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.

On-device Virtual Assistants (VAs) powered by Automatic Speech Recognition (ASR) require effective knowledge integration for the challenging entity-rich query recognition. In this paper, we conduct an empirical study of modeling strategies for server-side rescoring of spoken information domain queries using various categories of Language Models (LMs) (N-gram word LMs, sub-word neural LMs). We investigate the combination of on-device and server-side signals, and demonstrate significant WER improvements of 23%-35% on various entity-centric query subpopulations by integrating various server-side LMs compared to performing ASR on-device only. We also perform a comparison between LMs trained on domain data and a GPT-3 variant offered by OpenAI as a baseline. Furthermore, we also show that model fusion of multiple server-side LMs trained from scratch most effectively combines complementary strengths of each model and integrates knowledge learned from domain-specific data to a VA ASR system.

With the development of deep learning, automatic speech recognition (ASR) has made significant progress. To further enhance the performance, revising recognition results is one of the lightweight but efficient manners. Various methods can be roughly classified into N-best reranking methods and error correction models. The former aims to select the hypothesis with the lowest error rate from a set of candidates generated by ASR for a given input speech. The latter focuses on detecting recognition errors in a given hypothesis and correcting these errors to obtain an enhanced result. However, we observe that these studies are hardly comparable to each other as they are usually evaluated on different corpora, paired with different ASR models, and even use different datasets to train the models. Accordingly, we first concentrate on releasing an ASR hypothesis revising (HypR) dataset in this study. HypR contains several commonly used corpora (AISHELL-1, TED-LIUM 2, and LibriSpeech) and provides 50 recognition hypotheses for each speech utterance. The checkpoint models of the ASR are also published. In addition, we implement and compare several classic and representative methods, showing the recent research progress in revising speech recognition results. We hope the publicly available HypR dataset can become a reference benchmark for subsequent research and promote the school of research to an advanced level.

In end-to-end (E2E) speech recognition models, a representational tight-coupling inevitably emerges between the encoder and the decoder. We build upon recent work that has begun to explore building encoders with modular encoded representations, such that encoders and decoders from different models can be stitched together in a zero-shot manner without further fine-tuning. While previous research only addresses full-context speech models, we explore the problem in a streaming setting as well. Our framework builds on top of existing encoded representations, converting them to modular features, dubbed as Lego-Features, without modifying the pre-trained model. The features remain interchangeable when the model is retrained with distinct initializations. Though sparse, we show that the Lego-Features are powerful when tested with RNN-T or LAS decoders, maintaining high-quality downstream performance. They are also rich enough to represent the first-pass prediction during two-pass deliberation. In this scenario, they outperform the N-best hypotheses, since they do not need to be supplemented with acoustic features to deliver the best results. Moreover, generating the Lego-Features does not require beam search or auto-regressive computation. Overall, they present a modular, powerful and cheap alternative to the standard encoder output, as well as the N-best hypotheses.

The recent emergence of joint CTC-Attention model shows significant improvement in automatic speech recognition (ASR). The improvement largely lies in the modeling of linguistic information by decoder. The decoder joint-optimized with an acoustic encoder renders the language model from ground-truth sequences in an auto-regressive manner during training. However, the training corpus of the decoder is limited to the speech transcriptions, which is far less than the corpus needed to train an acceptable language model. This leads to poor robustness of decoder. To alleviate this problem, we propose linguistic-enhanced transformer, which introduces refined CTC information to decoder during training process, so that the decoder can be more robust. Our experiments on AISHELL-1 speech corpus show that the character error rate (CER) is relatively reduced by up to 7%. We also find that in joint CTC-Attention ASR model, decoder is more sensitive to linguistic information than acoustic information.