This paper presents an efficient decoding approach for end-to-end automatic speech recognition (E2E-ASR) with large language models (LLMs). Although shallow fusion is the most common approach to incorporate language models into E2E-ASR decoding, we face two practical problems with LLMs. (1) LLM inference is computationally costly. (2) There may be a vocabulary mismatch between the ASR model and the LLM. To resolve this mismatch, we need to retrain the ASR model and/or the LLM, which is at best time-consuming and in many cases not feasible. We propose "delayed fusion," which applies LLM scores to ASR hypotheses with a delay during decoding and enables easier use of pre-trained LLMs in ASR tasks. This method can reduce not only the number of hypotheses scored by the LLM but also the number of LLM inference calls. It also allows re-tokenizion of ASR hypotheses during decoding if ASR and LLM employ different tokenizations. We demonstrate that delayed fusion provides improved decoding speed and accuracy compared to shallow fusion and N-best rescoring using the LibriHeavy ASR corpus and three public LLMs, OpenLLaMA 3B & 7B and Mistral 7B.

We propose a novel approach to optimizing a byte-level representation for end-to-end automatic speech recognition (ASR). Byte-level representation is often used by large scale multilingual ASR systems when the character set of the supported languages is large. The compactness and universality of byte-level representation allow the ASR models to use smaller output vocabularies and therefore, provide more flexibility. UTF-8 is a commonly used byte-level representation for multilingual ASR, but it is not designed to optimize machine learning tasks directly. By using auto-encoder and vector quantization, we show that we can optimize a byte-level representation for ASR and achieve better accuracy. Our proposed framework can incorporate information from different modalities, and provides an error correction mechanism. In an English/Mandarin dictation task, we show that a bilingual ASR model built with this approach can outperform UTF-8 representation by 5% relative in error rate.

Language models (LMs) have long been used to improve results of automatic speech recognition (ASR) systems, but they are unaware of the errors that ASR systems make. Error correction models are designed to fix ASR errors, however, they showed little improvement over traditional LMs mainly due to the lack of supervised training data. In this paper, we present Denoising LM (DLM), which is a $\textit{scaled}$ error correction model trained with vast amounts of synthetic data, significantly exceeding prior attempts meanwhile achieving new state-of-the-art ASR performance. We use text-to-speech (TTS) systems to synthesize audio, which is fed into an ASR system to produce noisy hypotheses, which are then paired with the original texts to train the DLM. DLM has several $\textit{key ingredients}$: (i) up-scaled model and data; (ii) usage of multi-speaker TTS systems; (iii) combination of multiple noise augmentation strategies; and (iv) new decoding techniques. With a Transformer-CTC ASR, DLM achieves 1.5% word error rate (WER) on $\textit{test-clean}$ and 3.3% WER on $\textit{test-other}$ on Librispeech, which to our knowledge are the best reported numbers in the setting where no external audio data are used and even match self-supervised methods which use external audio data. Furthermore, a single DLM is applicable to different ASRs, and greatly surpassing the performance of conventional LM based beam-search rescoring. These results indicate that properly investigated error correction models have the potential to replace conventional LMs, holding the key to a new level of accuracy in ASR systems.

This work studies the use of attention masking in transformer transducer based speech recognition for building a single configurable model for different deployment scenarios. We present a comprehensive set of experiments comparing fixed masking, where the same attention mask is applied at every frame, with chunked masking, where the attention mask for each frame is determined by chunk boundaries, in terms of recognition accuracy and latency. We then explore the use of variable masking, where the attention masks are sampled from a target distribution at training time, to build models that can work in different configurations. Finally, we investigate how a single configurable model can be used to perform both first pass streaming recognition and second pass acoustic rescoring. Experiments show that chunked masking achieves a better accuracy vs latency trade-off compared to fixed masking, both with and without FastEmit. We also show that variable masking improves the accuracy by up to 8% relative in the acoustic re-scoring scenario.

The neural transducer is an end-to-end model for automatic speech recognition (ASR). While the model is well-suited for streaming ASR, the training process remains challenging. During training, the memory requirements may quickly exceed the capacity of state-of-the-art GPUs, limiting batch size and sequence lengths. In this work, we analyze the time and space complexity of a typical transducer training setup. We propose a memory-efficient training method that computes the transducer loss and gradients sample by sample. We present optimizations to increase the efficiency and parallelism of the sample-wise method. In a set of thorough benchmarks, we show that our sample-wise method significantly reduces memory usage, and performs at competitive speed when compared to the default batched computation. As a highlight, we manage to compute the transducer loss and gradients for a batch size of 1024, and audio length of 40 seconds, using only 6 GB of memory.