Guan-Ting Lin, Prashanth Gurunath Shivakumar, Aditya Gourav, Yile Gu, Ankur Gandhe, Hung-yi Lee, Ivan Bulyko. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Speech Recognition has seen a dramatic shift towards adopting Large Language Models (LLMs). This shift is partly driven by good scalability properties demonstrated by LLMs, ability to leverage large amounts of labelled, unlabelled speech and text data, streaming capabilities with autoregressive framework and multi-tasking with instruction following characteristics of LLMs. However, simple next-token prediction objective, typically employed with LLMs, have certain limitations in performance and challenges with hallucinations. In this paper, we propose application of Group Relative Policy Optimization (GRPO) to enable reinforcement learning from human feedback for automatic speech recognition (ASR). We design simple rule based reward functions to guide the policy updates. We demonstrate significant improvements in word error rate (upto 18.4% relative), reduction in hallucinations, increased robustness on out-of-domain datasets and effectiveness in domain adaptation.

Large language models (LLM) have demonstrated the ability to understand human language by leveraging large amount of text data. Automatic speech recognition (ASR) systems are often limited by available transcribed speech data and benefit from a second pass rescoring using LLM. Recently multi-modal large language models, particularly speech and text foundational models have demonstrated strong spoken language understanding. Speech-Text foundational models leverage large amounts of unlabelled and labelled data both in speech and text modalities to model human language. In this work, we propose novel techniques to use multi-modal LLM for ASR rescoring. We also explore discriminative training to further improve the foundational model rescoring performance. We demonstrate cross-modal knowledge transfer in speech-text LLM can benefit rescoring. Our experiments demonstrate up-to 20% relative improvements over Whisper large ASR and up-to 15% relative improvements over text-only LLM.

In the realm of spoken language understanding (SLU). numerous natural language understanding (NLU) methodologies have been adapted by supplying large language models (LLMs) with transcribed speech instead of conventional written text. In real-world scenarios, prior to input into an LLM. an automated speech recognition (ASR) system generates an output transcript hypothesis, where inherent errors can degrade subsequent SLU tasks. Here we introduce a method that utilizes the ASR system's lattice output instead of relying solely on the top hypothesis, aiming to encapsulate speech ambiguities and enhance SLU outcomes. Our in-context learning experiments, covering spoken question answering and intent classification. underline the LLM's resilience to noisy speech transcripts with the help of word confusion networks from lattices, bridging the SLU performance gap between using the top ASR hypothesis and an oracle upper bound. Additionally, we delve into the LLM's robustness to varying ASR performance conditions and scrutinize the aspects of in-context learning which prove the most influential.

Large Language Models (LLMs) have demonstrated superior abilities in tasks such as chatting, reasoning, and question-answering. However, standard LLMs may ignore crucial paralinguistic information, such as sentiment, emotion, and speaking style, which are essential for achieving natural, human-like spoken conversation, especially when such information is conveyed by acoustic cues. We therefore propose Paralinguistics-enhanced Generative Pretrained Transformer (ParalinGPT), an LLM that utilizes text and speech modalities to better model the linguistic content and paralinguistic attributes of spoken dialogue. The model takes the conversational context of text, speech embeddings, and paralinguistic attributes as input prompts within a serialized multitasking multimodal framework. Specifically, our framework serializes tasks in the order of current paralinguistic attribute prediction, response paralinguistic attribute prediction, and response text generation with autoregressive conditioning. We utilize the Switchboard-1 corpus, including its sentiment labels as the paralinguistic attribute, as our spoken dialogue dataset. Experimental results indicate the proposed serialized multitasking method outperforms typical sequence classification techniques on current and response sentiment classification. Furthermore, leveraging conversational context and speech embeddings significantly improves both response text generation and sentiment prediction. Our proposed framework achieves relative improvements of 6.7%, 12.0%, and 3.5% in current sentiment accuracy, response sentiment accuracy, and response text BLEU score, respectively.

Second pass rescoring is a critical component of competitive automatic speech recognition (ASR) systems. Large language models have demonstrated their ability in using pre-trained information for better rescoring of ASR hypothesis. Discriminative training, directly optimizing the minimum word-error-rate (MWER) criterion typically improves rescoring. In this study, we propose and explore several discriminative fine-tuning schemes for pre-trained LMs. We propose two architectures based on different pooling strategies of output embeddings and compare with probability based MWER. We conduct detailed comparisons between pre-trained causal and bidirectional LMs in discriminative settings. Experiments on LibriSpeech demonstrate that all MWER training schemes are beneficial, giving additional gains upto 8.5% WER. Proposed pooling variants achieve lower latency while retaining most improvements. Finally, our study concludes that bidirectionality is better utilized with discriminative training.

We propose a neural language modeling system based on low-rank adaptation (LoRA) for speech recognition output rescoring. Although pretrained language models (LMs) like BERT have shown superior performance in second-pass rescoring, the high computational cost of scaling up the pretraining stage and adapting the pretrained models to specific domains limit their practical use in rescoring. Here we present a method based on low-rank decomposition to train a rescoring BERT model and adapt it to new domains using only a fraction (0.08%) of the pretrained parameters. These inserted matrices are optimized through a discriminative training objective along with a correlation-based regularization loss. The proposed low-rank adaptation RescoreBERT (LoRB) architecture is evaluated on LibriSpeech and internal datasets with decreased training times by factors between 5.4 and 3.6.

Attention-based contextual biasing approaches have shown significant improvements in the recognition of generic and/or personal rare-words in End-to-End Automatic Speech Recognition (E2E ASR) systems like neural transducers. These approaches employ crossattention to bias the model towards specific contextual entities injected as bias-phrases to the model. Prior approaches typically relied on subword encoders for encoding the bias phrases. However, subword tokenizations are coarse and fail to capture granular pronunciation information which is crucial for biasing based on acoustic similarity. In this work, we propose to use lightweight character representations to encode fine-grained pronunciation features to improve contextual biasing guided by acoustic similarity between the audio and the contextual entities (termed acoustic biasing). We further integrate pretrained neural language model (NLM) based encoders to encode the utterance's semantic context along with contextual entities to perform biasing informed by the utterance’s semantic context (termed semantic biasing). Experiments using a Conformer Transducer model on the Librispeech dataset show a 4.62% - 9.26% relative WER improvement on different biasing list sizes over the baseline contextual model when incorporating our proposed acoustic and semantic biasing approach. On a large-scale in-house dataset, we observe 7.91% relative WER improvement compared to our baseline model. On tail utterances, the improvements are even more pronounced with 36.80% and 23.40% relative WER improvements on Librispeech rare words and an in-house testset respectively.

End-to-End (E2E) automatic speech recognition (ASR) systems used in voice assistants often have difficulties recognizing infrequent words personalized to the user, such as names and places. Rare words often have non-trivial pronunciations, and in such cases, human knowledge in the form of a pronunciation lexicon can be useful. We propose a PROnunCiation-aware conTextual adaptER (PROCTER) that dynamically injects lexicon knowledge into an RNN-T model by adding a phonemic embedding along with a textual embedding. The experimental results show that the proposed PROCTER architecture outperforms the baseline RNN-T model by improving the word error rate (WER) by 44% and 57% when measured on personalized entities and personalized rare entities, respectively, while increasing the model size (number of trainable parameters) by only 1%. Furthermore, when evaluated in a zero-shot setting to recognize personalized device names, we observe 7% WER improvement with PROCTER, as compared to only 1% WER improvement with text-only contextual attention.

End-to-end speech recognition models are improved by incorporating external text sources, typically by fusion with an external language model. Such language models have to be retrained whenever the corpus of interest changes. Furthermore, since they store the entire corpus in their parameters, rare words can be challenging to recall. In this work, we propose augmenting a transducer-based ASR model with a retrieval language model, which directly retrieves from an external text corpus plausible completions for a partial ASR hypothesis. These completions are then integrated into subsequent predictions by an adapter, which is trained once, so that the corpus of interest can be switched without incurring the computational overhead of retraining. Our experiments show that the proposed model significantly improves the performance of a transducer baseline on a pair of question-answering datasets. Further, it outperforms shallow fusion on recognition of named entities by about 7% relative; when the two are combined, the relative improvement increases to 13%.

In interactive automatic speech recognition (ASR) systems, low-latency requirements limit the amount of search space that can be explored during decoding, particularly in end-to-end neural ASR.In this paper, we present a novel streaming ASR architecture that outputs a confusion network while maintaining limited latency, as needed for interactive applications.We show that 1-best results of our model are on par with a comparable RNN-T system, while the richer hypothesis set allows secondpass rescoring to achieve 10-20% lower word error rate on the LibriSpeech task.We also show that our model outperforms a strong RNN-T baseline on a far-field voice assistant task.

End-to-end (E2E) automatic speech recognition models like Recurrent Neural Networks Transducer (RNN-T) are becoming a popular choice for streaming ASR applications like voice assistants. While E2E models are very effective at learning representation of the training data they are trained on, their accuracy on unseen domains remains a challenging problem. Additionally, these models require paired audio and text training data, are computationally expensive and are difficult to adapt towards the fast evolving nature of conversational speech. In this work, we explore a contextual biasing approach using likelihood-ratio that leverages text data sources to adapt RNN-T model to new domains and entities. We show that this method is effective in improving rare words recognition, and results in a relative improvement of 10% in 1-best word error rate (WER) and 10% in n-best Oracle1WER (n=8) on multiple out-of-domain datasets without any degradation on a general dataset. We also show that complementing the contextual biasing adaptation with adaptation of a second-pass rescoring model gives additive WER improvements.

Lattices form a compact representation of multiple hypotheses generated from an automatic speech recognition system and have been shown to improve performance of downstream tasks like spoken language understanding and speech translation, compared to using one-best hypothesis. In this work, we look into the effectiveness of lattice cues for rescoring n-best lists in second-pass. We encode lattices with a recurrent network and train an attention encoder-decoder model for n-best rescoring. The rescoring model with attention to lattices achieves 4-5% relative word error rate reduction over first-pass and 6-8% with attention to both lattices and acoustic features. We show that rescoring models with attention to lattices outperform models with attention to n-best hypotheses. We also study different ways to incorporate lattice weights in the lattice encoder and demonstrate their importance for n-best rescoring.

Second-pass rescoring is an important component in automatic speech recognition (ASR) systems that is used to improve the outputs from a first-pass decoder by implementing a lattice rescoring or n-best re-ranking. While pretraining with a masked language model (MLM) objective has received great success in various natural language understanding (NLU) tasks, it has not gained traction as a rescoring model for ASR. Specifically, training a bidirectional model like BERT on a discriminative objective such as minimum WER (MWER) has not been explored. Here we show how to train a BERT-based rescoring model with MWER loss, to incorporate the improvements of a discriminative loss into fine-tuning of deep bidirectional pretrained models for ASR. Specifically, we propose a fusion strategy that incorporates the MLM into the discriminative training process to effectively distill knowledge from a pretrained model. We further propose an alternative discriminative loss. This approach, which we call RescoreBERT, reduces WER by 6.6%/3.4% relative on the LibriSpeech clean/other test sets over a BERT baseline without discriminative objective. We also evaluate our method on an internal dataset from a conversational agent and find that it reduces both latency and WER (by 3 to 8% relative) over an LSTM rescoring model.

Improving end-to-end speech recognition by incorporating external text data has been a longstanding research topic. There has been a recent focus on training E2E ASR models that get the performance benefits of external text data without incurring the extra cost of evaluating an external language model at inference time. In this work, we propose training ASR model jointly with a set of text-to-text auxiliary tasks with which it shares a decoder and parts of the encoder. When we jointly train ASR and masked language model with the 960-hour Librispeech and Opensubtitles data respectively, we observe WER reductions of 16% and 20% on test-other and test-clean respectively over an ASR-only baseline without any extra cost at inference time, and reductions of 6% and 8% compared to a stronger MUTE-L baseline which trains the decoder with the same text data as our model. We achieve further improvements when we train masked language model on Librispeech data or when we use machine translation as the auxiliary task, without significantly sacrificing performance on the task itself.

Automatic Speech Recognition (ASR) systems have found their use in numerous industrial applications in very diverse domains creating a need to adapt to new domains with small memory and deployment overhead. In this work, we introduce domain prompts, a methodology that involves training a small number of domain embedding parameters to prime a Transformer-based Language Model (LM) to a particular domain. Using this domain-adapted LM for rescoring ASR hypotheses can achieve 7-13% WER reduction for a new domain with just 1000 unlabeled textual domain-specific sentences and a handful of additional parameters. Our method can match or even beat the performance of models fully fine-tuned towards a particular domain with only 0.02% of the parameters. Given the parameter efficiency and negligible deployment overhead, our experiments showcase that such a method is an ideal choice for on-the-fly adaptation of LMs used in ASR systems to progressively scale it to new domains.

End-to-end automatic speech recognition (ASR) systems are increasingly popular due to their relative architectural simplicity and competitive performance. However, even though the average accuracy of these systems may be high, the performance on rare content words often lags behind hybrid ASR systems. To address this problem, second-pass rescoring is often applied leveraging upon language modeling (LM). In this paper, we propose a second-pass system with multi-task learning, utilizing semantic targets (such as intent and slot prediction) to improve speech recognition performance. We show that our rescoring model trained with these additional tasks outperforms the baseline rescoring model, trained with only the LM task, by 1.4% on a general test and by 2.6% on a rare word test set in terms of word-error-rate relative (WERR). Our best ASR system with multi-task LM shows 4.6% WERR deduction compared with RNN Transducer only ASR baseline for rare words recognition.

The recognition of personalized content, such as contact names, remains a challenging problem for end-to-end speech recognition systems. In this work, we demonstrate how first- and second-pass rescoring strategies can be leveraged together to improve the recognition of such words. Following previous work, we use a shallow fusion approach to bias towards recognition of personalized content in the first-pass decoding. We show that such an approach can improve personalized content recognition by up to 16% with minimum degradation on the general use case. We describe a fast and scalable algorithm that enables our biasing models to remain at the word-level, while applying the biasing at the subword level. This has the advantage of not requiring the biasing models to be dependent on any subword symbol table. We also describe a novel second-pass de-biasing approach: used in conjunction with a first-pass shallow fusion that optimizes on oracle WER, we can achieve an additional 14% improvement on personalized content recognition, and even improve accuracy for the general use case by up to 2.5%.

As voice assistants become more ubiquitous, they are increasingly expected to support and perform well on a wide variety of use-cases across different domains. We present a domain-aware rescoring framework suitable for achieving domain-adaptation during second-pass rescoring in production settings. In our framework, we fine-tune a domain-general neural language model on several domains, and use an LSTM-based domain classification model to select the appropriate domain-adapted model to use for second-pass rescoring. This domain-aware rescoring improves the word error rate by up to 2.4% and slot word error rate by up to 4.1% on three individual domains – shopping, navigation, and music – compared to domain general rescoring. These improvements are obtained while maintaining accuracy for the general use case.

End-to-end approaches for automatic speech recognition (ASR) benefit from directly modeling the probability of the word sequence given the input audio stream in a single neural network. However, compared to conventional ASR systems, these models typically require more data to achieve comparable results. Well-known model adaptation techniques, to account for domain and style adaptation, are not easily applicable to end-to-end systems. Conventional HMM-based systems, on the other hand, have been optimized for various production environments and use cases. In this work, we propose to combine the benefits of end-to-end approaches with a conventional system using an attention-based discriminative language model that learns to rescore the output of a first-pass ASR system. We show that learning to rescore a list of potential ASR outputs is much simpler than learning to generate the hypothesis. The proposed model results in up to 8% improvement in word error rate even when the amount of training data is a fraction of data used for training the first-pass system.

Statistical language models (LM) play a key role in Automatic Speech Recognition (ASR) systems used by conversational agents. These ASR systems should provide a high accuracy under a variety of speaking styles, domains, vocabulary and argots. In this paper, we present a DNN-based method to adapt the LM to each user-agent interaction based on generalized contextual information, by predicting an optimal, context-dependent set of LM interpolation weights. We show that this framework for contextual adaptation provides accuracy improvements under different possible mixture LM partitions that are relevant for both (1) Goal-oriented conversational agents where it's natural to partition the data by the requested application and for (2) Non-goal oriented conversational agents where the data can be partitioned using topic labels that come from predictions of a topic classifier. We obtain a relative WER improvement of 3% with a 1-pass decoding strategy and 6% in a 2-pass decoding framework, over an unadapted model. We also show up to a 15% relative improvement in recognizing named entities which is of significant value for conversational ASR systems.

Language models (LM) for interactive speech recognition systems are trained on large amounts of data and the model parameters are optimized on past user data. New application intents and interaction types are released for these systems over time, imposing challenges to adapt the LMs since the existing training data is no longer sufficient to model the future user interactions. It is unclear how to adapt LMs to new application intents without degrading the performance on existing applications. In this paper, we propose a solution to (a) estimate n-gram counts directly from the hand-written grammar for training LMs and (b) use constrained optimization to optimize the system parameters for future use cases, while not degrading the performance on past usage. We evaluated our approach on new applications intents for a personal assistant system and find that the adaptation improves the word error rate by up to 15% on new applications even when there is no adaptation data available for an application.

In particular for “low resource” Keyword Search (KWS) and Speech-to-Text (STT) tasks, more untranscribed test data may be available than training data. Several approaches have been proposed to make this data useful during system development, even when initial systems have Word Error Rates (WER) above 70%. In this paper, we present a set of experiments on low-resource languages in telephony speech quality in Assamese, Bengali, Lao, Haitian, Zulu, and Tamil, demonstrating the impact that such techniques can have, in particular learning robust bottle-neck features on the test data. In the case of Tamil, when significantly more test data than training data is available, we integrated semi-supervised training and speaker adaptation on the test data, and achieved significant additional improvements in STT and KWS.

Recent works have shown Neural Network based Language Models (NNLMs) to be an effective modeling technique for Automatic Speech Recognition. Prior works have shown that these models obtain lower perplexity and word error rate (WER) compared to both standard n-gram language models (LMs) and more advanced language models including maximum entropy and random forest LMs. While these results are compelling, prior works were limited to evaluating NNLMs on perplexity and word error rate. Our initial results showed that while NNLMs improved speech recognition accuracy, the improvement in keyword search was negligible. In this paper we propose alternate optimizations of NNLMs for the task of keyword search. We evaluate the performance of the proposed methods for keyword search on the Vietnamese dataset provided in phase one of the BABEL1project and demonstrate that by penalizing low frequency words during NNLM training, keyword search metrics such as actual term weighted value (ATWV) can be improved by up to 9.3% compared to the standard training methods.

For resource rich languages, recent works have shown Neural Network based Language Models (NNLMs) to be an effective modeling technique for Automatic Speech Recognition, out performing standard n-gram language models (LMs). For low resource languages, however, the performance of NNLMs has not been well explored. In this paper, we evaluate the effectiveness of NNLMs for low resource languages and show that NNLMs learn better word probabilities than state-of-theart n-gram models even when the amount of training data is severely limited. We show that interpolated NNLMs obtain a lower WER than standard n-gram models, no mater the amount of training data. Additionally, we observe that with small amounts of data (approx. 100k training tokens), feed-forward NNLMs obtain lower perplexity than recurrent NNLMs, while for the larger data condition (500k-1M training tokens), recurrent NNLMs can obtain lower perplexity than feed-forward models.

For low resource languages, collecting sufficient training data to build acoustic and language models is time consuming and often expensive. But large amounts of text data, such as online newspapers, web forums or online encyclopedias, usually exist for languages that have a large population of native speakers. This text data can be easily collected from the web and then used to both expand the recognizer's vocabulary and improve the language model. One challenge, however, is normalizing and filtering the web data for a specific task. In this paper, we investigate the use of online text resources to improve the performance of speech recognition specifically for the task of keyword spotting. For the five languages provided in the base period of the IARPA BABEL project, we automatically collected text data from the web using only Limited LP resources. We then compared two methods for filtering the web data, one based on perplexity ranking and the other based on out-of-vocabulary (OOV) word detection. By integrating the web text into our systems, we observed significant improvements in keyword spotting accuracy for four out of the five languages. The best approach obtained an improvement in actual term weighted value (ATWV) of 0.0424 compared to a baseline system trained only on LimitedLP resources. On average, ATWV was improved by 0.0243 across five languages.