AI Briefing#asr
The latest AI and developer news about #asr, with the original source and a short summary.
The latest AI and developer news about #asr, with the original source and a short summary.
The 26.6B parameter model supports seven languages and offers natural-language control over transcription style and context biasing.
V4.1-Flash enables server-side tool calls for search, file input, and safety checks, with standalone APIs available for OCR, ASR, and TTS.
Adapts Qwen3-ASR-1.7B with Longest Stable Prefix learning on 1,500 hours of audio to enable monotonic, real-time transcription.
The architecture comprises nine layers including ASR, turn-taking, LLM orchestration, and telephony, with Eleven v4 Turbo leading the Artificial Analysis Provider Voice Arena Leaderboard.
The guide identifies audio input, speaker dynamics, and lexical issues as the primary causes of transcription errors and provides specific fixes for each.
NVIDIA has released Nemotron 3.5 Transcribe, which supports 40 languages, as open source.
The ~70 MB model uses 200+ trained heads on a shared encoder to output transcripts, linked names, and metadata every 100 ms.
The open-source desktop app supports real-time voice conversations, TTS, and ASR using models like Qwen3-TTS and Parakeet TDT v3 without requiring Python.
New 2B model ranks #2 on Open ASR Leaderboard while updated Turbo CTC achieves 19,870 RTFx.
The model achieves 11.43% WER on LibriSpeech test-clean with just 1.06 MB of INT8 weights.
The model achieves 250 ms median endpoint latency and integrates with Gemma 4 E4B and VoxCPM2 for voice agents.
Scribe outputs structured token arrays with word, spacing, and audio_event types in a single API call, supporting up to 5 independent channels.
The new method improves accuracy by 20.8% in-domain and 10.0% cross-domain over prior methods by coupling online teachers with server-side update stabilization.
NVIDIA has released Nemotron 3 Diarization, a 100M parameter open-weight model that ranked first on the Voice Arena Diarization-Bench.
ElevenLabs has officially launched Scribe v2 Medical, optimized for clinical audio, significantly improving medical ASR accuracy.
ElevenLabs released Scribe v2 Medical, a speech recognition model specialized for medical audio, improving clinical audio WER by 18%.
Wispr Advanced Interfaces Lab released 'Canto', a voice model that achieves low Word Error Rate (WER) in real-time transcription environments by applying GRPO reinforcement learning.
Orukeet, a 25-language ASR model based on NVIDIA Parakeet TDT, has been released, improving average WER by 10.6%.
Microsoft has released MAI-Transcribe-2, which supports 60 languages and transcribes one hour of audio in just 10 seconds.
VoiceStudio, an open-source tool for performing voice cloning, video dubbing, and transcription locally, has been released.
Meta released the real-time streaming ASR model 'Muse Voice Transcribe', achieving a top benchmark ranking with a 3.1% WER.
Hugging Face has released Monsoon, a speech recognition evaluation dataset for Hindi and Indian English, to support Global South languages.
The local audio AI framework audio.cpp 0.7 has been released, supporting 62 model families and an Arena UI for comparison.
A study quantifies the phenomenon of ASR models inflating scores by learning benchmark patterns rather than actual speech.