Edge0 open-sources Audio8-ASR-Infinite for 24/7 streaming speech recognition

Audio8-ASR-Infinite from Edge0 is an open-source, native streaming speech recognition model built specifically for nonstop bilingual transcription in Chinese and English.

Instead of choking on long audio streams, it keeps both transcription delay (240 to 560 ms) and GPU memory consumption completely flat across 24/7 continuous runs.

How it stays fast without running out of VRAM

Most streaming ASR models suffer from memory bloat or timestamp drifting when listening to indefinite audio. Audio8 fixes this with a rolling KV cache operating on a 30-second context window, powered by exact RoPE re-basing.

It features a configurable audio clock set to 80, 120, or 160 ms. At the fastest 80 ms step, the architecture decodes 12.5 times per second, letting developers dial in the exact balance between latency and compute cost.

It also integrates semantic Voice Activity Detection (VAD). Unlike standard acoustic energy gates, its semantic heads evaluate context over horizons from 0.5 to 3.0 seconds to distinguish between normal pauses, stuttering, and genuine turn completion.

Architecture and benchmark numbers

Under the hood, Audio8 merges the Voxtral realtime architecture with DSM-style streaming. The causal audio tower originates from Voxtral Realtime 4B, while the text generation side is driven by Qwen2.5-3B-Instruct.

The entire model sits at around 8.17 GB in bfloat16 format, complete with a 151,936-token vocabulary and dedicated semantic VAD weights.

At an 80 ms audio clock and 480 ms latency, it clearly outperforms competitors in Chinese transcription. It scored a 1.750% Character Error Rate (CER) on aishell1 and 2.893% on aishell4, compared to over 16% on Voxtral-Mini-4B and over 12% on Nemotron-3.5-ASR. On English Librispeech test.clean, it clocked a 3.042% Word Error Rate (WER), finishing with an overall benchmark average error of 3.623%.

How to deploy it

The full checkpoint is live on Hugging Face. You can load it directly via the standard Transformers pipeline helper by enabling trust_remote_code.

For persistent 24/7 production use, Edge0 provides an adapted vLLM build orchestrated through Docker Compose. Spinning it up launches a local web demo and a WebSocket streaming endpoint listening on port 18191.

This release represents the preview ASR base. The roadmap indicates that frame-level semantic perception tasks will be stacked onto the same acoustic grid in future updates.

Why it matters

If you are building live broadcast captioning, voice agent frontends, or meeting monitors, Audio8 solves the painful engineering overhead of chunking audio and clearing caches. Getting sub-500ms latency without drift makes real-time agent interruptions much more natural.

Pairing a Voxtral audio encoder with Qwen2.5-3B strikes a very smart balance for bilingual Chinese-English accuracy. You will need an enterprise-grade or mid-range GPU to host the 8GB footprint, but for production streaming, it is one of the cleanest self-hosted options available right now.