SayedShaun/Qwen3-ASR-0.6B-Bengali

Qwen3-ASR-0.6B — Bengali (merged)

Qwen3-ASR-0.6B fine-tuned to transcribe Bengali — a language the base model wasn't pretrained on — with the LoRA adapter merged into the base weights. This repo holds a plain, standalone Qwen3ASRForConditionalGeneration checkpoint: no PEFT/adapter loading step, so it drops straight into any tool that already knows how to serve Qwen3-ASR, including vLLM.

Trained on Bengali audio only, so non-Bengali audio comes out as garbled phonetic Bengali rather than that language. Merging is one-way — unlike the LoRA repo, there is no model.disable_adapter() escape hatch here to recover the original multilingual model. If you need both Bengali and the base model's other languages from one set of weights, use the LoRA adapter instead.

Results

split	utterances	WER	CER
test (full, held-out)	10,485	14.74%	5.34%

Scored on the full held-out test set (pre-merge, on the LoRA adapter this was merged from), with the generation budget sized to each clip's length. Training was monitored throughout on a small dev sample and stopped at 72k of a planned 90k steps once that curve plateaued.

Training

• Base: Qwen/Qwen3-ASR-0.6B-hf, frozen, bf16
• LoRA: r=64, alpha=64, dropout 0.05, on attention + MLP in both the decoder and audio encoder — 59.2M trainable params (7.0% of 841.6M), merged into the base weights for this repo
• Optimizer: batch 32 (2 × 16 grad accumulation), LR 1e-4 cosine, bf16, gradient checkpointing
• Hardware: 1x NVIDIA RTX 5080
• Code: sayedshaun/qwen3-asr-finetune-bn

Data

753,378 Bengali utterances (1,174 h, 0.5–20s clips) from 9 sources:

source	utterances	hours	share
IndicVoices (AI4Bharat)	309,683	599.7 h	51.1%
OpenSLR-53 — Large Bengali ASR training data set	214,324	211.2 h	18.0%
Vaani (ARTPARK / IISc)	114,628	159.0 h	13.5%
Kathbath (AI4Bharat)	46,458	82.3 h	7.0%
Mozilla Common Voice (via Mozilla Data Collective)	43,892	73.2 h	6.2%
Shrutilipi (AI4Bharat)	17,521	27.4 h	2.3%
FLEURS bn_in (Google)	3,921	13.2 h	1.1%
OpenSLR-37 — High quality TTS data for Bengali	1,850	2.9 h	0.2%
Ben-10 (Bengali.AI)	1,101	4.9 h	0.4%

IndicVoices and OpenSLR-53 alone make up 69% of the hours — worth keeping in mind when judging what conditions this model is well calibrated for.

Usage (transformers)

Requires transformers>=5.13.0.

import torch
from transformers import AutoProcessor, Qwen3ASRForConditionalGeneration

MODEL = "SayedShaun/Qwen3-ASR-0.6B-Bengali"

processor = AutoProcessor.from_pretrained(MODEL)
model = Qwen3ASRForConditionalGeneration.from_pretrained(
    MODEL, dtype=torch.bfloat16
).eval().to("cuda")

conversation = [
    {"role": "user", "content": [{"type": "audio", "path": "clip.wav"}]},
    {"role": "assistant", "content": [{"type": "text", "text": "language Bengali<asr_text>"}]},
]
inputs = processor.apply_chat_template(
    conversation, tokenize=True, return_dict=True,
    padding=True, continue_final_message=True,
).to(model.device, model.dtype)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=256, do_sample=False)

text = processor.batch_decode(
    out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True
)[0]
print(text.split("<asr_text>", 1)[-1].strip())

For clips longer than a few seconds, size max_new_tokens to the audio's duration (~30 tokens/second is a safe rate) — a small fixed value will silently truncate the transcript.

Usage (vLLM)

Because this repo is a plain Qwen3ASRForConditionalGeneration checkpoint (no adapter to load), it serves like any other Qwen3-ASR model. vLLM lists Qwen3ASRForConditionalGeneration as a supported transcription architecture with an OpenAI-compatible API:

vllm serve SayedShaun/Qwen3-ASR-0.6B-Bengali

Then transcribe against the OpenAI-compatible transcriptions endpoint:

curl http://localhost:8000/v1/audio/transcriptions \
  -H "Authorization: Bearer EMPTY" \
  -F file=@clip.wav \
  -F model=SayedShaun/Qwen3-ASR-0.6B-Bengali

Check vllm serve --help / the vLLM release you're on for exact flags — serving details for newer architectures can shift between versions. vLLM's supported-models list also shows LoRA support for this architecture, so the unmerged adapter repo can alternatively be served directly on top of the base model without a separate merge step.

License

Released under the Apache License 2.0, the same license as the base model Qwen/Qwen3-ASR-0.6B-hf. Free for commercial and non-commercial use, modification, and redistribution, with attribution.