Qwen3-ASR-1.7B-hf Orze

This is a self-contained, merged English ASR specialization of Qwen/Qwen3-ASR-1.7B-hf, selected by Orze experiments for the Open ASR Leaderboard. It loads directly with native Transformers; no PEFT adapter or custom code is required.

Usage

import torch
from transformers import AutoProcessor, Qwen3ASRForConditionalGeneration

model_id = "bosonai/Qwen3-ASR-1.7B-hf-orze"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3ASRForConditionalGeneration.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)

inputs = processor.apply_transcription_request(
    audio="audio.wav",
    language="en",
).to(model.device, model.dtype)
output_ids = model.generate(**inputs, max_new_tokens=512)
generated_ids = output_ids[:, inputs["input_ids"].shape[1]:]
text = processor.decode(
    generated_ids, return_format="transcription_only"
)[0]
print(text)

The included generation configuration uses deterministic four-beam decoding with a 0.7 length penalty. This setting was selected using the Open ASR public test scores and public proxy evaluations; it changes decoding only and does not use private data.

Evaluation status and benchmark-fitting disclosure

The results below are self-evaluated public-set results, not maintainer-verified leaderboard results. The public test scores were used during checkpoint and decoding selection, so they are benchmark-fitted development results rather than an independent estimate of generalization. This model has no official private scripted or private conversational score or rank.

Neither the private Appen nor DataOcean leaderboard datasets was accessed. AppTek Call-Center Dialogues is a separate public dataset and is unrelated to Appen.

Self-evaluated public benchmark results

The results below were produced with the official leaderboard Transformers evaluator and normalizer at commit d1e99b25524814332d6868a5645e568670834cfb.

Measured throughput was 74.01 RTFx on the local A100 evaluation run. The same decoding settings were used for every dataset.

Public proxy checks used during development

The final weights were compared at length penalties 0.8 and 0.7 on public English proxy sets used during development. These checks informed selection and are not independent validation. Lower WER is better.

The three AppTek selections use different calls and were never used for training.

Training disclosure

Orze first retained two compatible acoustic-side updates: an audio projector update at step 50 and an audio-tower linear-layer update at step 75. A final rank-16 audio-tower update was trained for 25 steps on a balanced public-data mix and fused at one-quarter strength. The final update's sampler drew from:

• Switchboard train: 6,000 utterances
• accented Common Voice English train: 5,000
• TED-LIUM train: 4,000
• AMI train: 2,000
• LibriSpeech train-other-500: 1,000
• VoxPopuli English train: 500
• Earnings22 public train split: 500
• LibriSpeech train-clean-100: 500
• SPGISpeech train: 500

No Open ASR Leaderboard test audio, private Appen or DataOcean leaderboard data, selected AppTek proxy audio, or selected MInDS proxy partition was used for training. Candidate and decoding selection used the Open ASR public test scores and public proxy sets. The published checkpoint is fully merged and contains exactly 2,038,052,480 parameters in its safetensors metadata (2.04B when rounded to two decimals).

Scope and limitations

This specialization was optimized and evaluated for English short-form ASR. The multilingual abilities of the base model were not evaluated after tuning, so use the original Qwen checkpoint when broad multilingual performance is the priority. See the base model card for architecture, supported languages, safety considerations, and additional usage details.

License

Apache-2.0, following the base checkpoint.