---

license: apache-2.0 pipeline_tag: automatic-speech-recognition library_name: transformers base_model: Musci-research/Musci-ASR-qwen-1019

Musci-ASR-qwen-1019-hf

This is a Transformers-native conversion of Musci-ASR-qwen-1019. No additional training was performed during conversion. The source model is a GRPO fine-tune of Qwen3-ASR and is not an official Qwen release.

Conversion Details

• Source checkpoint revision: 0fdbc84b7dcccba011158700bf0860125e5de40a.
• Converted using the official Transformers convert_qwen3_asr_to_hf.py at commit 2e12b325d5c4532f3ec5b0ec54e1ff0b7e0db397.
• Environment: Transformers 5.21.0.dev0 and huggingface_hub 1.31.0, running offline in an H200 container.
• The source checkpoint has tie_word_embeddings=True and stores only the shared embedding tensor. Before the official converter's strict=True weight loading, the missing thinker.lm_head.weight entry was restored as an alias of thinker.model.embed_tokens.weight. No missing weights were randomly initialized.
• The converted model retains tied input embeddings and output head. Saving only one copy of the shared tensor in safetensors is expected.

Loading

Use a Transformers version that includes the native Qwen3-ASR implementation; the exact conversion commit is listed above. Neither the qwen-asr package nor trust_remote_code=True is required.

import torch
from transformers import Qwen3ASRForConditionalGeneration, Qwen3ASRProcessor

model_id = "Musci-research/Musci-ASR-qwen-1019-hf"
processor = Qwen3ASRProcessor.from_pretrained(model_id)
model = Qwen3ASRForConditionalGeneration.from_pretrained(
    model_id, dtype=torch.bfloat16, attn_implementation="sdpa"
).to("cuda")
model.eval()

Validation Status

The checkpoint passed the converter's strict weight-loading checks, as well as native configuration, processor, and weight-name checks.

Open ASR Leaderboard Results

Evaluated with transformers/run_eval.py from the Open ASR Leaderboard repository (Transformers backend, bf16, SDPA, batch size 256, max_new_tokens=512) on a single H200, scored with normalizer/eval_utils.score_results.

Dataset	WER	RTFx
AMI-Cleaned	6.53	653.0
Earnings22-Cleaned-AA-chunked	5.83	436.4
GigaSpeech-Cleaned	6.41	460.6
LibriSpeech test.clean	1.03	621.4
LibriSpeech test.other	2.18	579.4
SPGISpeech	1.86	813.3
URGENT2024	10.49	708.7
Voice Arena Monsoon en-IN	3.44	588.0
VoxPopuli-AA-Cleaned	1.47	390.6
Average (public, 9 sets)	4.36	658.3

Training Data and Attribution

This checkpoint is fine-tuned from Qwen3-ASR-1.7B-HF on an internal mixture built from the public Open ASR Leaderboard training datasets and non-native English speaker training datasets.

Retain attribution to Qwen3-ASR and the source fine-tuned model, and comply with the Apache-2.0 license.