT, sentence-level representation is accumulated from speech frame features through self-attention layers using a special CLS token. However, we observe that the information aggregated in the CLS token correlates more with speaker identity than with linguistic content. To address this, we propose a speech-only self-supervised fine-tuning approach that separates syllabic units from speaker information. Our method introduces speaker perturbation as data augmentation and adopts a frame-level training objective to prevent the CLS token from aggregating paralinguistic information. Experimental results show that our approach surpasses the current state-of-the-art method in most syllable segmentation and syllabic unit quality metrics on Librispeech, underscoring its effectiveness in promoting syllabic organization within speech-only models.
We present mHuBERT-147, the first general-purpose massively multilingual HuBERT speech representation model trained on 90K hours of clean, open-license data. To scale up the multi-iteration HuBERT approach, we use faiss-based clustering, achieving 5.2x faster label assignment than the original method. We also apply a new multilingual batching up-sampling strategy, leveraging both language and dataset diversity. After 3 training iterations, our compact 95M parameter mHuBERT-147 outperforms larger models trained on substantially more data. We rank second and first on the ML-SUPERB 10min and 1h leaderboards, with SOTA scores for 3 tasks. Across ASR/LID tasks, our model consistently surpasses XLS-R (300M params; 436K hours) and demonstrates strong competitiveness against the much larger MMS (1B params; 491K hours). Our findings indicate that mHuBERT-147 is a promising model for multilingual speech tasks, offering an unprecedented balance between high performance and parameter efficiency.
Self-supervised speech representation learning enables the extraction of meaningful features from raw waveforms. These features can then be efficiently used across multiple downstream tasks. However, two significant issues arise when considering the deployment of such methods ``in-the-wild": (i) Their large size, which can be prohibitive for edge applications; and (ii) their robustness to detrimental factors, such as noise and/or reverberation, that can heavily degrade the performance of such systems. In this work, we propose RobustDistiller, a novel knowledge distillation mechanism that tackles both problems jointly. Simultaneously to the distillation recipe, we apply a multi-task learning objective to encourage the network to learn noise-invariant representations by denoising the input. The proposed mechanism is evaluated on twelve different downstream tasks. It outperforms several benchmarks regardless of noise type, or noise and reverberation levels. Experimental results show that the new Student model with 23M parameters can achieve results comparable to the Teacher model with 95M parameters. Lastly, we show that the proposed recipe can be applied to other distillation methodologies, such as the recent DPWavLM. For reproducibility, code and model checkpoints will be made available at \mbox{\url{https://github.com/Hguimaraes/robustdistiller}}.
Foundation models in speech are often trained using many GPUs, which implicitly leads to large effective batch sizes. In this paper we study the effect of batch size on pre-training, both in terms of statistics that can be monitored during training, and in the effect on the performance of a downstream fine-tuning task. By using batch sizes varying from 87.5 seconds to 80 minutes of speech we show that, for a fixed amount of iterations, larger batch sizes result in better pre-trained models. However, there is lower limit for stability, and an upper limit for effectiveness. We then show that the quality of the pre-trained model depends mainly on the amount of speech data seen during training, i.e., on the product of batch size and number of iterations. All results are produced with an independent implementation of the wav2vec 2.0 architecture, which to a large extent reproduces the results of the original work (arXiv:2006.11477). Our extensions can help researchers choose effective operating conditions when studying self-supervised learning in speech, and hints towards benchmarking self-supervision with a fixed amount of seen data. Code and model checkpoints are available at https://github.com/nikvaessen/w2v2-batch-size.
Multilingual speech processing requires understanding emotions, a task made difficult by limited labelled data. CLARA, minimizes reliance on labelled data, enhancing generalization across languages. It excels at fostering shared representations, aiding cross-lingual transfer of speech and emotions, even with little data. Our approach adeptly captures emotional nuances in speech, overcoming subjective assessment issues. Using a large multilingual audio corpus and self-supervised learning, CLARA develops speech representations enriched with emotions, advancing emotion-aware multilingual speech processing. Our method expands the data range using data augmentation, textual embedding for visual understanding, and transfers knowledge from high- to low-resource languages. CLARA demonstrates excellent performance in emotion recognition, language comprehension, and audio benchmarks, excelling in zero-shot and few-shot learning. It adapts to low-resource languages, marking progress in multilingual speech representation learning.
In this paper, we present a methodology for linguistic feature extraction, focusing particularly on automatically syllabifying words in multiple languages, with a design to be compatible with a forced-alignment tool, the Montreal Forced Aligner (MFA). In both the textual and phonetic domains, our method focuses on the extraction of phonetic transcriptions from text, stress marks, and a unified automatic syllabification (in text and phonetic domains). The system was built with open-source components and resources. Through an ablation study, we demonstrate the efficacy of our approach in automatically syllabifying words from several languages (English, French and Spanish). Additionally, we apply the technique to the transcriptions of the CMU ARCTIC dataset, generating valuable annotations available online\footnote{\url{https://github.com/noetits/MUST_P-SRL}} that are ideal for speech representation learning, speech unit discovery, and disentanglement of speech factors in several speech-related fields.
Recent years have witnessed significant advancements in self-supervised learning (SSL) methods for speech-processing tasks. Various speech-based SSL models have been developed and present promising performance on a range of downstream tasks including speech recognition. However, existing speech-based SSL models face a common dilemma in terms of computational cost, which might hinder their potential application and in-depth academic research. To address this issue, we first analyze the computational cost of different modules during HuBERT pre-training and then introduce a stack of efficiency optimizations, which is named Fast-HuBERT in this paper. The proposed Fast-HuBERT can be trained in 1.1 days with 8 V100 GPUs on the Librispeech 960h benchmark, without performance degradation, resulting in a 5.2x speedup, compared to the original implementation. Moreover, we explore two well-studied techniques in the Fast-HuBERT and demonstrate consistent improvements as reported in previous work.
In this paper, we introduce self-distillation and online clustering for self-supervised speech representation learning (DinoSR) which combines masked language modeling, self-distillation, and online clustering. We show that these concepts complement each other and result in a strong representation learning model for speech. DinoSR first extracts contextualized embeddings from the input audio with a teacher network, then runs an online clustering system on the embeddings to yield a machine-discovered phone inventory, and finally uses the discretized tokens to guide a student network. We show that DinoSR surpasses previous state-of-the-art performance in several downstream tasks, and provide a detailed analysis of the model and the learned discrete units.
In this paper, we present a multimodal and dynamical VAE (MDVAE) applied to unsupervised audio-visual speech representation learning. The latent space is structured to dissociate the latent dynamical factors that are shared between the modalities from those that are specific to each modality. A static latent variable is also introduced to encode the information that is constant over time within an audiovisual speech sequence. The model is trained in an unsupervised manner on an audiovisual emotional speech dataset, in two stages. In the first stage, a vector quantized VAE (VQ-VAE) is learned independently for each modality, without temporal modeling. The second stage consists in learning the MDVAE model on the intermediate representation of the VQ-VAEs before quantization. The disentanglement between static versus dynamical and modality-specific versus modality-common information occurs during this second training stage. Extensive experiments are conducted to investigate how audiovisual speech latent factors are encoded in the latent space of MDVAE. These experiments include manipulating audiovisual speech, audiovisual facial image denoising, and audiovisual speech emotion recognition. The results show that MDVAE effectively combines the audio and visual information in its latent space. They also show that the learned static representation of audiovisual speech can be used for emotion recognition with few labeled data, and with better accuracy compared with unimodal baselines and a state-of-the-art supervised model based on an audiovisual transformer architecture.
This paper presents FaceXHuBERT, a text-less speech-driven 3D facial animation generation method that allows to capture personalized and subtle cues in speech (e.g. identity, emotion and hesitation). It is also very robust to background noise and can handle audio recorded in a variety of situations (e.g. multiple people speaking). Recent approaches employ end-to-end deep learning taking into account both audio and text as input to generate facial animation for the whole face. However, scarcity of publicly available expressive audio-3D facial animation datasets poses a major bottleneck. The resulting animations still have issues regarding accurate lip-synching, expressivity, person-specific information and generalizability. We effectively employ self-supervised pretrained HuBERT model in the training process that allows us to incorporate both lexical and non-lexical information in the audio without using a large lexicon. Additionally, guiding the training with a binary emotion condition and speaker identity distinguishes the tiniest subtle facial motion. We carried out extensive objective and subjective evaluation in comparison to ground-truth and state-of-the-art work. A perceptual user study demonstrates that our approach produces superior results with respect to the realism of the animation 78% of the time in comparison to the state-of-the-art. In addition, our method is 4 times faster eliminating the use of complex sequential models such as transformers. We strongly recommend watching the supplementary video before reading the paper. We also provide the implementation and evaluation codes with a GitHub repository link.
The transformer is a fundamental building block in deep learning, and the attention mechanism is the transformer's core component. Self-supervised speech representation learning (SSRL) represents a popular use-case for the transformer architecture. Due to transformers' acausal behavior, the use of transformers for SSRL has been predominantly focused on acausal applications. However, several media processing problems, such as speech processing, require real-time solutions. In this paper, we present an implementation of the attention module that enables training of SSRL architectures with low compute and memory requirements, while allowing real-time inference with low and fixed latency. The attention module proposed in this paper includes two components, streaming attention (SA) and low-latency streaming attention (LLSA). The SA represents our proposal for an efficient streaming SSRL implementation, while the LLSA solves the latency build-up problem of other streaming attention architectures, such as the masked acausal attention (MAA), guaranteeing a latency equal to one layer even when multiple layers are stacked. We present a comparative analysis between the vanilla attention, which we will refer here as acausal attention (AA), the SA, and the LLSA, by training a streaming SSRL with automatic speech recognition as downstream task. When training on librispeech-clean-100 and testing on librispeech-test-clean, our low-latency attention module has a word error rate (WER) of 5.84%, which represents a significant improvement over the MAA (WER = 13.82%). Our implementation also reduces the inference latency from 1.92 to 0.16 seconds. The proposed low-latency module preserves many of the benefits of conventional acausal transformers, but also enables latency characteristics that make it applicable to real-time streaming applications.
In this paper, we provide a new perspective on self-supervised speech models from how the training targets are obtained. We generalize the targets extractor into Offline Targets Extractor (Off-TE) and Online Targets Extractor (On-TE). Based on this, we propose a new multi-tasking learning framework for self-supervised learning, MT4SSL, which stands for Boosting Self-Supervised Speech Representation Learning by Integrating Multiple Targets. MT4SSL uses the K-means algorithm as an Off-TE and a teacher network without gradients as an On-TE, respectively. Our model outperforms previous SSL methods by nontrivial margins on the LibriSpeech benchmark, and is comparable to or even better than the best-performing models with fewer data. Furthermore, we find that using both Off-TE and On-TE results in better convergence in the pre-training phase. With both effectiveness and efficiency, we think doing multi-task learning on self-supervised speech models from our perspective is a promising trend.
Speech representation learning has improved both speech understanding and speech synthesis tasks for single language. However, its ability in cross-lingual scenarios has not been explored. In this paper, we extend the pretraining method for cross-lingual multi-speaker speech synthesis tasks, including cross-lingual multi-speaker voice cloning and cross-lingual multi-speaker speech editing. We propose a speech-text joint pretraining framework, where we randomly mask the spectrogram and the phonemes given a speech example and its transcription. By learning to reconstruct the masked parts of the input in different languages, our model shows great improvements over speaker-embedding-based multi-speaker TTS methods. Moreover, our framework is end-to-end for both the training and the inference without any finetuning effort. In cross-lingual multi-speaker voice cloning and cross-lingual multi-speaker speech editing tasks, our experiments show that our model outperforms speaker-embedding-based multi-speaker TTS methods.
In this paper, we propose a new Self-Supervised Learning (SSL) algorithm called data2vec-aqc, for speech representation learning from unlabeled speech data. Our goal is to improve SSL for speech in domains where both unlabeled and labeled data are limited. Building on the recently introduced data2vec, we introduce additional modules to the data2vec framework that leverage the benefit of data augmentations, quantized representations, and clustering. The interaction between these modules helps solve the cross-contrastive loss as an additional self-supervised objective. data2vec-aqc achieves up to 14.1% and 20.9% relative WER improvement over the existing state-of-the-art data2vec system over the test-clean and test-other sets, respectively of LibriSpeech, without the use of any language model (LM). Our proposed model also achieves up to 17.8\% relative WER gains over the baseline data2vec when fine-tuned on a subset of the Switchboard dataset. Code: https://github.com/Speech-Lab-IITM/data2vec-aqc.
Self-supervised pre-training methods based on contrastive learning or regression tasks can utilize more unlabeled data to improve the performance of automatic speech recognition (ASR). However, the robustness impact of combining the two pre-training tasks and constructing different negative samples for contrastive learning still remains unclear. In this paper, we propose a noise-robust data2vec for self-supervised speech representation learning by jointly optimizing the contrastive learning and regression tasks in the pre-training stage. Furthermore, we present two improved methods to facilitate contrastive learning. More specifically, we first propose to construct patch-based non-semantic negative samples to boost the noise robustness of the pre-training model, which is achieved by dividing the features into patches at different sizes (i.e., so-called negative samples). Second, by analyzing the distribution of positive and negative samples, we propose to remove the easily distinguishable negative samples to improve the discriminative capacity for pre-training models. Experimental results on the CHiME-4 dataset show that our method is able to improve the performance of the pre-trained model in noisy scenarios. We find that joint training of the contrastive learning and regression tasks can avoid the model collapse to some extent compared to only training the regression task.
Self-supervised learning (SSL) speech models, which can serve as powerful upstream models to extract meaningful speech representations, have achieved unprecedented success in speech representation learning. However, their effectiveness on non-speech datasets is relatively less explored. In this work, we propose an ensemble framework, with a combination of ensemble techniques, to fuse SSL speech models' embeddings. Extensive experiments on speech and non-speech audio datasets are conducted to investigate the representation abilities of our ensemble method and its single constituent model. Ablation studies are carried out to evaluate the performances of different ensemble techniques, such as feature averaging and concatenation. All experiments are conducted during NeurIPS 2021 HEAR Challenge as a standard evaluation pipeline provided by competition officials. Results demonstrate SSL speech models' strong abilities on various non-speech tasks, while we also note that they fail to deal with fine-grained music tasks, such as pitch classification and note onset detection. In addition, feature ensemble is shown to have great potential on producing more holistic representations, as our proposed framework generally surpasses state-of-the-art SSL speech/audio models and has superior performance on various datasets compared with other teams in HEAR Challenge. Our code is available at https://github.com/tony10101105/HEAR-2021-NeurIPS-Challenge -- NTU-GURA.
Contrastive learning enables learning useful audio and speech representations without ground-truth labels by maximizing the similarity between latent representations of similar signal segments. In this framework various data augmentation techniques are usually exploited to help enforce desired invariances within the learned representations, improving performance on various audio tasks thanks to more robust embeddings. Now, selecting the most relevant augmentations has proven crucial for better downstream performances. Thus, this work introduces a conditional independance-based method which allows for automatically selecting a suitable distribution on the choice of augmentations and their parametrization from a set of predefined ones, for contrastive self-supervised pre-training. This is performed with respect to a downstream task of interest, hence saving a costly hyper-parameter search. Experiments performed on two different downstream tasks validate the proposed approach showing better results than experimenting without augmentation or with baseline augmentations. We furthermore conduct a qualitative analysis of the automatically selected augmentations and their variation according to the considered final downstream dataset.