🎙️

Tamil ASR

தமிழ் பேச்சு-to-எழுத்து

State-of-the-art Tamil speech-to-text powered by osmapi/tamil-asr-qwen3

#1 FLEURS-Tamil Qwen3-ASR-1.7B Apache-2.0 1,475h training

ℹ️ Live Demo

This is a model showcase. For a live interactive demo, you can run the model locally or deploy your own Gradio Space (requires Hugging Face PRO).

🏆Benchmark Results

FLEURS-Tamil test set — 586 utterances, held out from training

Model	WER ↓	CER ↓
Tamil-ASR-Qwen3 (this model)	25.27%	7.95%
IndicConformer-600M (AI4Bharat)	25.49%	9.33%
Whisper-large-v3 (OpenAI)	48.58%	13.75%
Qwen3-ASR-1.7B (base, no Tamil)	153.29%	113.24%

📊Model Stats

1.7B

Parameters

25.27%

WER (FLEURS)

1,475h

Training Data

BF16

Precision

🚀How to Use

Install dependencies and run inference:

pip install transformers==4.57.6 qwen-asr soundfile

from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("osmapi/tamil-asr-qwen3")
model = AutoModelForMultimodalLM.from_pretrained(
    "osmapi/tamil-asr-qwen3", 
    device_map="auto"
)

Audio should be 16 kHz mono. Use beam search (num_beams=5) for best results.

Or use with ms-swift:

pip install ms-swift transformers==4.57.6 qwen-asr soundfile

swift infer \
  --model osmapi/tamil-asr-qwen3 \
  --val_dataset your_data.jsonl \
  --infer_backend pt --max_new_tokens 256 --num_beams 5

📌Model Details

Task: Automatic Speech Recognition (speech → Tamil text)
Language: Tamil (ta) only
Architecture: Qwen3-ASR — audio encoder + Qwen3 LLM decoder
Base model: Qwen/Qwen3-ASR-1.7B
Audio input: 16 kHz mono

🗂️Training Data

Source	Hours	Style
IndicVoices (AI4Bharat)	803	Natural / conversational
Shrutilipi (AI4Bharat)	462	Read / news
Kathbath (AI4Bharat)	172	Read
FLEURS-Tamil + TTS	~38	Read

🔗Links

Model Card Base Model Qwen3-ASR GitHub