Breaking

Edge0 Releases Audio8 ASR Infinite for Always-On Streaming Speech Recognition

September 24, 2026 at 21:16 EDT

Foundation Models
Open Source

Edge0 has released Audio8 ASR Infinite, an Apache 2.0 streaming automatic speech recognition model aimed at continuous, low-latency transcription in Chinese and English. The release claims always-on operation through a rolling KV cache, with published benchmarks showing a strong lead on Chinese test sets but weaker English results than Voxtral on LibriSpeech.

Sep. 25, 2026 / Edge0

Always-on ASR goes open-weight

Audio8 ASR Infinite is built for continuous Chinese and English transcription, combining low-latency streaming with a rolling cache designed to avoid drift over long sessions.

80 ms

fastest audio clock

12.5/s

decoding decisions

24/7

target operation

8.17 GB

main weights file

AISHELL-1 error-rate comparison

Character error rate; lower is better. Column height is proportional to the reported value.

1.750

Audio8

12.927

Nemotron

16.795

Voxtral

Why it can keep listening

STEP 1

Shared audio clock

one token per clock step

STEP 2

Rolling KV cache

constant memory and latency

STEP 3

Semantic turn cues

pause, stutter or end of turn

What looks strong

Self-hosted continuous transcription
Large Chinese benchmark margin
Docker and vLLM deployment path

What still needs proof

Preview release, not a full speech platform
English WER trails Voxtral
Production setup requires local serving work

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.