We introduce a large language model (LLM) capable of processing speech inputs and show that tuning it further with reinforcement learning on human preference (RLHF) enables it to adapt better to disordered speech than traditional fine-tuning. Our method replaces low-frequency text tokens in an LLM’s vocabulary with audio tokens and enables the model to recognize speech by fine-tuning it on speech with transcripts. We then use RL with rewards based on syntactic and semantic accuracy measures generalizing the LLM further to recognize disordered speech. While the resulting LLM does not outperform existing systems for speech recognition we find that tuning with reinforcement learning using custom rewards leads to substantially better performance than supervised fine-tuning of the language model, specifically when adapting to speech in a different setting. This presents a compelling alternative tuning strategy for speech recognition using large language models.

This study investigates the impact of integrating a dataset of disordered speech recordings (~1,000 hours) into the fine-tuning of a near state-of-the-art ASR baseline system. Contrary to what one might expect, despite the data being less than 1% of the training data of the ASR system, we find a considerable improvement in disordered speech recognition accuracy. Specifically, we observe a 33% improvement on prompted speech, and a 26% improvement on a newly gathered spontaneous, conversational dataset of disordered speech. Importantly, there is no significant performance decline on standard speech recognition benchmarks. Further, we observe that the proposed tuning strategy helps close the gap between the baseline system and personalized models by 64% highlighting the significant progress as well as the room for improvement. Given the substantial benefits of our findings, this experiment suggests that from a fairness perspective, incorporating a small fraction of high quality disordered speech data in a training recipe is an easy step that could be done to make speech technology more accessible for users with speech disabilities.

We explore a strategy to handle controversial topics in LLM-based chatbots based on Wikipedia’s Neutral Point of View (NPOV) principle: acknowledge the absence of a single true answer and surface multiple perspectives. We frame this as retrieval augmented generation, where perspectives are retrieved from a knowledge base and the LLM is tasked with generating a fluent and faithful response from the given perspectives. As a starting point, we use a deterministic retrieval system and then focus on common LLM failure modes that arise during this approach to text generation, namely hallucination and coverage errors. We propose and evaluate three methods to detect such errors based on (1) word-overlap, (2) salience, and (3) LLM-based classifiers. Our results demonstrate that LLM-based classifiers, even when trained only on synthetic errors, achieve high error detection performance, with ROC AUC scores of 95.3% for hallucination and 90.5% for coverage error detection on unambiguous error cases. We show that when no training data is available, our other methods still yield good results on hallucination (84.0%) and coverage error (85.2%) detection.

Automatic Speech Recognition (ASR) systems, despite significant advances in recent years, still have much room for improvement particularly in the recognition of disordered speech. Even so, erroneous transcripts from ASR models can help people with disordered speech be better understood, especially if the transcription doesn’t significantly change the intended meaning. Evaluating the efficacy of ASR for this use case requires a methodology for measuring the impact of transcription errors on the intended meaning and comprehensibility. Human evaluation is the gold standard for this, but it can be laborious, slow, and expensive. In this work, we tune and evaluate large language models for this task and find them to be a much better proxy for human evaluators than other metrics commonly used. We further present a case-study using the presented approach to assess the quality of personalized ASR models to make model deployment decisions and correctly set user expectations for model quality as part of our trusted tester program.

Project Euphonia, a Google initiative, is dedicated to improving automatic speech recognition (ASR) of disordered speech. A central objective of the project is to create a large, high-quality, and diverse speech corpus. This report describes the project's latest advancements in data collection and annotation methodologies, such as expanding speaker diversity in the database, adding human-reviewed transcript corrections and audio quality tags to 350K (of the 1.2M total) audio recordings, and amassing a comprehensive set of metadata (including more than 40 speech characteristic labels) for over 75% of the speakers in the database. We report on the impact of transcript corrections on our machine-learning (ML) research, inter-rater variability of assessments of disordered speech patterns, and our rationale for gathering speech metadata. We also consider the limitations of using automated off-the-shelf annotation methods for assessing disordered speech.

Although personalized automatic speech recognition (ASR) models have recently been improved to recognize even severely impaired speech, model performance may degrade over time for persons with degenerating speech. The aims of this study were to (1) analyze the change of performance of ASR over time in individuals with degrading speech, and (2) explore mitigation strategies to optimize recognition throughout disease progression. Speech was recorded by four individuals with degrading speech due to amyotrophic lateral sclerosis (ALS). Word error rates (WER) across recording sessions were computed for three ASR models: Unadapted Speaker Independent (U-SI), Adapted Speaker Independent (A-SI), and Adapted Speaker Dependent (A-SD or personalized). The performance of all models degraded significantly over time as speech became more impaired, but the A-SD model improved markedly when updated with recordings from the severe stages of speech progression. Recording additional utterances early in the disease before significant speech degradation did not improve the performance of A-SD models. This emphasizes the importance of continuous recording (and model retraining) when providing personalized models for individuals with progressive speech impairments.

This study investigates the performance of personalized automatic speech recognition (ASR) for recognizing disordered speech using small amounts of per-speaker adaptation data. We trained personalized models for 195 individuals with different types and severities of speech impairment with training sets ranging in size from <1 minute to 18-20 minutes of speech data. Word error rate (WER) thresholds were selected to determine Success Percentage (the percentage of personalized models reaching the target WER) in different application scenarios. For the home automation scenario, 79% of speakers reached the target WER with 18-20 minutes of speech; but even with only 3-4 minutes of speech, 63% of speakers reached the target WER. Further evaluation found similar improvement on test sets with conversational and out-of-domain, unprompted phrases. Our results demonstrate that with only a few minutes of recordings, individuals with disordered speech could benefit from personalized ASR.

Shanqing Cai, Subhashini Venugopalan, Katrin Tomanek, Ajit Narayanan, Meredith Morris, Michael Brenner. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Automatic Speech Recognition (ASR) systems are often optimized to work best for speakers with canonical speech patterns.Unfortunately, these systems perform poorly when tested on atypical speech and heavily accented speech.It has previously been shown that personalization through model fine-tuning substantially improves performance.However, maintaining such large models per speaker is costly and difficult to scale.We show that by adding a relatively small number of extra parameters to the encoder layers via socalled residual adapter, we can achieve similar adaptation gains compared to model finetuning, while only updating a tiny fraction (less than 0.5%) of the model parameters.We demonstrate this on two speech adaptation tasks (atypical and accented speech) and for two state-of-the-art ASR architectures.

Automatic classification of disordered speech can provide an objective tool for identifying the presence and severity of speech impairment. Classification approaches can also help identify hard-to-recognize speech samples to teach ASR systems about the variable manifestations of impaired speech. Here, we develop and compare different deep learning techniques to classify the intelligibility of disordered speech on selected phrases. We collected samples from a diverse set of 661 speakers with a variety of self-reported disorders speaking 29 words or phrases, which were rated by speech-language pathologists for their overall intelligibility using a five-point Likert scale. We then evaluated classifiers developed using 3 approaches: (1) a convolutional neural network (CNN) trained for the task, (2) classifiers trained on non-semantic speech representations from CNNs that used an unsupervised objective [1], and (3) classifiers trained on the acoustic (encoder) embeddings from an ASR system trained on typical speech [2]. We found that the ASR encoder's embeddings considerably outperform the other two on detecting and classifying disordered speech. Further analysis shows that the ASR embeddings cluster speech by the spoken phrase, while the non-semantic embeddings cluster speech by speaker. Also, longer phrases are more indicative of intelligibility deficits than single words.

In lots of natural language processing tasks, the classes to be dealt with often occur heavily imbalanced in the underlying data set and classifiers trained on such skewed data tend to exhibit poor performance for low-frequency classes. We introduce and compare different approaches to reduce class imbalance by design within the context of active learning (AL). Our goal is to compile more balanced data sets up front during annotation time when AL is used as a strategy to acquire training material. We situate our approach in the context of named entity recognition. Our experiments reveal that we can indeed reduce class imbalance and increase the performance of classifiers on minority classes while preserving a good overall performance in terms of macro F-score.

MOTIVATION: The recognition and normalization of textual mentions of gene and protein names is both particularly important and challenging. Its importance lies in the fact that they constitute the crucial conceptual entities in biomedicine. Their recognition and normalization remains a challenging task because of widespread gene name ambiguities within species, across species, with common English words and with medical sublanguage terms. RESULTS: We present GeNo, a highly competitive system for gene name normalization, which obtains an F-measure performance of 86.4% (precision: 87.8%, recall: 85.0%) on the BioCreAtIvE-II test set, thus being on a par with the best system on that task. Our system tackles the complex gene normalization problem by employing a carefully crafted suite of symbolic and statistical methods, and by fully relying on publicly available software and data resources, including extensive background knowledge based on semantic profiling. A major goal of our work is to present GeNo's architecture in a lucid and perspicuous way to pave the way to full reproducibility of our results. AVAILABILITY: GeNo, including its underlying resources, will be available from www.julielab.de. It is also currently deployed in the Semedico search engine at www.semedico.org.