Lost in Transcription, Found in Distribution Shift: Demystifying Hallucination in Speech Foundation Models

This article concerns comparative studies on the Automatic Speech Recognition (ASR) model incorporated with the Large Language Model (LLM) used for medical interviews. The proposed solution is tested on polish language benchmarks and dataset with medical interviews. The latest ASR technologies are based on convolutional neural networks (CNNs), recurrent neural networks (RNNs) and Transformers. Most of them work as end-to-end solutions. The presented approach in the case of the Whisper model shows a two-stage solution with End-To-End ASR and LLM working together in a pipeline. The ASR output is an input for LLM. The LLM is a component by which the output from ASR is corrected and improved. Comparative studies for automatic recognition of the Polish language between modern End-To-End deep learning architectures and the ASR hybrid model were performed. The medical interview tests were performed with two state-of-the-art ASR models: OpenAI Whisper incorporated with LLM and Scribe ElevenLabs. Additionally, the results were compared with five more end-to-end models (QuartzNet, FastConformer, Wav2Vec 2.0 XLSR and ESPnet Model Zoo) on Mozilla Common Voice and VoxPopuli databases. Tests were conducted for clean audio signal, signal with bandwidth limitation, and degraded. The tested models were evaluated on the basis of Word Error Rate (WER) and Character Error Rate (CER). The results show that the Whisper model performs by far the best among the open-source models. ElevenLabs Scribe model, on the other hand, performs best for Polish on both general benchmark and medical data.

Error correction techniques have been used to refine the output sentences from automatic speech recognition (ASR) models and achieve a lower word error rate (WER). Previous works usually adopt end-to-end models and has strong dependency on Pseudo Paired Data and Original Paired Data. But when only pre-training on Pseudo Paired Data, previous models have negative effect on correction. While fine-tuning on Original Paired Data, the source side data must be transcribed by a well-trained ASR model, which takes a lot of time and not universal. In this paper, we propose UCorrect, an unsupervised Detector-Generator-Selector framework for ASR Error Correction. UCorrect has no dependency on the training data mentioned before. The whole procedure is first to detect whether the character is erroneous, then to generate some candidate characters and finally to select the most confident one to replace the error character. Experiments on the public AISHELL-1 dataset and WenetSpeech dataset show the effectiveness of UCorrect for ASR error correction: 1) it achieves significant WER reduction, achieves 6.83% even without fine-tuning and 14.29% after fine-tuning; 2) it outperforms the popular NAR correction models by a large margin with a competitive low latency; and 3) it is an universal method, as it reduces all WERs of the ASR model with different decoding strategies and reduces all WERs of ASR models trained on different scale datasets.

In this paper, the improvements in the recently implemented Kannada speech recognition system is demonstrated in detail. The Kannada automatic speech recognition (ASR) system consists of ASR models which are created by using Kaldi, IVRS call flow and weather and agricultural commodity prices information databases. The task specific speech data used in the recently developed spoken dialogue system had high level of different background noises. The different types of noises present in collected speech data had an adverse effect on the on line and off line speech recognition performances. Therefore, to improve the speech recognition accuracy in Kannada ASR system, a noise reduction algorithm is developed which is a fusion of spectral subtraction with voice activity detection (SS-VAD) and minimum mean square error spectrum power estimator based on zero crossing (MMSE-SPZC) estimator. The noise elimination algorithm is added in the system before the feature extraction part. An alternative ASR models are created using subspace Gaussian mixture models (SGMM) and deep neural network (DNN) modeling techniques. The experimental results show that, the fusion of noise elimination technique and SGMM/DNN based modeling gives a better relative improvement of 7.68% accuracy compared to the recently developed GMM-HMM based ASR system. The least word error rate (WER) acoustic models could be used in spoken dialogue system. The developed spoken query system is tested from Karnataka farmers under uncontrolled environment.

As the output quality of neural networks in the fields of automatic speech recognition (ASR) and text-to-speech (TTS) continues to improve, new opportunities are becoming available to train models in a weakly supervised fashion, thus minimizing the manual effort required to annotate new audio data for supervised training. While weak supervision has recently shown very promising results in the domain of ASR, speech synthesis has not yet been thoroughly investigated regarding this technique despite requiring the equivalent training dataset structure of aligned audio-transcript pairs. In this work, we compare the performance of TTS models trained using a well-curated and manually labeled training dataset to others trained on the same audio data with text labels generated using both grapheme- and phoneme-based ASR models. Phoneme-based approaches seem especially promising, since even for wrongly predicted phonemes, the resulting word is more likely to sound similar to the originally spoken word than for grapheme-based predictions. For evaluation and ranking, we generate synthesized audio outputs from all previously trained models using input texts sourced from a selection of speech recognition datasets covering a wide range of application domains. These synthesized outputs are subsequently fed into multiple state-of-the-art ASR models with their output text predictions being compared to the initial TTS model input texts. This comparison enables an objective assessment of the intelligibility of the audio outputs from all TTS models, by utilizing metrics like word error rate and character error rate. Our results not only show that models trained on data generated with weak supervision achieve comparable quality to models trained on manually labeled datasets, but can outperform the latter, even for small, well-curated speech datasets. These findings suggest that the future creation of labeled datasets for supervised training of TTS models may not require any manual annotation but can be fully automated.

This study explores the feasibility of constructing a small-scale speech recognition system capable of competing with larger, modern automated speech recognition (ASR) systems in both performance and word error rate (WER). Our central hypothesis posits that a compact transformer-based ASR model can yield comparable results, specifically in terms of WER, to traditional ASR models while challenging contemporary ASR systems that boast significantly larger computational sizes. The aim is to extend ASR capabilities to under-resourced languages with limited corpora, catering to scenarios where practitioners face constraints in both data availability and computational resources. The model, comprising a compact convolutional neural network (CNN) and transformer architecture with 2.214 million parameters, challenges the conventional wisdom that large-scale transformer-based ASR systems are essential for achieving high accuracy. In comparison, contemporary ASR systems often deploy over 300 million parameters. Trained on a modest dataset of approximately 3000 h – significantly less than the 50,000 h used in larger systems – the proposed model leverages the Common Voice and LibriSpeech datasets. Evaluation on the LibriSpeech test-clean and test-other datasets produced character error rates (CERs) of 6.40% and 16.73% and WERs of 16.03% and 35.51% respectively. Comparisons with existing architectures showcase the efficiency of our model. A gated recurrent unit (GRU) architecture, albeit achieving lower error rates, incurred a computational cost 24 times larger than our proposed model. Large-scale transformer architectures, while achieving marginally lower WERs (2%–4% on LibriSpeech test-clean), require 200 times more parameters and 53,000 additional hours of training data. Modern large language models are used to improve the WERs, but require large computational resources. To further enhance performance, a small 4-g language model was integrated into our end-to-end ASR model, resulting in improved WERs. The overarching goal of this work is to provide a practical solution for practitioners dealing with limited datasets and computational resources, particularly in the context of under-resourced languages.

Reverberation negatively impacts the performance of automatic speech recognition (ASR). Prior work on quantifying the effect of reverberation has shown that clarity (C50), a parameter that can be estimated from the acoustic impulse response, is correlated with ASR performance. In this paper we propose predicting ASR performance in terms of the word error rate (WER) directly from acoustic parameters via a polynomial, sigmoidal, or neural network fit, as well as blindly from reverberant speech samples using a convolutional neural network (CNN). We carry out experiments on two state-of-the-art ASR models and a large set of acoustic impulse responses (AIRs). The results confirm C50 and C80 to be highly correlated with WER, allowing WER to be predicted with the proposed fitting approaches. The proposed non-intrusive CNN model outperforms C50-based WER prediction, indicating that WER can be estimated blindly, i.e., directly from the reverberant speech samples without knowledge of the acoustic parameters.

State-of-the-art multilingual Automatic Speech Recognition (ASR) models produce systematic errors when applied to low-resource languages like Rajasthani, for which they lack dedicated training data. This article addresses this challenge by introducing a post-ASR correction framework that leverages the complementary error patterns in the outputs (termed as views) from two distinct models: Whisper-large-v3 and MMS-1B-All. We propose a multi-view, 1 character-level sequence-to-sequence (Seq2Seq) model that uses a gated fusion mechanism to dynamically weigh information from the two ASR outputs. On a new benchmark created from the IndicTTS Rajasthani corpus, our gated model achieves a Character Error Rate (CER) of 7.86% and a Word Error Rate (WER) of 29.98%. This outperforms the best single-view baselines (8.01% CER and 30.33% WER), simple multi-view concatenation (8.21% CER and 30.05% WER), as well as Llama-3.2-3B and mBART-50-large, both fine-tuned on Whisper and MMS inputs. It also surpasses powerful Large Language Models (LLMs) like GPT-4o and Gemini 2.5 Pro in a zero-shot setting. This work establishes the first baseline for post-ASR correction in Rajasthani, demonstrating that a compact, specialized model is more effective than general-purpose LLMs for this targeted, low-resource task.

This work presents a dialect-aware and noise-robust Kannada automatic speech recognition (ASR) system that bridges the gap between low-resource linguistic contexts and state-of-the-art deep learning models. We design a two-stage approach: (i) a scratch-built convolutional neural network (CNN)–Transformer hybrid trained on curated Kannada speech data with fast Fourier transform-based noise reduction and (ii) fine-tuning OpenAI’s Whisper-small model on a dialect-diverse corpus. The proposed pipeline integrates adaptive noise suppression, subword tokenization, and beam-search decoding to handle agglutinative morphology, speaker variation, and environmental noise. Extensive experiments were conducted on two datasets: the Few-shot Learning Evaluation of Universal Representations of Speech (FLEURS) multilingual Kannada subset and a curated 150-sample, Custom-collected Kannada speech dataset covering both formal and conversational speech. On the FLEURS test set (≈ 2.5 hours), our fine-tuned Whisper model achieves a word error rate (WER) of 0.15, a character error rate (CER) of 0.255, and a BLEU score of 0.912, representing a 53% relative reduction in WER and 36% reduction in CER compared to the scratch CNN–Transformer baseline. For the customized dataset, the fine-tuned Whisper model achieves a WER of 0.2311 and a CER of 0.0453, outperforming Google Speech-to-Text Application Programming Interface (API) by 16.8% (relative WER reduction) and surpassing the scratch transformer by over 70% in WER. We further evaluate robustness under dialectal variation and noisy recordings, providing detailed error analysis and computational efficiency metrics. To our knowledge, this is the first comprehensive evaluation of Whisper fine-tuning for Kannada, demonstrating its viability for real-time, edge-deployable applications in education, accessibility, and public administration.

Code-switching is the frequent alternation between two or more languages within a single utterance and is a widespread phenomenon among bilingual and multilingual speakers. In India, more than 250 million people are estimated to engage in code-switched communication, especially blending English with Hindi (Hinglish), making it one of the largest bilingual populations globally, making challenging for developing accurate and robust Automatic Speech Recognition (ASR) systems. Existing ASR models, typically trained on monolingual corpus, struggle with code-switched input due to a lack of large, balanced, and representative datasets—particularly for diverse age groups. Recent evaluations have shown that ASR models experience a relative increase in Word Error Rate (WER) of 30–50 % when exposed to code-switched speech compared to monolingual input. To address this resource gap, we introduce a benchmark Hinglish speech corpus, HiACC, to improve ASR performance in resource-constrained settings. While several monolingual Hindi and English corpus exist, publicly available code-switched datasets remain scarce, and none till date include children's speech. Our corpus fills this gap by providing the first code-switched Hinglish speech dataset with recordings from both adults and children. It comprises 3,318 audio segments from adult participants and 1,858 segments from children, covering 5.24 hours of read and spontaneous speech. The transcriptions include detailed annotations and code-switching tags to assist in linguistic and computational analysis. The corpus is publicly available at [https://zenodo.org/records/15551669], offering segmented audio and aligned transcripts for open research. We also present baseline ASR experiments, which show that standard models trained on monolingual data underperform by approximately 42 % WER on our test set, highlighting the complexity of the task. To our knowledge, this is the first publicly available resource on code-switched Hinglish speech encompassing both adult and child speakers, designed to catalyse progress in this challenging yet important area of speech recognition.

This research introduces a comprehensive strategy to enhance the performance of an existing automatic speech recognition (ASR) model, which has been previously documented in published articles. The study sets out to achieve several objectives. Firstly, it concentrates on updating the ASR model by retraining it with new datasets. This involves integrating samples from the latest Common Voice corpus release and data collected independently via the armspeech.com web application. Another key focus lies in optimizing the ASR model for near-real-time processing, intending to improve its speed and efficiency. The proposed adjustments to the model’s architecture aim to balance accuracy and processing speed, which is essential for applications requiring prompt speech recognition. Furthermore, the research explores the integration of Transformer models into the post-processing pipeline to introduce punctuation and capitalization into the ASR output. This step not only enhances the linguistic quality of transcriptions but also improves their readability and usability. In tandem with these advancements, the research presents a systematic approach to gathering, annotating, and storing datasets specifically tailored for punctuation and capitalization tasks. The methodology outlines the acquisition and organization of a dataset conducive to training Transformer models for these linguistic tasks. This comprehensive approach, which encompasses dataset enrichment, architectural modifications, and post-processing enhancements, aims to elevate the ASR model’s accuracy, speed, and linguistic refinement, with a particular focus on addressing the intricacies of the Armenian language. The research contributes valuable insights into the optimization of ASR systems, tackling both language-specific challenges and broader issues related to linguistic post-processing.

In this letter, a spoken dialogue system is demonstrated to disseminate the agricultural commodity price information in Kannada language/dialects. The spoken dialogue system mainly includes interactive voice response system (IVRS), automatic speech recognition (ASR) models and price information database. The IVRS call flow is used for speech data collection from the farmers across various dialect regions of Karnataka. The collected speech data is transcribed and validated using Indic transliteration tool (IT3: UTF-8). The dictionary and corresponding phoneme set is created for Kannada language. The ASR acoustic models are created using the speech recognition toolkit called Kaldi. The 85% and 15% of transcribed and validated speech corpus is used for Kaldi system training and decoding respectively. The agricultural marketing network (AGMARKNET) website is used as the price information database for different agricultural commodities. The postgresql database is used for managing the crawled data. The word error rates (WERs) of ASR models for districts, mandis, commodities and mixed data are discussed and the spoken dialogue system is developed and tested from the real farmers to access the real time agricultural commodity price information in language/dialects of Kannada.

Automatic speech recognition (ASR) is a vital technology that transforms spoken language into written text, facilitating effective accessibility and communication. Despite the ongoing development of deep learning approaches, speech recognition remains a formidable task, especially for languages with limited data resources, such as Tamil. This work presents the development of an ASR system by utilizing the real-time spontaneous Tamil speech data collected from various types of people's communications in public places. The corpus is trained by fine-tuning the pre-trained wav2vec2 XLSR model. This model captures the diverse acoustic features and patterns and even applied to multiple dialects, making it adaptable to real-world speech. The implemented model is evaluated on various noisy environments like markets, hospitals, shops, etc. In terms of various evaluation metrics such as word error rate (WER) and character error rate (CER), the designed model exhibits an optimal performance by achieving a lower error rate when compared to the baseline ASR models.

In VoIP applications, such as Interactive Voice Response and VoIP-phone conversation transcription, speech signals are degraded not only by environmental noise but also by transmission network quality, and distortions induced by encoding and decoding algorithms. Therefore, there is a need for automatic speech recognition (ASR) systems to handle integrated noise-network distorted speech. In this study, we present a comparative analysis of a speech-to-text system trained on clean speech against one trained on integrated noise-network distorted speech. Training an ASR model on noise-network distorted speech dataset improves its robustness. Although the performance of an ASR model trained on clean speech depends on noise type, this is not the case when noise is further distorted by network transmission. The model trained on noise-network distorted speech exhibited a 60% improvement rate in the word error rate (WER), word match rate (MER), and word information lost (WIL) over the model trained on clean speech. Furthermore, the ASR model trained with noise-network distorted speech could tolerate a jitter of less than 20% and a packet loss of less than 15%, without a decrease in performance. However, WER, MER, and WIL increased in proportion to the jitter and packet loss as they exceeded 20% and 15%, respectively. Additionally, the model trained on noise-network distorted speech exhibited higher robustness compared to that trained on clean speech. The ASR model trained on noise-network distorted speech can also tolerate signal-to-noise (SNR) values of 5 dB and above, without the loss of performance, independent of noise type.

Ambient clinical documentation technology uses automatic speech recognition (ASR) and natural language processing (NLP) to turn patient-clinician conversations into clinical documentation. It is a promising approach to reducing clinician burden and improving documentation quality. However, the performance of current-generation ASR remains inadequately validated. In this study, we investigated the impact of non-lexical conversational sounds (NLCS) on ASR performance. NLCS, such as Mm-hm and Uh-uh, are commonly used to convey important information in clinical conversations, for example, Mm-hm as a "yes" response from the patient to the clinician question "are you allergic to antibiotics?" In this study, we evaluated 2 contemporary ASR engines, Google Speech-to-Text Clinical Conversation ("Google ASR"), and Amazon Transcribe Medical ("Amazon ASR"), both of which have their language models specifically tailored to clinical conversations. The empirical data used were from 36 primary care encounters. We conducted a series of quantitative and qualitative analyses to examine the word error rate (WER) and the potential impact of misrecognized NLCS on the quality of clinical documentation. Out of a total of 135647 spoken words contained in the evaluation data, 3284 (2.4%) were NLCS. Among these NLCS, 76 (0.06% of total words, 2.3% of all NLCS) were used to convey clinically relevant information. The overall WER, of all spoken words, was 11.8% for Google ASR and 12.8% for Amazon ASR. However, both ASR engines demonstrated poor performance in recognizing NLCS: the WERs across frequently used NLCS were 40.8% (Google) and 57.2% (Amazon), respectively; and among the NLCS that conveyed clinically relevant information, 94.7% and 98.7%, respectively. Current ASR solutions are not capable of properly recognizing NLCS, particularly those that convey clinically relevant information. Although the volume of NLCS in our evaluation data was very small (2.4% of the total corpus; and for NLCS that conveyed clinically relevant information: 0.06%), incorrect recognition of them could result in inaccuracies in clinical documentation and introduce new patient safety risks.

Although over 300M around the world speak Bangla, scant work has been done in improving Bangla voice-to-text transcription due to Bangla being a low-resource language. However, with the introduction of the Bengali Common Voice 9.0 speech dataset, Automatic Speech Recognition (ASR) models can now be significantly improved. With 399hrs of speech recordings, Bengali Common Voice is the largest and most diversified open-source Bengali speech corpus in the world. In this paper, we outperform the State-of-the-Art (SOTA) pretrained Bengali ASR models by finetuning a pretrained wav2vec2 model on the common voice dataset. We also demonstrate how to significantly improve the performance of an ASR model by adding an n-gram language model as a post-processor. Finally, we do some experiments and hyperparameter tuning to generate a robust Bangla ASR model that is better than the existing ASR models.