Open-source ASR built on a Qwen3-4B decoder posts 3.54 average WER across 5 languages

rohanpaul_ai · x · 2026-10-16

Hojo-ASR-Multi-V1, an open multilingual ASR model (Apache 2.0, on Hugging Face), posted a 3.54 average WER across 5 European languages — 7th on the leaderboard, behind Azure Speech (2.61) and ElevenLabs Scribe v2 (2.65).

The architectural twist: an Encoder-Adapter-LLM setup where the acoustic encoder and adapter feed features into a Qwen3-4B LLM that produces the final transcript. That gives decoding much richer language context for messy, informal speech, but also raises hallucination and over-correction risks on weak audio. The model was trained with multi-stage modular training plus reinforcement learning and claims support for German, French, Spanish, Portuguese, Japanese, Italian, Arabic, Korean and Russian.

Caveats from @rohanpaulai:

3.54 isn't universal — the leaderboard covers only 5 European languages; the claimed Japanese/Arabic/Korean/Russian support isn't reflected in this result, and averages can hide large per-dataset gaps.
Local inference ≠ edge-ready — for glasses, earbuds or recorders you still need on-device measurements of memory, GPU needs, throughput, streaming behavior and latency; better tests should mix accents, noise, far-field audio, code-switching, and publish ground truth next to raw transcripts.

Installable via pip install hojo-asr; accepts WAV paths, byte streams and SCP inputs.

Original post →

Open-source ASR built on a Qwen3-4B decoder posts 3.54 average WER across 5 languages

More from Models