Ordalin
Submit a tool

Reviewed tool profile

Audio8 ASR

Edge0's open speech-to-text models: live 24/7 Chinese/English streaming, plus a 0.1B model for CPUs and iPhones.

Sources checked

Visit website ↗
Pricing
Both models are free to download and run yourself, but their licences differ. Audio8 ASR Infinite is under Apache 2.0, which allows commercial use. Audio8-ASR-0.1B and its ONNX and iOS packages are under Creative Commons Attribution-NonCommercial 4.0, so they cannot be used commercially without other permission. Your cost is your own hardware; the pages checked list no hosted API. Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗
Available as
Infinite, server: an adapted vLLM build started with Docker Compose, which also serves a web demo and a streaming socket you can drive from a terminal. · Infinite, Python: simulated streaming decode in PyTorch with the model's embedded remote code; the bfloat16 weights are 8.17 GB. · 0.1B, CPU: a self-contained ONNX Runtime package (Python 3.10+) with a browser UI, a local HTTP API and a command-line transcriber. · 0.1B, iPhone: a Swift SDK and demo app that transcribe on-device with Core ML on the Apple Neural Engine (iOS 18+ / macOS 15+), built in Xcode. · 0.1B, Python: the Hugging Face Transformers checkpoint, loaded with trust_remote_code. Official source 1 ↗ Official source 1 ↗ Official source 1 ↗ Official source 1 ↗ Official source 1 ↗
Primary group
Voice & Speech

Also consider

See all 6 similar tools ↓

What it does

Audio8 ASR is Edge0's pair of open speech-to-text models. Audio8 ASR Infinite transcribes Chinese and English live as you speak: it emits text every 80, 120 or 160 ms with a delay you set between 240 and 560 ms, and a rolling 30-second memory lets it run 24/7 on an adapted vLLM server. It is a 4B Voxtral audio tower with a Qwen2.5-3B decoder, under Apache 2.0. Audio8-ASR-0.1B is a much smaller model for seven languages that runs on a CPU through ONNX in about 1.1 GB, or on an iPhone in about 200 MB, under a non-commercial licence.

Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗

Features

  • Live transcription with a delay you choose. Infinite decodes up to 12.5 times a second. You pick an audio clock of 80, 120 or 160 ms and a delay from 240 to 560 ms; a longer delay buys accuracy. Edge0 post-trained specific clock and delay pairs and says others may perform worse. Official source 1 ↗
  • Runs 24/7 without drifting. The model's native context is 30 seconds, but a rolling KV cache with RoPE re-basing keeps memory and latency flat, so the vLLM server can transcribe an endless stream. Official source 1 ↗
  • Semantic end-of-turn detection. Infinite ships semantic voice-activity heads that tell a thinking pause or a stutter from the real end of a speaker's turn, which simple silence detection gets wrong. Official source 1 ↗ · Official source 2 ↗
  • Strong Chinese streaming scores. At 480 ms delay Edge0 reports 1.75% and 2.89% character error on AISHELL-1 and AISHELL-4, against 16.8% and 16.5% for Voxtral Mini 4B Realtime. On English LibriSpeech it reports 3.04% / 6.81% word error, behind Voxtral's 2.21% / 5.55%. Official source 1 ↗ · Official source 2 ↗
  • A 0.1B model for edge devices. Audio8-ASR-0.1B has a 0.104B language model (0.324B parameters end to end) on a Qwen3-ASR audio encoder. It covers English, Chinese, Cantonese, French, German, Japanese and Korean and scores a 7.03% mean word error on the seven Open ASR Leaderboard English splits. Official source 1 ↗
  • Hotword boosting. The 0.1B models can be nudged towards names and terms you supply at decode time, without fine-tuning. The ONNX runtime offers int8 (default), int4 and fp32 precision choices. Official source 1 ↗ · Official source 2 ↗

Pricing & free limits

Audio8 ASR Infinite

FreeOpen weights; Apache License 2.0, self-hosted

Download the 8.17 GB checkpoint and run it with the adapted vLLM Docker stack or PyTorch on your own hardware. Commercial use is allowed under Apache 2.0. Official source 1 ↗ · Official source 2 ↗

Audio8-ASR-0.1B (Transformers, ONNX, iOS)

FreeOpen weights; CC BY-NC 4.0, non-commercial

Download the checkpoint, the CPU ONNX package or the iPhone package. The Creative Commons NonCommercial licence rules out commercial use. Official source 1 ↗ · Official source 2 ↗

Billing details

Billing details were not established by the checked official sources.

Free access & limits

There is no hosted service with quotas; what you can transcribe depends on your own hardware. The 0.1B models are free only for non-commercial use. Official source 1 ↗ · Official source 2 ↗

Best for & limitations

Live captions or meeting transcription in Chinese and English that must run continuously with low delay, using Infinite under Apache 2.0. Official source 1 ↗ · Official source 2 ↗

Voice agents that need to know when a speaker has really finished, using Infinite's semantic end-of-turn detection. Official source 1 ↗

Offline, private transcription on a laptop CPU or an iPhone in seven languages for personal or research projects, using the non-commercial 0.1B packages. Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗

  • Infinite is a preview, Chinese and English only. This release is the transcription base; the semantic perception Edge0 plans is still being built, and Infinite handles only Chinese and English. Official source 1 ↗
  • The small model is non-commercial. Audio8-ASR-0.1B and its ONNX and iOS packages are under CC BY-NC 4.0, unlike Infinite's Apache 2.0. Official source 1 ↗ · Official source 2 ↗
  • Setup is for developers. Infinite needs the full merged 8.17 GB checkpoint and a Docker vLLM stack; the iPhone build has to be compiled and signed in Xcode with an Apple Developer account. There is no ready-made app. Official source 1 ↗ · Official source 2 ↗
  • 0.1B examples cap audio at 30 seconds. The default 0.1B examples target short clips and cut audio at 30 seconds, and hotword boosts set too high can bias the transcript. Official source 1 ↗
  • Vendor-run benchmarks. All scores are Edge0's own, at specific settings (Infinite at an 80 ms clock and 480 ms delay; the 0.1B Chinese scores on an internal evaluation). On English LibriSpeech Infinite trails Voxtral Mini 4B Realtime. Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗
Audio8 ASR product screenshot
Website previewFixed desktop viewport · 1440 × 900

Similar tools

Published tools that share its categories.

Audio8 TTSEdge0's small open-source text-to-speech model that clones a voice from one reference recording, in 23 languages.
Voice & SpeechFreeChecked Oct 31
Visit
CluinnAI dictation for any Mac app, plus a call assistant that answers questions mid-call and lists action items after.
Voice & SpeechPaidChecked Oct 31
Visit
TwinMindAI memory app that transcribes your meetings and day, builds a second brain, and gives any AI your context.
ProductivityFree planChecked Oct 22
Visit
Voxtral 3Mistral AI's speech-to-text and text-to-speech model family with a voice agent SDK for real-time voice agents.
Voice & SpeechPaidChecked Oct 21
Visit