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 explore the ability of large language models (LLMs) to act as speech recognition post-processors that perform rescoring and error correction. Our first focus is on instruction prompting to let LLMs perform these task without fine-tuning, for which we evaluate different prompting schemes, both zeroand few-shot in-context learning, and a novel “task activation“ prompting method that combines causal instructions and demonstration to increase its context windows. Next, we show that rescoring only by in-context learning with frozen LLMs achieves results that are competitive with rescoring by domain-tuned LMs, using a pretrained first-pass recognition system and rescoring output on two out-of-domain tasks (ATIS and WSJ). By combining prompting techniques with fine-tuning we achieve error rates below the N-best oracle level, showcasing the generalization power of the LLMs.

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.

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%.

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.

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.

In this work, we aim to enhance the system robustness of end-to-end automatic speech recognition (ASR) against adversarially-noisy speech examples. We focus on a rigorous and empirical "closed-model adversarial robustness" setting (e.g., on-device or cloud applications). The adversarial noise is only generated by closed-model optimization (e.g., evolutionary and zeroth-order estimation) without accessing gradient information of a targeted ASR model directly. We propose an advanced Bayesian neural network (BNN) based adversarial detector, which could model latent distributions against adaptive adversarial perturbation with divergence measurement. We further simulate deployment scenarios of RNN Transducer, Conformer, and wav2vec-2.0 based ASR systems with proposed adversarial detection system. Leveraging the proposed BNN based detection system, we improve detection rate by +2.77 to +5.42% (relative +3.03 to +6.26%) and reduce the word error rate by 5.02 to 7.47% on LibriSpeech datasets compared to the current model enhancement methods against the adversarial speech examples.

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.

As shown in [1, 2], score normalization is of crucial importance for improving the Average Term-Weighted Value (ATWV) measure that is commonly used for evaluating keyword spotting systems. In this paper, we compare three different methods for score normalization within a keyword spotting system that employs phonetic search. We show that a new unsupervised linear fit method results in better-estimated posterior scores, that, when fed into the keyword-specific normalization of [1], result in ATWV gains of 3% on average. Furthermore, when these scores are used as features within a supervised machine learning framework, they result in additional gains of 3.8% on average over the five languages used in the first year of the IARPA-funded project Babel.

We present two techniques that are shown to yield improved Keyword Spotting (KWS) performance when using the ATWV/MTWV performance measures: (i) score normalization, where the scores of different keywords become commensurate with each other and they more closely correspond to the probability of being correct than raw posteriors; and (ii) system combination, where the detections of multiple systems are merged together, and their scores are interpolated with weights which are optimized using MTWV as the maximization criterion. Both score normalization and system combination approaches show that significant gains in ATWV/MTWV can be obtained, sometimes on the order of 8-10 points (absolute), in five different languages. A variant of these methods resulted in the highest performance for the official surprise language evaluation for the IARPA-funded Babel project in April 2013.

We present a Mandarin keyword search system that uses a large vocabulary recognizer to generate consensus networks at various resolutions: word, character, syllable and phone. In order to achieve fast and accurate search, we propose the use of an efficient approximate-match dynamic programming algorithm that finds the best alignment between the target query and the consensus network. Experiments with Mandarin conversational telephone speech show that the approximate-match search improves detection accuracy by more than 10% for rare words that are not present in the recognizer's dictionary (OOV terms). We also found OOV terms to benefit most from system combination, where we observe a roughly 10% improvement relative to the best single system.

We present a semi-supervised language modeling technique to improve search performance on words without training data. Probabilities estimated from automatic transcripts of a large corpus of in-domain audio are added to an existing LM. Requiring only pronunciations and indomain audio, our method achieves 70% of the possible gain without any supervised data or development data. This method is contrasted with techniques using additional resources. If human effort is available, we also present a transcription regime to efficiently close the gap.

Automatic speech recognition (ASR) for low resource languages continues to be a difficult problem. In particular, colloquial dialects of Arabic, Farsi, and Pashto pose significant challenges in pronunciation dictionary creation. Therefore, most state-of-the-art ASR engines rely on the grapheme-as-phoneme approach for creating pronunciation dictionaries in these languages. While the grapheme approach simplifies ASR training, it performs significantly worse than a system trained with a high-quality phonetic dictionary. In this paper, we explore two techniques for bridging the performance gap between the grapheme and the phonetic approaches, without requiring manual pronunciations for all the words in the training data. The first approach is based on learning letter-to-sound rules from a small set of manual pronunciations in Pashto, and the second approach uses a hybrid phoneme/grapheme representation for recognition. Through experimental results on colloquial Pashto, we demonstrate that both techniques perform as well as a full phonetic system while requiring manual pronunciations for only a small fraction of the words in the acoustic training data.

Named entity (NE) extraction is traditionally applied to written text. Some recent works extend the extraction to broadcast speech but most of these approaches simply cascade a speech-to-text (STT) engine and a named entity tagger without any information sharing between the two. In this paper, we extract named entities from conversational speech and explore approaches to couple the STT and NE extraction beyond a simple cascade. We propose a new approach that adapts the STT language model based on the extracted named entities. This steers the STT to focus on the subset of the vocabulary that is most important to NE extraction. We performed a number of experiments on English conversational speech in the Fisher Corpus with different STT recognition speeds. We show that the language model (LM) adaptation approach increases NE extraction recall and improves NE performance by as much as 15% as measured by F-score, by significantly improving the word error rate (WER) of NE words, with minimal impact on the overall WER.

This paper investigates the use of several language model adaptation techniques applied to the task of machine translation from Arabic broadcast speech. Unsupervised and discriminative approaches slightly outperform the traditional perplexity-based optimization technique. Language model adaptation, when used for n-best rescoring, improves machine translation performance by 0.3-0.4 BLEU and reduces translation edit rate (TER) by 0.2-0.5% compared to an unadapted LM.

This paper presents a set of experiments that we conducted in order to optimize the performance of an Arabic/English machine translation system on broadcast news and conversational speech data. Proper integration of speech-to-text (STT) and machine translation (MT) requires special attention to issues such as sentence boundary detection, punctuation, STT accuracy, tokenization, conversion of spoken numbers and dates to written form, optimization of MT decoding weights, and scoring. We discuss these issues, and show that a carefully tuned STT/MT integration can lead to significant translation accuracy improvements compared to simply feeding the regular STT output to a text MT system.

Lack of data is a problem in training language models for conversational speech recognition, particularly for languages other than English. Experiments in English have successfully used Web-based text collection, targeted for a conversational style, to augment small sets of transcribed speech; we look at extending these techniques to Mandarin. In addition, we investigate different techniques for topic adaptation. Experiments in recognizing Mandarin telephone conversations show that the use of filtered Web data leads to a 28% reduction in perplexity and 7% reduction in character error rate, with most of the gain due to the general filtered Web data.

We describe the ICSI-SRI-UW team’s entry in the Spring 2004 NIST Meeting Recognition Evaluation. The system was derived from SRI’s 5xRT Conversational Telephone Speech (CTS) recognizer by adapting CTS acoustic and language models to the Meeting domain, adding noise reduction and delay-sum array processing for far-field recognition, and postprocessing for cross-talk suppression. A modified MAP adaptation procedure was developed to make best use of discriminatively trained (MMIE) prior models. These meeting-specific changes yielded an overall 9% and 22% relative improvement as compared to the original CTS system, and 16% and 29% relative improvement as compared to our 2002 Meeting Evaluation system, for the individual-headset and multiple-distant microphones conditions, respectively.

N-gram language modeling typically requires large quantities of in-domain training data, i.e., data that matches the task in both topic and style. For conversational speech applications, particularly meeting transcription, obtaining large volumes of speech transcripts is often unrealistic; topics change frequently and collecting conversational-style training data is time-consuming and expensive. In particular, new topics introduce new vocabulary items which are not included in existing models. In this work, we use a variety of data sources (reflecting different sizes and styles), combined using mixture n-gram models. We study the impact of the different sources on vocabulary expansion and recognition accuracy, and investigate possible indicators of the usefulness of a data source. For the task of recognizing meeting speech, we obtain a 9% relative reduction in the overall word error rate and a 61% relative reduction in the word error rate for "new" words added to the vocabulary over a baseline language model trained from general conversational speech data.

With the growing popularity of corpus-based methods for concatenative speech synthesis, a large amount of interest has been placed on borrowing techniques from the ASR community. This paper explores the applications of Buried Markov Models (BMM) to speech synthesis. We show that BMMs are more efficient than HMMs as a synthesis model, and focus on using BMM dependencies for computing splicing costs. We also show how the computational complexity of the dynamic search can be significantly reduced by constraining the splicing points with a negligible loss in synthesis quality.

We describe how prosody prediction can be efficiently integrated with the unit selection process in a concatenative speech synthesizer under a weighted finite-state transducer (WFST) architecture. WFSTs representing prosody prediction and unit selection can be composed during synthesis, thus effectively expanding the space of possible prosodic targets. We implemented a symbolic prosody prediction module and a unit selection database as the synthesis components of a travel planning system. Results of perceptual experiments show that by combining the steps of prosody prediction and unit selection we are able to achieve improved naturalness of synthetic speech compared to the sequential implementation.

In this paper we describe how unit selection for concatenative speech synthesis can be implemented efficiently for sub-phonetic units using weighted finite state transducers (WFST). We also introduce splicing costs as a measure to indicate which unit boundaries are particularly good or poor joint points. Splicing costs extend the flexibility offered by the unit selection paradigm. Through a perceptual experiment we demonstrate an improvement in speech quality achieved by using splicing costs during unit selection.

This paper uses an information-based approach to conduct feature types selection for language modeling in a systematic manner.We describe a quantitative analysis of the information gain and the information redundancy for various combinations of feature types inspired by both dependency structure and bigram structure through analyzing an English treebank corpus and taking word prediction as the object.The experiments yield several conclusions on the predictive value of several feature types and feature types combinations for word prediction, which are expected to provide reliable reference for feature type selection in language modeling.