qwen3-asr-0.6b-assamese-lora
LoRA adapter weights fine-tuned on top of Qwen/Qwen3-ASR-0.6B-hf for low-resource Assamese Automatic Speech Recognition (ASR).
Model Overview
- Base Architecture:
Qwen/Qwen3-ASR-0.6B-hf - Language: Assamese (
as) - Details: Sub-1B dedicated model demonstrating acoustic front-end capacity bottleneck on zero-shot phonetic transfer.
- PEFT Setup: LoRA (r=16, alpha=32, dropout=0.05) applied to all linear projections with acoustic encoder frozen.
Benchmark Performance
Evaluated on a strictly speaker-disjoint held-out test split (2,266 utterances, 31 unique speakers, 3.18 hours, 22,596 total words) with standardized Unicode NFC and punctuation stripping:
| Metric | Score | 95% Bootstrap CI |
|---|---|---|
| Word Error Rate (WER) | 107.90% | [106.87%, 108.93%] |
| Character Error Rate (CER) | 52.07% | — |
| Match Error Rate (MER) | 90.61% | — |
| Word Information Lost (WIL) | 98.93% | — |
| Training Wall-Clock Time | 6.07 hours | (NVIDIA A40 48GB) |
Word-Level Error Decomposition
- Hits: 2,526
- Substitutions: 19,479
- Deletions: 591
- Insertions: 4,311
Usage
from transformers import AutoProcessor, AutoModelForCausalLM
from peft import PeftModel
base_model_id = "Qwen/Qwen3-ASR-0.6B-hf"
adapter_id = "bijaykumarsingh/qwen3-asr-0.6b-assamese-lora"
processor = AutoProcessor.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(base_model_id, device_map="auto")
model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()
Citation
@article{singh2026assamese_lalm,
title={Architectural Inductive Bias Trumps Parameter Scale: Adapting Large Audio-Language Models to Low-Resource Assamese ASR},
author={Singh, Bijay Kumar},
journal={arXiv preprint},
year={2026}
}