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.

Neural contextual biasing allows speech recognition models to leverage contextually relevant information, leading to improved transcription accuracy. However, the biasing mechanism is typically based on a cross-attention module between the audio and a catalogue of biasing entries, which means computational complexity can pose severe practical limitations on the size of the biasing catalogue and consequently on accuracy improvements. This work proposes an approximation to cross-attention scoring based on vector quantization and enables compute- and memory-efficient use of large biasing catalogues. We propose to use this technique jointly with a retrieval based contextual biasing approach. First, we use an efficient quantized retrieval module to shortlist biasing entries by grounding them on audio. Then we use retrieved entries for biasing. Since the proposed approach is agnostic to the biasing method, we investigate using full cross-attention, LLM prompting, and a combination of the two. We show that retrieval based shortlisting allows the system to efficiently leverage biasing catalogues of several thousands of entries, resulting in up to 71% relative error rate reduction in personal entity recognition. At the same time, the proposed approximation algorithm reduces compute time by 20% and memory usage by 85-95%, for lists of up to one million entries, when compared to standard dot-product cross-attention.

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.

This paper presents an extension to train end-to-end Context-Aware Transformer Transducer ( CATT ) models by using a simple, yet efficient method of mining hard negative phrases from the latent space of the context encoder. During training, given a reference query, we mine a number of similar phrases using approximate nearest neighbour search. These sampled phrases are then used as negative examples in the context list alongside random and ground truth contextual information. By including approximate nearest neighbour phrases (ANN-P) in the context list, we encourage the learned representation to disambiguate between similar, but not identical, biasing phrases. This improves biasing accuracy when there are several similar phrases in the biasing inventory. We carry out experiments in a large-scale data regime obtaining up to 7% relative word error rate reductions for the contextual portion of test data. We also extend and evaluate CATT approach in streaming applications.

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.

Very deep CNNs achieve state-of-the-art results in both computer vision and speech recognition, but are difficult to train. The most popular way to train very deep CNNs is to use shortcut connections (SC) together with batch normalization (BN). Inspired by Self-Normalizing Neural Networks, we propose the self-normalizing deep CNN (SNDCNN) based acoustic model topology, by removing the SC/BN and replacing the typical RELU activations with scaled exponential linear unit (SELU) in ResNet-50. SELU activations make the network self-normalizing and remove the need for both shortcut connections and batch normalization. Compared to ResNet-50, we can achieve the same or lower word error rate (WER) while at the same time improving both training and inference speed by 60%-80%. We also explore other model inference optimizations to further reduce latency for production use.