Edge0/Audio8-ASR-Infinite — new model trending #30 on Hugging Face
Edge0 released Audio8-ASR-Infinite, a bilingual streaming speech model with constant-memory unlimited transcription.
Edge0 released Audio8-ASR-Infinite, a preview streaming automatic-speech-recognition model for Chinese and English. It offers 80, 120, or 160 millisecond audio clocks and uses a rolling KV cache so memory and latency stay constant on unlimited-length audio. The audio tower is initialized from Voxtral Realtime 4B and the decoder from Qwen2.5-3B-Instruct; weights are 8.17 GB under Apache 2.0. At a 480 millisecond delay, Edge0 reports 1.750 CER on AISHELL-1 and 3.042 WER on LibriSpeech test-clean.
- Selectable audio clocks are 80, 120, or 160 milliseconds.
- Rolling KV cache keeps memory and latency constant for long audio.
- Decoder starts from Qwen2.5-3B; audio tower from Voxtral Realtime 4B.
- At 480 ms delay, AISHELL-1 CER is 1.750 versus Voxtral 16.795.
- Preview weights are 8.17 GB and licensed under Apache 2.0.
Full article929 words · extracted from huggingface.co · click to collapse
<div align="center">
# Audio8 ASR Infinite
[](https://huggingface.co/Edge0/Audio8-ASR-Infinite)
[](https://github.com/Edge0-AI/Audio8-ASR-Infinite)
[](https://github.com/Edge0-AI/Audio8-ASR-Infinite)
[](https://github.com/Edge0-AI/Audio8-ASR-Infinite/blob/main/LICENSE)
</div>
**Audio8 ASR Infinite** is a native streaming speech recognition model built to be
as responsive as possible. It offers a selectable audio clock (80/120/160 ms) and
a transcription delay (240–560 ms).
With our adapted vLLM build it transcribes unlimited-length audio **24/7** without drifting.
## Highlights
- **Super responsive** — the native streaming architecture decodes 12.5 times per second.
- **Unlimited-length transcription** — a rolling KV Cache keeps both **memory and
latency constant**, even in **24/7 operation**.
- **Selectable streaming clock** — one text token per clock step
(12.5 / 8.3 / 6.25 decisions per second), balancing perception granularity and resource cost.
- **Configurable transcription delay** — set how much delay to trade for accuracy.
- **Semantic VAD** — distinguishes thinking pauses, stuttering and real end of turn, where traditional acoustic VAD usually fails.
- **Bilingual** — Chinese and English.
## See Audio8-ASR-Infinite in action
The checkpoint has a native context of 30 seconds. But with Rolling KV Cache, it can transcribe 24/7 nonstop.
<video controls playsinline width="100%" preload="metadata"
src="https://huggingface.co/Edge0/Audio8-ASR-Infinite/resolve/main/Audio8-Asr-Infinite-Demo.mp4"></video>
## Optimized operation points
The following combinations of frame length and delay are post-trained. Other combinations can be used but performance may not be optimum.
| audio clock | `frame_len` | `streaming_n_left_pad_tokens` | selectable `target_delay_ms` |
| --- | --- | --- | --- |
| 80 ms | 4 | 18 | 240 / 320 / 480 / 560 |
| 120 ms | 6 | 12 | 240 / 480 |
| 160 ms | 8 | 9 | 320 / 480 |