Edge0 released Audio8 ASR Infinite, a native streaming speech recognition model. It supports selectable audio clocks of 80, 120 or 160 ms. Users can set transcription delay between 240 and 560 ms. A rolling KV cache enables unlimited-length transcription without drift. It handles Chinese and English. Semantic VAD distinguishes pauses and stuttering from speech ends. The model inherits Voxtral’s realtime audio architecture. It uses a 30-second context window by default. Deployment is via Docker Compose with vLLM for 24/7 operation. This is a preview release focused on transcription. Realtime semantic perception is planned next. Benchmarks show lower error rates than comparable models on key datasets.
What changed
Streaming ASR models typically drifted in accuracy or latency over long audio sessions.
What it unlocks
Transcribing unlimited-length audio 24/7 with constant memory and latency.
- 4B parameters
- 8.17 GB model size
- 30-second rolling KV cache
- 1.75% CER on aishell1/test
What you need to act on it
- vLLM or Torch for inference
- selectable audio clock (80/120/160 ms)
- configurable transcription delay (240–560 ms)
Sources