AI Briefing#speech-recognition
The latest AI and developer news about #speech-recognition, with the original source and a short summary.
The latest AI and developer news about #speech-recognition, 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.
The Apache-2.0 licensed model family offers three builds, with the Flash variant achieving 0.31 s first audio latency and supporting 40 concurrent users per 80 GB GPU.
Adapts Qwen3-ASR-1.7B with Longest Stable Prefix learning on 1,500 hours of audio to enable monotonic, real-time transcription.
New private splits from Liva AI include 1.34 hours of Egyptian and 1.55 hours of Nigerian conversational speech to improve evaluation reliability.
The model features a hybrid mixture-of-experts architecture activating approximately 4.3B parameters and supports 40 languages.
The larger balbal-v2-manas model, a CPU-optimized Turkish ASR, ranks second on Mu-Bench TR behind Google Chirp 3, outperforming Gemini 3.5 Transcribe and ElevenLabs Scribe v2.
AnySpeechRecognizer, a Swift package that reimplements the Apple Speech API to enable the use of open models, and AnyDecisionModel for decision tasks have been released.
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 model uses a Gemma 4 backbone and Whisper encoder to handle speech, reasoning, and tool calls, outperforming other open-source systems on voice benchmarks.
The patch release updates the Granite Speech ASR model for 10 languages and improves the Decision adapter without recomputing the base model.
The packages wrap Apple's new SpeechAnalyzer API with NVIDIA and Alibaba models, achieving 1.55% WER on LibriSpeech test-clean, and enable typed decision outputs via six backends including MLX and llama.cpp.
Open-weight model supports 40 languages and speaker diarization, with a 30B parameter architecture using 4B active parameters per token.
New 2B model ranks #2 on Open ASR Leaderboard while updated Turbo CTC achieves 19,870 RTFx.
The compact Mixture-of-Experts model supports speech recognition, audio reasoning, and optional synthesized speech responses in Chinese and English.
The model activates 6.5B parameters per inference step and supports speech recognition, audio reasoning, and synthesized speech responses under an MIT license.
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.
The method achieves a mean F1 score of 0.798 across 14 language pairs in CS-FLEURS using the Whisper-small encoder.
Voice Activity Detection (VAD) classifies audio frames as speech or silence in real-time, enabling efficient transcription, telephony, and voice agent interactions.
Neuphonic has open sourced NeuDecide, a 43 MB audio-to-tool model that achieves 72.4% tool accuracy on SLURP benchmarks without transcription, running efficiently on a single CPU thread.
The model achieves a 20.92% average Word Error Rate on Arabic benchmarks, outperforming the previous best published result by 2.25 percentage points.
The model achieves an 11.1 ms time-to-first-token on an Apple M4 Pro CPU and supports seven languages.
Interventions like auxiliary classifiers reduced phone discrimination error from 11.6% to 10.4% in English/French HuBERT models.