Recent work suggests that large language models (LLMs) can improve performance of speech tasks compared to existing systems. To support their claims, results on LibriSpeech and Common Voice are often quoted. However, this work finds that a substantial amount of the LibriSpeech and Common Voice evaluation sets appear in public LLM pretraining corpora. This calls into question the reliability of findings drawn from these two datasets. To measure contamination impact, LLMs trained with/without contamination are compared. A contaminated LLM is more likely to generate test sentences it has seen during training. Then, speech recognisers based on LLMs are compared. They show only subtle error rate differences if the LLM is contaminated, but assign significantly higher probabilities to transcriptions seen during LLM training. Results show that LLM outputs can be biased by tiny amounts of data contamination, highlighting the importance of evaluating LLM-based speech systems with held-out data.

Self-attention relies on positional embeddings to encode input order. Relative Position (RelPos) embeddings are widely used in Automatic Speech Recognition (ASR). However, RelPos has quadratic time complexity to input length and is often incompatible with fast GPU implementations of attention. In contrast, Rotary Positional Embedding (RoPE) rotates each input vector based on its absolute position, taking linear time to sequence length, implicitly encoding relative distances through self-attention dot products. Thus, it is usually compatible with efficient attention. However, its use in ASR remains underexplored. This work evaluates RoPE across diverse ASR tasks with training data ranging from 100 to 50,000 hours, covering various speech types (read, spontaneous, clean, noisy) and different accents in both streaming and non-streaming settings. ASR error rates are similar or better than RelPos, while training time is reduced by up to $21 \%$. Code is available via the SpeechBrain toolkit.

Automatic speech recognition (ASR) with an encoder equipped with self-attention, whether streaming or non-streaming, takes quadratic time in the length of the speech utterance. This slows down training and decoding, increase the cost, and limits the deployment of the ASR in constrained devices. SummaryMixing is a promising linear-time complexity alternative to self-attention for non-streaming speech recognition that, for the first time, preserves or outperforms the accuracy of self-attention models. Unfortunately, the original definition of SummaryMixing is not suited to streaming speech recognition. Hence, this work extends SummaryMixing to a Conformer Transducer that works in both a streaming and an offline mode. It shows that this new linear-time complexity speech encoder outperforms self-attention in both scenarios while requiring less compute and memory during training and decoding.

SpeechBrain is an open-source Conversational AI toolkit based on PyTorch, focused particularly on speech processing tasks such as speech recognition, speech enhancement, speaker recognition, text-to-speech, and much more. It promotes transparency and replicability by releasing both the pre-trained models and the complete recipes of code and algorithms required for training them. This paper presents SpeechBrain 1.0, a significant milestone in the evolution of the toolkit, which now has over 200 recipes for speech, audio, and language processing tasks, and more than 100 models available on Hugging Face. SpeechBrain 1.0 introduces new technologies to support diverse learning modalities, Large Language Model (LLM) integration, and advanced decoding strategies, along with novel models, tasks, and modalities. It also includes a new benchmark repository, offering researchers a unified platform for evaluating models across diverse tasks.

International audience

In this paper, we propose an online attention mechanism, known as cumulative attention (CA), for streaming Transformer-based automatic speech recognition (ASR). Inspired by monotonic chunk-wise attention (MoChA) and head-synchronous decoder-end adaptive computation steps (HS-DACS) algorithms, CA triggers the ASR outputs based on the acoustic information accumulated at each encoding timestep, where the decisions are made using a trainable device, referred to as halting selector. In CA, all the attention heads of the same decoder layer are synchronised to have a unified halting position. This feature effectively alleviates the problem caused by the distinct behaviour of individual heads, which may otherwise give rise to severe latency issues as encountered by MoChA. The ASR experiments conducted on AIShell-1 and Librispeech datasets demonstrate that the proposed CA-based Transformer system can achieve on par or better performance with significant reduction in latency during inference, when compared to other streaming Transformer systems in literature.

Although the lower layers of a deep neural network learn features which are transferable across datasets, these layers are not transferable within the same dataset. That is, in general, freezing the trained feature extractor (the lower layers) and retraining the classifier (the upper layers) on the same dataset leads to worse performance. In this paper, for the first time, we show that the frozen classifier is transferable within the same dataset. We develop a novel top-down training method which can be viewed as an algorithm for searching for high-quality classifiers. We tested this method on automatic speech recognition (ASR) tasks and language modelling tasks. The proposed method consistently improves recurrent neural network ASR models on Wall Street Journal, self-attention ASR models on Switchboard, and AWD-LSTM language models on WikiText-2.

Recently, Transformer based models have shown competitive automatic speech recognition (ASR) performance. One key factor in the success of these models is the multi-head attention mechanism. However, for trained models, we have previously observed that many attention matrices are close to diagonal, indicating the redundancy of the corresponding attention heads. We have also found that some architectures with reduced numbers of attention heads have better performance. Since the search for the best structure is time prohibitive, we propose to randomly remove attention heads during training and keep all attention heads at test time, thus the final model is an ensemble of models with different architectures. The proposed method also forces each head independently learn the most useful patterns. We apply the proposed method to train Transformer based and Convolution-augmented Transformer (Conformer) based ASR models. Our method gives consistent performance gains over strong baselines on the Wall Street Journal, AISHELL, Switchboard and AMI datasets. To the best of our knowledge, we have achieved state-of-the-art end-to-end Transformer based model performance on Switchboard and AMI.

Self-attention models such as Transformers, which can capture temporal relationships without being limited by the distance between events, have given competitive speech recognition results. However, we note the range of the learned context increases from the lower to upper self-attention layers, whilst acoustic events often happen within short time spans in a left-to-right order. This leads to a question: for speech recognition, is a global view of the entire sequence useful for the upper self-attention encoder layers in Transformers? To investigate this, we train models with lower self-attention/upper feed-forward layers encoders on Wall Street Journal and Switchboard. Compared to baseline Transformers, no performance drop but minor gains are observed. We further developed a novel metric of the diagonality of attention matrices and found the learned diagonality indeed increases from the lower to upper encoder self-attention layers. We conclude the global view is unnecessary in training upper encoder layers.

End-to-end models yield impressive speech recognition results on clean datasets while having inferior performance on noisy datasets. To address this, we propose transfer learning from a clean dataset (WSJ) to a noisy dataset (CHiME4) for connectionist temporal classification models. We argue that the clean classifier (the upper layers of a neural network trained on clean data) can force the feature extractor (the lower layers) to learn the underlying noise invariant patterns in the noisy dataset. While training on the noisy dataset, the clean classifier is either frozen or trained with a small learning rate. The feature extractor is trained with no learning rate re-scaling. The proposed method gives up to 15.5% relative character error rate (CER) reduction compared to models trained only on CHiME-4. Furthermore, we use the test sets of Aurora-4 to perform evaluation on unseen noisy conditions. Our method has significantly lower CERs (11.3% relative on average) on all 14 Aurora-4 test sets compared to the conventional transfer learning method (no learning rate rescale for any layer), indicating our method enables the model to learn noise invariant features.

The usual attention mechanisms used for encoder-decoder models do not constrain the relationship between input and output sequences to be monotonic. To address this we explore windowed attention mechanisms which restrict attention to a block of source hidden states. Rule-based windowing restricts attention to a (typically large) fixed-length window. The performance of such methods is poor if the window size is small. In this paper, we propose a fully-trainable windowed attention and provide a detailed analysis on the factors which affect the performance of such an attention mechanism. Compared to the rule-based window methods, the learned window size is significantly smaller yet the model's performance is competitive. On the TIMIT corpus this approach has resulted in a 17% (relative) performance improvement over the traditional attention model. Our model also yields comparable accuracies to the joint CTC-attention model on the Wall Street Journal corpus.

Jointly optimised attention-based encoder-decoder models have yielded impressive speech recognition results. The recurrent neural network (RNN) encoder is a key component in such models – it learns the hidden representations of the inputs.However, it is difficult for RNNs to model the long sequences characteristic of speech recognition. To address this, subsampling between stacked recurrent layers of the encoder is commonly employed. This method reduces the length of the input sequence and leads to gains in accuracy. However, static subsampling may both include redundant information and miss relevant information. We propose using a dynamic subsampling RNN (dsRNN) encoder. Unlike a statically subsampled RNN encoder, the dsRNN encoder can learn to skip redundant frames. Furthermore, the skip ratio may vary at different stages of training, thus allowing the encoder to learn the most relevant information for each epoch. Although the dsRNN is unidirectional, it yields lower phone error rates (PERs) than a bidirectional RNN on TIMIT. The dsRNN encoder has a 16.8% PER on the TIMIT test set, a considerable improvement over static subsampling methods used with unidirectional and bidirectional RNN encoders (23.5% and 20.4% PER respectively).