Self-supervised learning (SSL) models reshaped our approach to speech, language and vision. However their huge size and the opaque relations between their layers and tasks result in slow inference and network overthinking, where predictions made from the last layer of large models is worse than those made from intermediate layers. Early exit (EE) strategies can solve both issues by dynamically reducing computations at inference time for certain samples. Although popular for classification tasks in vision and language, EE has seen less use for sequence-to-sequence speech recognition (ASR) tasks where outputs from early layers are often degenerate. This challenge is further compounded when speech SSL models are applied on out-of-distribution (OOD) data. This paper first shows that SSL models do overthinking in ASR. We then motivate further research in EE by computing an optimal bound for performance versus speed trade-offs. To approach this bound we propose two new strategies for ASR: (1) we adapt the recently proposed patience strategy to ASR; and (2) we design a new EE strategy specific to ASR that performs better than all strategies previously introduced.
Automatic assessment of reading fluency using automatic speech recognition (ASR) holds great potential for early detection of reading difficulties and subsequent timely intervention. Precise assessment tools are required, especially for languages other than English. In this study, we evaluate six state-of-the-art ASR-based systems for automatically assessing Dutch oral reading accuracy using Kaldi and Whisper. Results show our most successful system reached substantial agreement with human evaluations (MCC = .63). The same system reached the highest correlation between forced decoding confidence scores and word correctness (r = .45). This system's language model (LM) consisted of manual orthographic transcriptions and reading prompts of the test data, which shows that including reading errors in the LM improves assessment performance. We discuss the implications for developing automatic assessment systems and identify possible avenues of future research.
Dual-encoder structure successfully utilizes two language-specific encoders (LSEs) for code-switching speech recognition. Because LSEs are initialized by two pre-trained language-specific models (LSMs), the dual-encoder structure can exploit sufficient monolingual data and capture the individual language attributes. However, existing methods have no language constraints on LSEs and underutilize language-specific knowledge of LSMs. In this paper, we propose a language-specific characteristic assistance (LSCA) method to mitigate the above problems. Specifically, during training, we introduce two language-specific losses as language constraints and generate corresponding language-specific targets for them. During decoding, we take the decoding abilities of LSMs into account by combining the output probabilities of two LSMs and the mixture model to obtain the final predictions. Experiments show that either the training or decoding method of LSCA can improve the model's performance. Furthermore, the best result can obtain up to 15.4% relative error reduction on the code-switching test set by combining the training and decoding methods of LSCA. Moreover, the system can process code-switching speech recognition tasks well without extra shared parameters or even retraining based on two pre-trained LSMs by using our method.
Streaming speech recognition architectures are employed for low-latency, real-time applications. Such architectures are often characterized by their causality. Causal architectures emit tokens at each frame, relying only on current and past signal, while non-causal models are exposed to a window of future frames at each step to increase predictive accuracy. This dichotomy amounts to a trade-off for real-time Automatic Speech Recognition (ASR) system design: profit from the low-latency benefit of strictly-causal architectures while accepting predictive performance limitations, or realize the modeling benefits of future-context models accompanied by their higher latency penalty. In this work, we relax the constraints of this choice and present the Adaptive Non-Causal Attention Transducer (ANCAT). Our architecture is non-causal in the traditional sense, but executes in a low-latency, streaming manner by dynamically choosing when to rely on future context and to what degree within the audio stream. The resulting mechanism, when coupled with our novel regularization algorithms, delivers comparable accuracy to non-causal configurations while improving significantly upon latency, closing the gap with their causal counterparts. We showcase our design experimentally by reporting comparative ASR task results with measures of accuracy and latency on both publicly accessible and production-scale, voice-assistant datasets.
Recent research shows a big convergence in model architecture, training objectives, and inference methods across various tasks for different modalities. In this paper, we propose VioLA, a single auto-regressive Transformer decoder-only network that unifies various cross-modal tasks involving speech and text, such as speech-to-text, text-to-text, text-to-speech, and speech-to-speech tasks, as a conditional codec language model task via multi-task learning framework. To accomplish this, we first convert all the speech utterances to discrete tokens (similar to the textual data) using an offline neural codec encoder. In such a way, all these tasks are converted to token-based sequence conversion problems, which can be naturally handled with one conditional language model. We further integrate task IDs (TID) and language IDs (LID) into the proposed model to enhance the modeling capability of handling different languages and tasks. Experimental results demonstrate that the proposed VioLA model can support both single-modal and cross-modal tasks well, and the decoder-only model achieves a comparable and even better performance than the strong baselines.
This paper introduces an audio-visual speech enhancement system that leverages score-based generative models, also known as diffusion models, conditioned on visual information. In particular, we exploit audio-visual embeddings obtained from a self-super\-vised learning model that has been fine-tuned on lipreading. The layer-wise features of its transformer-based encoder are aggregated, time-aligned, and incorporated into the noise conditional score network. Experimental evaluations show that the proposed audio-visual speech enhancement system yields improved speech quality and reduces generative artifacts such as phonetic confusions with respect to the audio-only equivalent. The latter is supported by the word error rate of a downstream automatic speech recognition model, which decreases noticeably, especially at low input signal-to-noise ratios.
Handwriting recognition has seen significant success with the use of deep learning. However, a persistent shortcoming of neural networks is that they are not well-equipped to deal with shifting data distributions. In the field of handwritten text recognition (HTR), this shows itself in poor recognition accuracy for writers that are not similar to those seen during training. An ideal HTR model should be adaptive to new writing styles in order to handle the vast amount of possible writing styles. In this paper, we explore how HTR models can be made writer adaptive by using only a handful of examples from a new writer (e.g., 16 examples) for adaptation. Two HTR architectures are used as base models, using a ResNet backbone along with either an LSTM or Transformer sequence decoder. Using these base models, two methods are considered to make them writer adaptive: 1) model-agnostic meta-learning (MAML), an algorithm commonly used for tasks such as few-shot classification, and 2) writer codes, an idea originating from automatic speech recognition. Results show that an HTR-specific version of MAML known as MetaHTR improves performance compared to the baseline with a 1.4 to 2.0 improvement in word error rate (WER). The improvement due to writer adaptation is between 0.2 and 0.7 WER, where a deeper model seems to lend itself better to adaptation using MetaHTR than a shallower model. However, applying MetaHTR to larger HTR models or sentence-level HTR may become prohibitive due to its high computational and memory requirements. Lastly, writer codes based on learned features or Hinge statistical features did not lead to improved recognition performance.
We describe a comprehensive methodology for developing user-voice personalised ASR models by effectively training models on mobile phones, allowing user data and models to be stored and used locally. To achieve this, we propose a resource-aware sub-model based training approach that considers the RAM, and battery capabilities of mobile phones. We also investigate the relationship between available resources and training time, highlighting the effectiveness of using sub-models in such scenarios. By taking into account the evaluation metric and battery constraints of the mobile phones, we are able to perform efficient training and halt the process accordingly. To simulate real users, we use speakers with various accents. The entire on-device training and evaluation framework was then tested on various mobile phones across brands. We show that fine-tuning the models and selecting the right hyperparameter values is a trade-off between the lowest achievable performance metric, on-device training time, and memory consumption. Overall, our methodology offers a comprehensive solution for developing personalized ASR models while leveraging the capabilities of mobile phones, and balancing the need for accuracy with resource constraints.
Audio and visual modalities are inherently connected in speech signals: lip movements and facial expressions are correlated with speech sounds. This motivates studies that incorporate the visual modality to enhance an acoustic speech signal or even restore missing audio information. Specifically, this paper focuses on the problem of audio-visual speech inpainting, which is the task of synthesizing the speech in a corrupted audio segment in a way that it is consistent with the corresponding visual content and the uncorrupted audio context. We present an audio-visual transformer-based deep learning model that leverages visual cues that provide information about the content of the corrupted audio. It outperforms the previous state-of-the-art audio-visual model and audio-only baselines. We also show how visual features extracted with AV-HuBERT, a large audio-visual transformer for speech recognition, are suitable for synthesizing speech.
Speech representation learning approaches for non-semantic tasks such as language recognition have either explored supervised embedding extraction methods using a classifier model or self-supervised representation learning approaches using raw data. In this paper, we propose a novel framework of combining self-supervised representation learning with the language label information for the pre-training task. This framework, termed as Label Aware Speech Representation (LASR) learning, uses a triplet based objective function to incorporate language labels along with the self-supervised loss function. The speech representations are further fine-tuned for the downstream task. The language recognition experiments are performed on two public datasets - FLEURS and Dhwani. In these experiments, we illustrate that the proposed LASR framework improves over the state-of-the-art systems on language identification. We also report an analysis of the robustness of LASR approach to noisy/missing labels as well as its application to multi-lingual speech recognition tasks.