Audio8 ASR Infinite: An Open Streaming ASR Model Built to Never Stop Listening
Speech recognition is easy—until you ask it to listen forever.
Today we’re open-sourcing Audio8 ASR Infinite, a streaming ASR model designed for ultra-low latency, unlimited-length audio, and 24/7 transcription without drift.
It also includes built-in semantic turn detection, allowing the model to distinguish between a thinking pause, a stutter, and the actual end of a turn—closer to how a human listener understands conversation.
1. Open source
The preview release includes:
The Audio8 ASR Infinite streaming model
An adapted vLLM runtime
Docker deployment
Local inference examples
The model is built for:
24/7 transcription
Full-duplex, real-time voice interaction
Always-on audio applications
It is low-latency, bilingual in Chinese and English, and designed to keep listening continuously.
2. How does it listen 24/7 without drifting?
In simple terms:
One shared audio-text clock
Audio and text move on the same clock. Every 80 ms, the model emits exactly one token, resulting in 12.5 decisions per second.
Latency is configurable
The transcription delay can be set between 240 and 560 ms. You can trade latency for accuracy without changing the model weights.
A rolling context window
A 30-second rolling window keeps memory bounded. Older context is removed, and position indices rotate back to zero. After 10 hours of operation, the model behaves like it did during the first minute.
Semantic turn detection
Turn-taking is handled semantically rather than through acoustic volume alone. The model can recognize whether someone is still thinking, stuttering, or has actually finished speaking.
3. Benchmark results
We evaluated Audio8 ASR Infinite against two streaming baselines at the same operating point: 480 ms delay and 80 ms clock. Lower is better.
On Chinese speech, the gap is substantial:
AISHELL-1 CER: 1.75 vs. 16.80 for Voxtral-Mini-4B-Realtime
AISHELL-4 CER: 2.89 vs. 16.46
On English speech, the results are closer:
LibriSpeech test-clean WER: 3.04 for Audio8 ASR Infinite vs. 2.21 for Voxtral
We prioritized Chinese during post-training, so English clean-speech performance remains an area for improvement.
Across the four reported test sets, the average score was:
Audio8 ASR Infinite: 3.62
Voxtral-Mini-4B-Realtime: 10.25
Nemotron-3.5-ASR: 9.52
Nemotron-3.5 was measured at 560 ms delay, which gives it a longer latency budget and therefore a more favorable operating point.
Edge0 has open-sourced Audio8 ASR Infinite, a 4B-parameter streaming speech recognition model (Apache 2.0) built on the Voxtral realtime audio architecture with a Qwen2.5-3B-Instruct decoder. It's designed for unlimited-length, low-latency, 24/7 transcription in Chinese and English, using a shared audio-text clock that emits one token every 80ms (12.5 decisions per second), with transcription delay configurable between 240-560ms without retraining.
The preview release includes the streaming model, an adapted vLLM runtime, Docker deployment, and local inference examples. A semantic voice-activity-detection head — designed to distinguish a thinking pause, a stutter, and genuine end-of-turn across multiple time horizons — is included in the checkpoint, though the creators note the fuller frame-level semantic perception capability is still in progress toward a formal release, beyond this preview.
On a benchmark against Voxtral-Mini-4B-Realtime and Nemotron-3.5-ASR-Streaming (a smaller 0.6B model) at matched operating points, Audio8 shows a large advantage on Chinese speech and trails Voxtral slightly on clean English speech — consistent with the team's own note that Chinese was prioritized during post-training.