Tag: #asr16 results
Alibaba's Qwen has upgraded its speech capabilities with Fun-ASR-Realtime, a streaming model featuring sub-100ms latency. It supports 30 languages and 16 dialects with accuracy nearing offline models.
Alibaba has upgraded its Fun-ASR-Realtime model, now supporting 30 languages and 16 dialects with high accuracy. The model is available via the Alibaba Cloud Bailian platform and features sub-100ms latency.
NVIDIA introduces a new approach to evaluating clinical Automatic Speech Recognition (ASR) models using Agent Skills and Nemotron Speech. This solution addresses the difficulty of recognizing complex medical terminology, such as drug and procedure names, which often challenge off-the-shelf systems.
Cohere released Transcribe, its first open-source automatic speech recognition (ASR) model with around 2 billion parameters. Positioned as lightweight for consumer GPUs, it targets self-hosting for speech-to-text transcription and audio content analysis. Ideal for enterprises and developers needing on-premise deployments.
Whisper Accent adapts OpenAI Whisper for accented English using per-accent conditioning via Adaptive Layer Norm, keeping encoder/decoder frozen. It supports 20+ accents and achieves 13.4% WER on medium model, outperforming base Whisper. Full training code, eval pipeline, and checkpoints released on GitHub.
NagaTranslate is an initiative to develop translation and speech pipelines for Nagaland's low-resource languages like Nagamese, Ao, and Sema. The project currently utilizes a hybrid approach of commercial LLM APIs, fine-tuned VITS for TTS, and Whisper for ASR.
This research evaluates whether multilingual or English-only encoders are better for streaming ASR. It finds that multilingual initialization provides a significant advantage only in low-data regimes, with the benefit diminishing as target-language data increases.
Hugging Face has launched the FFASR Leaderboard, designed to benchmark Automatic Speech Recognition (ASR) systems in real-world conditions. This initiative aims to provide more accurate performance evaluations beyond standard academic datasets.
Developers are seeking effective strategies to fine-tune Whisper for technical or domain-specific speech, particularly in Spanish. The discussion evaluates methods beyond standard LoRA and QLoRA for better vocabulary recognition.
This study evaluates self-supervised Transformer models like Wav2Vec2.0 and XLS-R for Quranic ASR, achieving significant improvements in Word Error Rates. The research highlights the effectiveness of domain-specific fine-tuning on large-scale recitation datasets.