AirCaps launches A5S Async v1 speech model with 41% fewer errors, open-sources benchmark
dair_ai · x · 2026-11-10
AirCaps released A5S Async v1, a speaker-attributed speech recognition model claiming #1 on VoiceArena ahead of ElevenLabs, AssemblyAI, Deepgram, NVIDIA and Meta—the next-best model makes 41% more errors across 5 public benchmarks. Unusually, it open-sourced the benchmark harness code, dataset recipes, raw API responses and scores, a transparency practice dair-ai says more launches should copy. cpWER on Diarization-Bench v1: A5S 23.9% vs NVIDIA Nemotron 3.5 Transcribe 32.4%.
More from Multimodal
- Ming-3.0-TTS ships in Comfy-Org repo with no workflow and custom nodes won't load it — Slight_Tone_2188 · 2026-11-10
- Runway launches Sightline: generate 360 panoramas and shoot from any consistent angle — runwayml · 2026-11-10
- Lightricks open-sources LTX-2.5 world model: 22B weights, native 4K HDR, runs locally — dr_cintas · 2026-11-10
- ComfyUI realism struggles: Flux over-censors, Qwen lacks detail, seeking workflows — Accurate-Focus5276 · 2026-11-10
- Krea teases its biggest-ever team update, calling new features creative teammates — TitusTeatus · 2026-11-10
- Creator makes 5-minute sci-fi film 'I WAS HERE' using ChatGPT's new ImageGen — VraserX · 2026-11-10