Audio8 ASR Infinite: open streaming speech-to-text that never stops

The 4B Apache-2.0 model transcribes forever on one GPU, beats Voxtral and Nemotron on average, and ships with a WebSocket runtime — plus CLM-8B for agents.

Nowline OCT 1 4:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Transcribe 24/7 without drift

    Edge0 open-sourced Audio8 ASR Infinite, a 4B streaming model (Apache-2.0) that re-bases RoPE over a rolling 30-second KV cache, so memory and latency stay flat no matter how long it listens — hours of live audio, no restart, no positional drift. It decodes 12.5x per second with a 240–560 ms delay you set from a single checkpoint.

  • It out-transcribes Voxtral and Nemotron on average

    On the published evals it posts 1.75% CER on AISHELL-1 versus 16.8% for Voxtral Realtime and 12.9% for Nemotron, and 3.04% WER on LibriSpeech test.clean. Averaged over four test sets that is 3.62% error against Voxtral's 10.25% and Nemotron's 9.52% — Voxtral still edges it on clean English, but nothing else is close on Chinese.

  • A self-hosted voice loop for a weekend

    Weights are 8.17 GB in bf16 (one consumer GPU) and ship with an adapted vLLM runtime, Docker Compose, and a WebSocket realtime endpoint, plus semantic turn detection that tells a real end-of-turn from a pause. Bolt it to an open TTS and you have speech-in, speech-out with no per-minute transcription bill.

  • Also worth grabbing: CLM-8B scores agent actions fast

    Contrastive Language Models (Stanford/NVIDIA, Apache-2.0) rank candidate actions instead of generating them, up to 9x faster than Jev. Fine-tuned as a verifier, CLM-8B hits 81.6% on DeepSWE and 87.6% on Terminal-Bench 2.1 at 4–6x the speed — a cheap best-of-N checker to drop into an agent loop.