Downloads · 30 days
0
qwen-asr-0.6b-noisy-hinglish
addyo07/qwen-asr-0.6b-noisy-hinglish
qwen-asr-0.6b-noisy-hinglish is a automatic speech recognition model from addyo07. Use it when you need speech turned into text. It is set up for onnx. The card lists the license as apache-2.0.
Qwen3-ASR-0.6B fine-tuned on the Noisy Hinglish ASR Corpus using QLoRA. Optimized for streaming Hindi/Hinglish ASR in noisy desktop conditions with ONNX INT8 deployment.
Access
Public
Updated Jun 7, 2026
Repo size
15.6 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors3.2 GB · 75%
At a glance
- Task
- Automatic Speech Recognition
- Library
- onnx
- License
- apache-2.0
- Access
- Public
- Created
- May 19, 2026
- Updated
- Jun 7, 2026
- SHA
- 8c9011f1
Base models
Datasets
Model card
From the Hugging Face model README
Qwen3-ASR-0.6B — Noisy Hinglish Fine-Tune
Qwen3-ASR-0.6B fine-tuned on the Noisy Hinglish ASR Corpus using QLoRA. Optimized for streaming Hindi/Hinglish ASR in noisy desktop conditions with ONNX INT8 deployment.
Primary result: CJK hallucination reduced from 11% (base) to 0.4% on noisy Hindi audio.
Model Info
| Property | Value |
|---|---|
| Base model | Qwen3-ASR-0.6B |
| Fine-tuning method | QLoRA (rank=32, alpha=64, 4-bit NF4) |
| Epochs | 5 |
| Batch size (effective) | 32 (4 × 8 grad accum) |
| Learning rate | 2e-4 (cosine decay) |
| GPU | RTX 5070 Ti (16GB) |
| Training time | ~7 hours |
| ONNX export | INT8 dynamic quantization (encoder + decoder) |
Key Benchmark Results
| Test Set | WER | CER | CJK Halluc. Rate | RTF (CPU) |
|---|---|---|---|---|
| Clean Hindi (500 samples) | 26.5% | 12.8% | 0.2% | 1.36 |
| Hinglish (1036 samples) | 24.9% | 18.5% | 0.0% | 1.34 |
| Noisy Hindi (250 samples) | 46.1% | 28.8% | 0.4% | 1.90 |
| Negatives — false trigger rate | — | — | — | 4.4% FTR |
Streaming (window=2.0s, step=0.8s):
- TTFT: 292ms
- Transcript flip rate: 0
- Transient CJK in stream: 0
- WER vs offline: +10.9% (boundary artifact)
Model Files
model/
├── onnx/ # INT8 quantized ONNX (sherpa-onnx deployment)
│ ├── conv_frontend.onnx
│ ├── encoder.int8.onnx
│ ├── decoder.int8.onnx
│ └── tokenizer/
├── lora/ # Unmerged QLoRA adapters (checkpoint-1470)
│ ├── adapter_config.json
│ ├── adapter_model.safetensors
│ └── training state files
└── pytorch/ # Full merged weights (HF compatible)
├── model.safetensors
├── config.json
├── tokenizer.json
└── ...
Usage
sherpa-onnx (deployment)
import sherpa_onnx
recognizer = sherpa_onnx.OfflineRecognizer.from_transducer(
encoder="model/onnx/encoder.int8.onnx",
decoder="model/onnx/decoder.int8.onnx",
joiner="model/onnx/conv_frontend.onnx",
tokens="model/onnx/tokenizer/tokens.txt",
)
PyTorch / Transformers
from transformers import AutoModelForCTC, AutoProcessor
model = AutoModelForCTC.from_pretrained("addyo07/qwen-asr-0.6b-noisy-hinglish")
processor = AutoProcessor.from_pretrained("addyo07/qwen-asr-0.6b-noisy-hinglish")
Known Limitations
- RTF > 1.0 on CPU — not yet real-time on i5-class. INT4 requantization or thread tuning needed.
- Noisy WER gap — 46% vs 1.7B baseline's 30%. More noise augmentation needed.
- False triggers — 4.4% on silence/noise inputs. More diverse negatives needed.
Training & Export Scripts
All in scripts/:
| Script | Purpose |
|---|---|
train_lora.py | QLoRA fine-tuning |
merge_lora.py | Merge LoRA adapter into base weights |
export_onnx.py | ONNX export + INT8 quantization |
run_offline_eval.py | Parallelized benchmark evaluation |
streaming_sim.py | Streaming chunk simulation |
Training Data
Trained on the Noisy Hinglish ASR Corpus — a curated 28,681-sample, 36.6-hour dataset of Hindi, English, Hinglish, and noise-negative frames.
License
Apache 2.0
- Task
- Automatic Speech Recognition
- Library
- onnx
- License
- apache-2.0
- Languages
- hi, en
- Created
- May 19, 2026
- Updated
- Jun 7, 2026