asr-468m-apache-v2 — Apache-2.0 multilingual ASR that beats Audio8-ASR-0.1B

A 7-language (Chinese, English, French, German, Japanese, Korean, Cantonese) speech-to-text model distilled from Qwen3-ASR-0.6B (Apache-2.0), at 467.81M parameters (1.44x Audio8-ASR-0.1B's size) — outperforming Audio8-ASR-0.1B outright while remaining fully Apache-2.0 and commercially usable (Audio8 is CC-BY-NC and cannot be used commercially).

Architecture: Qwen3-ASR-0.6B's frozen 18×896 audio tower (186.38M) + an 8-layer × 1024 decoder with the full 151,936-token vocabulary (no vocab pruning — no vocab_remap.json needed, unlike the size-matched Luigi/asr-324m-apache).

Method: the best independently-trained 6-layer/42,000-vocab specialist had its 6 trained layers re-inserted at their original depth indices into the 8-layer parent (Luigi/asr-468m-apache-base), plus a full-vocabulary graft (the specialist's 42k rows kept, the other ~110k rows filled from the parent's own embedding) — no further training. Full derivation: docs/findings.md ("Stage 4 postscript").

Results (200-clip FLEURS test gate, all-refs; macro mixes CER for zh/ja/ko/yue and WER for en/fr/de)

language	Audio8-ASR-0.1B	this model	delta
French	20.88	16.40	−4.48 ✅
Cantonese	16.16	14.78	−1.38 ✅
Japanese	17.97	16.63	−1.34 ✅
Korean	13.97	11.89	−2.08 ✅
Chinese	11.94	12.08	+0.14
English	8.51	10.71	+2.20
German	17.73	19.35	+1.62
macro	15.31	14.55	−0.76 ✅ beats Audio8

Usage

Full vocabulary — no id remapping needed.

import torch
from qwen_asr.core.transformers_backend.modeling_qwen3_asr import Qwen3ASRForConditionalGeneration
from qwen_asr.core.transformers_backend.processing_qwen3_asr import Qwen3ASRProcessor

proc = Qwen3ASRProcessor.from_pretrained("Luigi/asr-468m-apache-v2")
model = Qwen3ASRForConditionalGeneration.from_pretrained(
    "Luigi/asr-468m-apache-v2", dtype=torch.bfloat16).cuda().eval()

Training data (all commercially usable)

Common Voice 17 (CC0), WenetSpeech4TTS (CC-BY-4.0), Multilingual LibriSpeech (CC-BY-4.0), LibriSpeech (CC-BY-4.0), FLEURS (CC-BY-4.0), AISHELL-1 (Apache-2.0), YouTube-Cantonese / OrcinusOrca (MIT). Attribution for CC-BY sources: trained in part on WenetSpeech4TTS, Multilingual LibriSpeech, LibriSpeech, and FLEURS, each licensed CC-BY-4.0 by their creators.

Audio8-ASR-0.1B is used only as a measurement reference — its weights are never loaded, merged, or distilled from.