Tag
This paper presents the first systematic study of federated training for SpeechLLM-based end-to-end ASR systems, evaluating on English and Italian tasks and achieving competitive word error rates with reduced communication costs.
Earnings25 is a 500-hour benchmark for evaluating automatic speech recognition on financial earnings calls, providing aligned transcripts and structured metadata for speaker- and industry-aware evaluation.
This paper presents an engineering study adapting NVIDIA Nemotron 3.5 ASR Streaming 0.6B to Kikuyu, Dholuo, and Kalenjin, achieving 42.97% and 33.98% WER on internal sets for Kikuyu and Dholuo, respectively, through data-centric techniques including corpus auditing, normalization, and streaming evaluation.
COALA is a robust framework for contextual biasing in automatic speech recognition (ASR) that uses a contrastive regularizer and biasing score estimation to improve recognition of domain-specific entities from large biasing lists. Experiments on LibriSpeech show consistent superior performance.
This paper proposes a multimodal framework that jointly improves Automatic Speech Recognition (ASR) and Dialect Identification (DID) for Indian languages, using a Bottleneck Encoder and RoBERTa with a gating mechanism. Evaluated on eight languages with 33 dialects, it achieves 81.63% DID accuracy and reduces CER/WER to 4.65%/17.73%.
Introduces a reinforcement learning with verifiable rewards recipe for data-efficient adaptation of audio-language models to code-switched ASR, achieving significant gains across 10 language pairs with minimal data.
This paper proposes Joint Speech-Text Interleaved Pretraining (JSTIP), a pretraining strategy that constructs word-level and segment-level interleaved speech-text sequences to improve ASR entity accuracy and reduce the modality gap between speech and text, showing competitive performance on domain adaptation and zero-shot speech question answering.
This paper evaluates the Mamba state space model for ASR on seven South African languages, finding it matches Conformer accuracy with fewer resources, and explores multilingual training strategies and low-resource settings.
Presents an open-source ASR system for assessing children's reading in Bambara, including field data collection, benchmark construction, model adaptation, and classroom validation, achieving significant word error rate reduction.
This paper presents a certification-inspired mechanism for automatic speech recognition that uses a dual-gate diagnostic pipeline (Two-Sided Atomic Audit and Rank-Based Tournament) to provide certified robustness and achieve up to a 55% relative reduction in word error rate across diverse architectures.
This paper audits multilingual clinical ASR systems on psychiatric interviews in Indian languages and proposes SamaVaani, a unified debiasing technique to improve performance and fairness across demographic groups.
This paper presents a systematic empirical study of fine-tuning pretrained Transformer models (Wav2Vec2.0, HuBERT, XLS-R) for Quranic Automatic Speech Recognition (ASR), achieving a WER of 0.08 on the EveryAyah subset and reducing training time from 140 to 40 hours, with Wav2Vec2-XLSR-53 providing the best representation.
Introduces MultiClin, a benchmark for evaluating ASR in multiscript clinical settings, showing that script unification improves performance over conventional single-reference metrics.
This paper addresses the problem of spoken language adherence in multimodal LLMs for ASR, proposing a soft prompting approach and novel metric to quantify language violations. It evaluates three mitigation strategies—zero-shot prompting, supervised fine-tuning, and chain-of-thought reasoning—across multiple languages to improve transcription fidelity.
This paper evaluates demographic and accent biases in phoneme-based ASR systems, specifically WhisperIPA and ZIPA, using phoneme error rate and a new Soft PER metric, revealing persistent disparities across languages and groups.
ServiceNow AI releases a benchmark and dataset for evaluating automatic speech recognition (ASR) on code-switched speech across four language pairs (Spanish-English, French-English, Canadian French-English, German-English) in enterprise HR and IT scenarios, finding that current frontier ASR models still struggle with code-switching, leading to higher error rates.
This paper introduces the Semantic Gambit attack, which uses LLM predictions to generate real-time adversarial perturbations for automatic speech recognition systems, achieving a three-fold increase in word error rate over prior methods.
This paper introduces Agentic ASR, an interactive speech recognition framework that uses semantic correction and reasoning-based editing to reduce semantic errors through multi-turn refinement. It also proposes a new sentence-level semantic error rate metric and an interactive simulation system for benchmarking.
SCRIBE is a diagnostic evaluation framework for automatic speech recognition that provides categorical error decomposition for Indic languages, releasing benchmarks and open-weight rich transcription models for Hindi, Malayalam, and Kannada.
FormalASR presents two compact end-to-end models that directly transcribe spoken Chinese into formal written text, achieving significant error reduction and eliminating the need for a separate LLM post-processing stage, enabling lightweight on-device deployment.