개인화된 구음장애 음성 인식(DSR) 모델 개발을 목표로, 본 연구에서는 중증 구음장애를 가진 단일 화자의 네덜란드어 연속 낭독 발화 및 자발화에 대해 인간 청취자와 세 가지 최신 범용 자동 음성 인식(ASR) 시스템(Whisper-large-V3, Google Chirp 3, Omnilingual)의 인식 성능을 비교하였다. 결과에 따르면 인간 청취자와 세 가지 범용 ASR 시스템 모두 평균 단어 오류율(WER)이 70%를 초과하여, DSR이 인간과 ASR 시스템 모두에게 매우 어려운 과제임을 보여준다. 구음장애 발화에 대한 미세조정은 WER을 크게 감소시켰다. 전체 WER이 여전히 상당히 높지만(>23%), 개인화된 DSR 모델은 인간 청취자를 능가하였으며, 구음장애 화자의 일상 의사소통 지원에 활용 가능한 수준에 점차 가까워지고 있다. 향후 연구는 자발화에서의 개인화 DSR 성능 향상과 낭독 발화에서의 긴 발화 처리에 집중해야 하며, 특정 음소에 대한 심층적인 분석이 필요하다.
구어 사실 주장은 종종 멀티턴 대화 안에서 등장하며, 주변 대화는 주장만으로는 얻을 수 없는 맥락을 제공할 수 있다. 그러나 대부분의 팩트체킹(fact-checking) 연구는 고립된 텍스트만 평가하여, 대화형 오디오는 연구가 부족한 상태로 남아 있다. 우리는 MAD2를 제안한다. 이는 구어 주장 검증을 위한 합성 멀티턴 오디오 대화(Multi-turn Audio Dialogues) 벤치마크로, 1,000개의 2화자 대화, 1,230개의 문장 수준 검증 가치 후보 어노테이션, 약 10시간 분량의 오디오를 포함한다. 또한 맥락 인식 오디오 인코더와 대화 인식 텍스트 모델의 보정(calibrated) 멀티모달 퓨전(fusion)을 제안한다. 대화 맥락을 추가하면 설정 전반에서 검증 성능이 향상되지만, 향상폭은 시나리오에 따라 달라진다. 과거 맥락만 사용할 때는 종종 지역적 오프라인(local offline) 성능에 근접하여, 미래 맥락을 사용할 수 없을 때 유용함을 시사한다. 퓨전은 전체 대화 맥락에서 텍스트 대비 평균 이점이 가장 크지만, 설정 전반에 걸쳐 텍스트를 일관되거나 유의미하게 능가하지는 못한다. 전체 대화 맥락 하에서, 탐색적 하위군 결과는 텍스트와 퓨전에서 대화 시나리오에 따른 성능 변동이 더 크고, 오디오에서는 spread 스타일에 따른 변동이 더 크다는 것을 보여준다.
Automatic Speech Recognition (ASR) technology has witnessed significant advancements in recent years, revolutionizing human-computer interactions. While major languages have benefited from these developments, lesser-resourced languages like Urdu face unique challenges. This paper provides an extensive exploration of the dynamic landscape of ASR research, focusing particularly on the resource-constrained Urdu language, which is widely spoken across South Asian nations. It outlines current research trends, technological advancements, and potential directions for future studies in Urdu ASR, aiming to pave the way for forthcoming researchers interested in this domain. By leveraging contemporary technologies, analyzing existing datasets, and evaluating effective algorithms and tools, the paper seeks to shed light on the unique challenges and opportunities associated with Urdu language processing and its integration into the broader field of speech research.
Current expressive speech synthesis models are constrained by the limited availability of open-source datasets containing diverse nonverbal vocalizations (NVs). In this work, we introduce NonverbalTTS (NVTTS), a 17-hour open-access dataset annotated with 10 types of NVs (e.g., laughter, coughs) and 8 emotional categories. The dataset is derived from popular sources, VoxCeleb and Expresso, using automated detection followed by human validation. We propose a comprehensive pipeline that integrates automatic speech recognition (ASR), NV tagging, emotion classification, and a fusion algorithm to merge transcriptions from multiple annotators. Fine-tuning open-source text-to-speech (TTS) models on the NVTTS dataset achieves parity with closed-source systems such as CosyVoice2, as measured by both human evaluation and automatic metrics, including speaker similarity and NV fidelity. By releasing NVTTS and its accompanying annotation guidelines, we address a key bottleneck in expressive TTS research. The dataset is available at https://huggingface.co/datasets/deepvk/NonverbalTTS.
Neural audio codecs, leveraging quantization algorithms, have significantly impacted various speech/audio tasks. While high-fidelity reconstruction is paramount for human perception, audio coding for machines (ACoM) prioritizes efficient compression and downstream task performance, disregarding perceptual nuances. This work introduces an efficient ACoM method that can compress and quantize any chosen intermediate feature representation of an already trained speech/audio downstream model. Our approach employs task-specific loss guidance alongside residual vector quantization (RVQ) losses, providing ultra-low bitrates (i.e., less than 200 bps) with a minimal loss of the downstream model performance. The resulting tokenizer is adaptable to various bitrates and model sizes for flexible deployment. Evaluated on automatic speech recognition and audio classification, our method demonstrates its efficacy and potential for broader task and architectural applicability through appropriate regularization.
Real-time Automatic Speech Recognition (ASR) is a fundamental building block for many commercial applications of ML, including live captioning, dictation, meeting transcriptions, and medical scribes. Accuracy and latency are the most important factors when companies select a system to deploy. We present WhisperKit, an optimized on-device inference system for real-time ASR that significantly outperforms leading cloud-based systems. We benchmark against server-side systems that deploy a diverse set of models, including a frontier model (OpenAI gpt-4o-transcribe), a proprietary model (Deepgram nova-3), and an open-source model (Fireworks large-v3-turbo).Our results show that WhisperKit matches the lowest latency at 0.46s while achieving the highest accuracy 2.2% WER. The optimizations behind the WhisperKit system are described in detail in this paper.
Realistic, high-fidelity 3D facial animations are crucial for expressive avatar systems in human-computer interaction and accessibility. Although prior methods show promising quality, their reliance on the mesh domain limits their ability to fully leverage the rapid visual innovations seen in 2D computer vision and graphics. We propose VisualSpeaker, a novel method that bridges this gap using photorealistic differentiable rendering, supervised by visual speech recognition, for improved 3D facial animation. Our contribution is a perceptual lip-reading loss, derived by passing photorealistic 3D Gaussian Splatting avatar renders through a pre-trained Visual Automatic Speech Recognition model during training. Evaluation on the MEAD dataset demonstrates that VisualSpeaker improves both the standard Lip Vertex Error metric by 56.1% and the perceptual quality of the generated animations, while retaining the controllability of mesh-driven animation. This perceptual focus naturally supports accurate mouthings, essential cues that disambiguate similar manual signs in sign language avatars.
With the advent of new sequence models like Mamba and xLSTM, several studies have shown that these models match or outperform state-of-the-art models in single-channel speech enhancement, automatic speech recognition, and self-supervised audio representation learning. However, prior research has demonstrated that sequence models like LSTM and Mamba tend to overfit to the training set. To address this issue, previous works have shown that adding self-attention to LSTMs substantially improves generalization performance for single-channel speech enhancement. Nevertheless, neither the concept of hybrid Mamba and time-frequency attention models nor their generalization performance have been explored for speech enhancement. In this paper, we propose a novel hybrid architecture, MambAttention, which combines Mamba and shared time- and frequency-multi-head attention modules for generalizable single-channel speech enhancement. To train our model, we introduce VoiceBank+Demand Extended (VB-DemandEx), a dataset inspired by VoiceBank+Demand but with more challenging noise types and lower signal-to-noise ratios. Trained on VB-DemandEx, our proposed MambAttention model significantly outperforms existing state-of-the-art LSTM-, xLSTM-, Mamba-, and Conformer-based systems of similar complexity across all reported metrics on two out-of-domain datasets: DNS 2020 and EARS-WHAM_v2, while matching their performance on the in-domain dataset VB-DemandEx. Ablation studies highlight the role of weight sharing between the time- and frequency-multi-head attention modules for generalization performance. Finally, we explore integrating the shared time- and frequency-multi-head attention modules with LSTM and xLSTM, which yields a notable performance improvement on the out-of-domain datasets. However, our MambAttention model remains superior on both out-of-domain datasets across all reported evaluation metrics.
This report details the NTU Speechlab system developed for the Interspeech 2025 Multilingual Conversational Speech and Language Model (MLC-SLM) Challenge (Task I), where we achieved 5th place. We present comprehensive analyses of our multilingual automatic speech recognition system, highlighting key advancements in model architecture, data selection, and training strategies. In particular, language-specific prompts and model averaging techniques were instrumental in boosting system performance across diverse languages. Compared to the initial baseline system, our final model reduced the average Mix Error Rate from 20.2% to 10.6%, representing an absolute improvement of 9.6% (a relative improvement of 48%) on the evaluation set. Our results demonstrate the effectiveness of our approach and offer practical insights for future Speech Large Language Models.
This paper presents SHTNet, a lightweight spherical harmonic transform (SHT) based framework, which is designed to address cross-array generalization challenges in multi-channel automatic speech recognition (ASR) through three key innovations. First, SHT based spatial sound field decomposition converts microphone signals into geometry-invariant spherical harmonic coefficients, isolating signal processing from array geometry. Second, the Spatio-Spectral Attention Fusion Network (SSAFN) combines coordinate-aware spatial modeling, refined self-attention channel combinator, and spectral noise suppression without conventional beamforming. Third, Rand-SHT training enhances robustness through random channel selection and array geometry reconstruction. The system achieves 39.26\% average CER across heterogeneous arrays (e.g., circular, square, and binaural) on datasets including Aishell-4, Alimeeting, and XMOS, with 97.1\% fewer computations than conventional neural beamformers.
Automatic speech recognition (ASR) systems have achieved strong performance on general transcription tasks. However, they continue to struggle with recognizing rare named entities and adapting to domain mismatches. In contrast, large language models (LLMs), trained on massive internet-scale datasets, are often more effective across a wide range of domains. In this work, we propose a reinforcement learning based approach for unsupervised domain adaptation, leveraging unlabeled data to enhance transcription quality, particularly the named entities affected by domain mismatch, through feedback from a LLM. Given contextual information, our framework employs a LLM as the reward model to score the hypotheses from the ASR model. These scores serve as reward signals to fine-tune the ASR model via reinforcement learning. Our method achieves a 21\% improvement on entity word error rate over conventional self-training methods.
Automatic speech recognition (ASR) models rely on high-quality transcribed data for effective training. Generating pseudo-labels for large unlabeled audio datasets often relies on complex pipelines that combine multiple ASR outputs through multi-stage processing, leading to error propagation, information loss and disjoint optimization. We propose a unified multi-ASR prompt-driven framework using postprocessing by either textual or speech-based large language models (LLMs), replacing voting or other arbitration logic for reconciling the ensemble outputs. We perform a comparative study of multiple architectures with and without LLMs, showing significant improvements in transcription accuracy compared to traditional methods. Furthermore, we use the pseudo-labels generated by the various approaches to train semi-supervised ASR models for different datasets, again showing improved performance with textual and speechLLM transcriptions compared to baselines.
Overlapping speech remains a major challenge for automatic speech recognition (ASR) in real-world applications, particularly in broadcast media with dynamic, multi-speaker interactions. We propose a light-weight, target-speaker-based extension to an existing streaming ASR system to enable practical transcription of overlapping speech with minimal computational overhead. Our approach combines a speaker-independent (SI) model for standard operation with a speaker-conditioned (SC) model selectively applied in overlapping scenarios. Overlap detection is achieved using a compact binary classifier trained on frozen SI model output, offering accurate segmentation at negligible cost. The SC model employs Feature-wise Linear Modulation (FiLM) to incorporate speaker embeddings and is trained on synthetically mixed data to transcribe only the target speaker. Our method supports dynamic speaker tracking and reuses existing modules with minimal modifications. Evaluated on a challenging set of Czech television debates with 16% overlap, the system reduced WER on overlapping segments from 68.0% (baseline) to 35.78% while increasing total computational load by only 44%. The proposed system offers an effective and scalable solution for overlap transcription in continuous ASR services.
Grammatical Error Correction (GEC) and feedback play a vital role in supporting second language (L2) learners, educators, and examiners. While written GEC is well-established, spoken GEC (SGEC), aiming to provide feedback based on learners' speech, poses additional challenges due to disfluencies, transcription errors, and the lack of structured input. SGEC systems typically follow a cascaded pipeline consisting of Automatic Speech Recognition (ASR), disfluency detection, and GEC, making them vulnerable to error propagation across modules. This work examines an End-to-End (E2E) framework for SGEC and feedback generation, highlighting challenges and possible solutions when developing these systems. Cascaded, partial-cascaded and E2E architectures are compared, all built on the Whisper foundation model. A challenge for E2E systems is the scarcity of GEC labeled spoken data. To address this, an automatic pseudo-labeling framework is examined, increasing the training data from 77 to over 2500 hours. To improve the accuracy of the SGEC system, additional contextual information, exploiting the ASR output, is investigated. Candidate feedback of their mistakes is an essential step to improving performance. In E2E systems the SGEC output must be compared with an estimate of the fluent transcription to obtain the feedback. To improve the precision of this feedback, a novel reference alignment process is proposed that aims to remove hypothesised edits that results from fluent transcription errors. Finally, these approaches are combined with an edit confidence estimation approach, to exclude low-confidence edits. Experiments on the in-house Linguaskill (LNG) corpora and the publicly available Speak & Improve (S&I) corpus show that the proposed approaches significantly boost E2E SGEC performance.
The recent rise in capabilities of AI-based music generation tools has created an upheaval in the music industry, necessitating the creation of accurate methods to detect such AI-generated content. This can be done using audio-based detectors; however, it has been shown that they struggle to generalize to unseen generators or when the audio is perturbed. Furthermore, recent work used accurate and cleanly formatted lyrics sourced from a lyrics provider database to detect AI-generated music. However, in practice, such perfect lyrics are not available (only the audio is); this leaves a substantial gap in applicability in real-life use cases. In this work, we instead propose solving this gap by transcribing songs using general automatic speech recognition (ASR) models. We do this using several detectors. The results on diverse, multi-genre, and multi-lingual lyrics show generally strong detection performance across languages and genres, particularly for our best-performing model using Whisper large-v2 and LLM2Vec embeddings. In addition, we show that our method is more robust than state-of-the-art audio-based ones when the audio is perturbed in different ways and when evaluated on different music generators. Our code is available at https://github.com/deezer/robust-AI-lyrics-detection.
The ability to dynamically adjust the computational load of neural models during inference in a resource aware manner is crucial for on-device processing scenarios, characterised by limited and time-varying computational resources. Early-exit architectures represent an elegant and effective solution, since they can process the input with a subset of their layers, exiting at intermediate branches (the upmost layers are hence removed from the model). From a different perspective, for automatic speech recognition applications there are memory-efficient neural architectures that apply variable frame rate analysis, through downsampling/upsampling operations in the middle layers, reducing the overall number of operations and improving significantly the performance on well established benchmarks. One example is the Zipformer. However, these architectures lack the modularity necessary to inject early-exit branches. With the aim of improving the performance in early-exit models, we propose introducing parallel layers in the architecture that process downsampled versions of their inputs. % in conjunction with standard processing layers. We show that in this way the speech recognition performance on standard benchmarks significantly improve, at the cost of a small increase in the overall number of model parameters but without affecting the inference time.
Automatic Speech Recognition (ASR) has reached impressive accuracy for high-resource languages, yet its utility in linguistic fieldwork remains limited. Recordings collected in fieldwork contexts present unique challenges, including spontaneous speech, environmental noise, and severely constrained datasets from under-documented languages. In this paper, we benchmark the performance of two fine-tuned multilingual ASR models, MMS and XLS-R, on five typologically diverse low-resource languages with control of training data duration. Our findings show that MMS is best suited when extremely small amounts of training data are available, whereas XLS-R shows parity performance once training data exceed one hour. We provide linguistically grounded analysis for further provide insights towards practical guidelines for field linguists, highlighting reproducible ASR adaptation approaches to mitigate the transcription bottleneck in language documentation.
Whispered speech recognition presents significant challenges for conventional automatic speech recognition systems, particularly when combined with dialect variation. However, utilizing an efficient method to solve this problem using a low-range dataset and processing load is beneficial. This paper proposes a solution using a Mamba-based state-space model and four fine-tuned self-supervised models consisting of Wav2Vec2, WavLM, HuBERT, and Whisper to address the dual challenges of whispered speech and dialect diversity. Based on our knowledge, this represents the best performance reported on the wTIMIT and CHAINS datasets for whispered speech recognition. We trained the models using whispered and normal speech data across Singaporean, US, and Irish dialects. The findings demonstrated that utilizing the proposed Mamba-based model could work as a highly efficient model trained with low amounts of whispered data to simultaneously work on whispered and normal speech recognition. The code for this work is freely available.
With the rapid progress of speech language models (SLMs), discrete speech tokens have emerged as a core interface between speech and text, enabling unified modeling across modalities. Recent speech tokenization approaches aim to isolate semantic information from low-level acoustics to better align with language models. In particular, previous methods use SSL teachers such as HuBERT to extract semantic representations, which are then distilled into a semantic quantizer to suppress acoustic redundancy as well as capture content-related latent structures. However, they still produce speech token sequences significantly longer than their textual counterparts, creating challenges for efficient speech-language modeling. Reducing the frame rate is a natural solution, but standard techniques, such as rigid average pooling across frames, can distort or dilute the semantic structure required for effective LM alignment. To address this, we propose LM-SPT, a speech tokenization method that introduces a novel semantic distillation. Instead of directly matching teacher and student features via pooling, we reconstruct speech solely from semantic tokens and minimize the discrepancy between the encoded representations of the original and reconstructed waveforms, obtained from a frozen automatic speech recognition (ASR) encoder. This indirect yet data-driven supervision enables the tokenizer to learn discrete units that are more semantically aligned with language models. LM-SPT further incorporates architectural improvements to the encoder and decoder for speech tokenization, and supports multiple frame rates, including 25Hz, 12.5Hz, and 6.25Hz. Experimental results show that LM-SPT achieves superior reconstruction fidelity compared to baselines, and that SLMs trained with LM-SPT tokens achieve competitive performances on speech-to-text and consistently outperform baselines on text-to-speech tasks.
Automatic Speech Recognition (ASR) systems struggle with regional dialects due to biased training which favours mainstream varieties. While previous research has identified racial, age, and gender biases in ASR, regional bias remains underexamined. This study investigates ASR performance on Newcastle English, a well-documented regional dialect known to be challenging for ASR. A two-stage analysis was conducted: first, a manual error analysis on a subsample identified key phonological, lexical, and morphosyntactic errors behind ASR misrecognitions; second, a case study focused on the systematic analysis of ASR recognition of the regional pronouns ``yous'' and ``wor''. Results show that ASR errors directly correlate with regional dialectal features, while social factors play a lesser role in ASR mismatches. We advocate for greater dialectal diversity in ASR training data and highlight the value of sociolinguistic analysis in diagnosing and addressing regional biases.