Edge0 Releases Audio8-ASR-0.1B to Run Speech Recognition on Any Phone
A tiny 0.1B-parameter multilingual speech recognition model runs entirely on-device with a 200MB memory footprint on iPhones.
- Audio8-ASR-0.1B is a 324M-param multilingual ASR model with a 0.1B LM decoder, released on Hugging Face.
- Supports Chinese, English, French, German, Japanese, Korean, and Cantonese with hotword boosting at decode time.
- Achieves 7.03% mean WER on the Open ASR Leaderboard's seven English splits at 741x realtime on H200.
- iOS Neural Engine variant runs with roughly 200MB peak memory; ONNX Runtime variant needs about 1.1GB.
- Built on a Qwen3-ASR encoder plus a Qwen-style 8-layer causal LM decoder, trained with on-policy distillation.
- Released under CC-BY-NC-4.0 as part of the broader Audio8 ASR and TTS model family.
Audio8-ASR-0.1B targets multilingual transcription on edge devices
Edge0 has released Audio8-ASR-0.1B, a compact autoregressive speech-recognition model for phones, laptops, and other memory-constrained hardware. The release pairs a base Transformers checkpoint with ONNX Runtime, INT8, and iOS Neural Engine deployment options.
The model recognizes Chinese, English, French, German, Japanese, Korean, and Cantonese. Its decoder produces a transcript one token at a time from encoded audio and previously generated tokens. Local execution can reduce network latency and keep raw audio off remote servers when the surrounding application also avoids uploads and telemetry.
A 103.5M decoder inside a 324M system
The 0.1B label identifies the language-model component, which contains about 103.5 million parameters. The complete speech-recognition pipeline contains about 324 million unique parameters after the audio encoder, adapter, and projector are included.
| Component | Implementation |
|---|---|
| Audio encoder | Backbone derived from Qwen3-ASR-0.6B |
| Adapter and projector | MLP components trained for Audio8-ASR |
| Language decoder | Eight-layer Qwen-style causal model based on Ref-Pretrain-Qwen-104M |
| Decoder size | Approximately 103.5 million parameters |
| Full pipeline size | Approximately 324 million unique parameters |
| Audio rate | 16 kHz |
The architecture reuses a larger model’s audio representation while keeping the text decoder small. That split explains how the release can carry a 0.1B name while requiring more memory than a standalone 100-million-parameter language model.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
