Byte pair encoding (BPE) emerges as an effective tokenization method for tackling the out-of-vocabulary (OOV) challenge in various natural language and speech processing tasks. Recent research highlights the dependency of BPE subword tokenization's efficacy on the morphological nature of the language, particularly in languages rich in inflectional morphology, where fewer BPE merges suffice for generating highly productive tokens. Motivated by this, our study empirically identifies the optimal number of BPE tokens for Bengali, a language known for its morphological complexity, thus enhancing out-of-distribution automatic speech recognition (ASR) performance. Experimental evaluation reveals that an excessively high number of BPE tokens can lead to overfitting, while approximately 500-1000 tokens result in superior OOV performance. Furthermore, we conduct a comparative analysis of BPE with character-based and unigram-based tokenization methods. By introducing BPE tokenization to Bengali ASR, we achieve a substantial reduction in the word error rate (WER) from 66.44% in our character-based baseline system to 63.80% on the LB-ASRTD eval set and from 46.34% to 42.80% on the SHRUTI eval set, both of which include out-of-distribution data.
In this study, we present SeMaScore, generated using a segment-wise mapping and scoring algorithm that serves as an evaluation metric for automatic speech recognition tasks. SeMaScore leverages both the error rate and a more robust similarity score. We show that our algorithm's score generation improves upon the state-of-the-art BERTscore. Our experimental results show that SeMaScore corresponds well with expert human assessments, signal-to-noise ratio levels, and other natural language metrics. We outperform BERTscore by 41x in metric computation speed. Overall, we demonstrate that SeMaScore serves as a more dependable evaluation metric, particularly in real-world situations involving atypical speech patterns.
Automatic speech recognition (ASR) outcomes serve as input for downstream tasks, substantially impacting the satisfaction level of end-users. Hence, the diagnosis and enhancement of the vulnerabilities present in the ASR model bear significant importance. However, traditional evaluation methodologies of ASR systems generate a singular, composite quantitative metric, which fails to provide comprehensive insight into specific vulnerabilities. This lack of detail extends to the post-processing stage, resulting in further obfuscation of potential weaknesses. Despite an ASR model's ability to recognize utterances accurately, subpar readability can negatively affect user satisfaction, giving rise to a trade-off between recognition accuracy and user-friendliness. To effectively address this, it is imperative to consider both the speech-level, crucial for recognition accuracy, and the text-level, critical for user-friendliness. Consequently, we propose the development of an Error Explainable Benchmark (EEB) dataset. This dataset, while considering both speech- and text-level, enables a granular understanding of the model's shortcomings. Our proposition provides a structured pathway for a more `real-world-centric' evaluation, a marked shift away from abstracted, traditional methods, allowing for the detection and rectification of nuanced system weaknesses, ultimately aiming for an improved user experience.
Self-supervised learning (SSL) for automated speech recognition in terms of its emotional content, can be heavily degraded by the presence noise, affecting the efficiency of modeling the intricate temporal and spectral informative structures of speech. Recently, SSL on large speech datasets, as well as new audio-specific SSL proxy tasks, such as, temporal and frequency masking, have emerged, yielding superior performance compared to classic approaches drawn from the image augmentation domain. Our proposed contribution builds upon this successful paradigm by introducing CochCeps-Augment, a novel bio-inspired masking augmentation task for self-supervised contrastive learning of speech representations. Specifically, we utilize the newly introduced bio-inspired cochlear cepstrogram (CCGRAM) to derive noise robust representations of input speech, that are then further refined through a self-supervised learning scheme. The latter employs SimCLR to generate contrastive views of a CCGRAM through masking of its angle and quefrency dimensions. Our experimental approach and validations on the emotion recognition K-EmoCon benchmark dataset, for the first time via a speaker-independent approach, features unsupervised pre-training, linear probing and fine-tuning. Our results potentiate CochCeps-Augment to serve as a standard tool in speech emotion recognition analysis, showing the added value of incorporating bio-inspired masking as an informative augmentation task for self-supervision. Our code for implementing CochCeps-Augment will be made available at: https://github.com/GiannisZgs/CochCepsAugment.
Multi-talker automatic speech recognition plays a crucial role in scenarios involving multi-party interactions, such as meetings and conversations. Due to its inherent complexity, this task has been receiving increasing attention. Notably, the serialized output training (SOT) stands out among various approaches because of its simplistic architecture and exceptional performance. However, the frequent speaker changes in token-level SOT (t-SOT) present challenges for the autoregressive decoder in effectively utilizing context to predict output sequences. To address this issue, we introduce a masked t-SOT label, which serves as the cornerstone of an auxiliary training loss. Additionally, we utilize a speaker similarity matrix to refine the self-attention mechanism of the decoder. This strategic adjustment enhances contextual relationships within the same speaker's tokens while minimizing interactions between different speakers' tokens. We denote our method as speaker-aware SOT (SA-SOT). Experiments on the Librispeech datasets demonstrate that our SA-SOT obtains a relative cpWER reduction ranging from 12.75% to 22.03% on the multi-talker test sets. Furthermore, with more extensive training, our method achieves an impressive cpWER of 3.41%, establishing a new state-of-the-art result on the LibrispeechMix dataset.
Far-field speech recognition is a challenging task that conventionally uses signal processing beamforming to attack noise and interference problem. But the performance has been found usually limited due to heavy reliance on environmental assumption. In this paper, we propose a unified multichannel far-field speech recognition system that combines the neural beamforming and transformer-based Listen, Spell, Attend (LAS) speech recognition system, which extends the end-to-end speech recognition system further to include speech enhancement. Such framework is then jointly trained to optimize the final objective of interest. Specifically, factored complex linear projection (fCLP) has been adopted to form the neural beamforming. Several pooling strategies to combine look directions are then compared in order to find the optimal approach. Moreover, information of the source direction is also integrated in the beamforming to explore the usefulness of source direction as a prior, which is usually available especially in multi-modality scenario. Experiments on different microphone array geometry are conducted to evaluate the robustness against spacing variance of microphone array. Large in-house databases are used to evaluate the effectiveness of the proposed framework and the proposed method achieve 19.26\% improvement when compared with a strong baseline.
This paper delineates the visual speech recognition (VSR) system introduced by the NPU-ASLP-LiAuto (Team 237) in the first Chinese Continuous Visual Speech Recognition Challenge (CNVSRC) 2023, engaging in the fixed and open tracks of Single-Speaker VSR Task, and the open track of Multi-Speaker VSR Task. In terms of data processing, we leverage the lip motion extractor from the baseline1 to produce multi-scale video data. Besides, various augmentation techniques are applied during training, encompassing speed perturbation, random rotation, horizontal flipping, and color transformation. The VSR model adopts an end-to-end architecture with joint CTC/attention loss, comprising a ResNet3D visual frontend, an E-Branchformer encoder, and a Transformer decoder. Experiments show that our system achieves 34.76% CER for the Single-Speaker Task and 41.06% CER for the Multi-Speaker Task after multi-system fusion, ranking first place in all three tracks we participate.
While automatic speech recognition (ASR) systems degrade significantly in noisy environments, audio-visual speech recognition (AVSR) systems aim to complement the audio stream with noise-invariant visual cues and improve the system's robustness. However, current studies mainly focus on fusing the well-learned modality features, like the output of modality-specific encoders, without considering the contextual relationship during the modality feature learning. In this study, we propose a multi-layer cross-attention fusion based AVSR (MLCA-AVSR) approach that promotes representation learning of each modality by fusing them at different levels of audio/visual encoders. Experimental results on the MISP2022-AVSR Challenge dataset show the efficacy of our proposed system, achieving a concatenated minimum permutation character error rate (cpCER) of 30.57% on the Eval set and yielding up to 3.17% relative improvement compared with our previous system which ranked the second place in the challenge. Following the fusion of multiple systems, our proposed approach surpasses the first-place system, establishing a new SOTA cpCER of 29.13% on this dataset.
This paper explores sentence-level Multilingual Visual Speech Recognition with a single model for the first time. As the massive multilingual modeling of visual data requires huge computational costs, we propose a novel strategy, processing with visual speech units. Motivated by the recent success of the audio speech unit, the proposed visual speech unit is obtained by discretizing the visual speech features extracted from the self-supervised visual speech model. To correctly capture multilingual visual speech, we first train the self-supervised visual speech model on 5,512 hours of multilingual audio-visual data. Through analysis, we verify that the visual speech units mainly contain viseme information while suppressing non-linguistic information. By using the visual speech units as the inputs of our system, we pre-train the model to predict corresponding text outputs on massive multilingual data constructed by merging several VSR databases. As both the inputs and outputs are discrete, we can greatly improve the training efficiency compared to the standard VSR training. Specifically, the input data size is reduced to 0.016% of the original video inputs. In order to complement the insufficient visual information in speech recognition, we apply curriculum learning where the inputs of the system begin with audio-visual speech units and gradually change to visual speech units. After pre-training, the model is finetuned on continuous features. We set new state-of-the-art multilingual VSR performances by achieving comparable performances to the previous language-specific VSR models, with a single trained model.
Recent developments in natural language processing (NLP) have highlighted the need for substantial amounts of data for models to capture textual information accurately. This raises concerns regarding the computational resources and time required for training such models. This paper introduces Semantics for data SAliency in Model performance Estimation (SeSaME). It is an efficient data sampling mechanism solely based on textual information without passing the data through a compute-heavy model or other intensive pre-processing transformations. The application of this approach is demonstrated in the use case of low-resource automated speech recognition (ASR) models, which excessively rely on text-to-speech (TTS) calls when using augmented data. SeSaME learns to categorize new incoming data points into speech recognition difficulty buckets by employing semantic similarity-based graph structures and discrete ASR information from homophilous neighbourhoods through message passing. The results indicate reliable projections of ASR performance, with a 93% accuracy increase when using the proposed method compared to random predictions, bringing non-trivial information on the impact of textual representations in speech models. Furthermore, a series of experiments show both the benefits and challenges of using the ASR information on incoming data to fine-tune the model. We report a 7% drop in validation loss compared to random sampling, 7% WER drop with non-local aggregation when evaluating against a highly difficult dataset, and 1.8% WER drop with local aggregation and high semantic similarity between datasets.