README.md

---
language: th
license: apache-2.0
tags:
  - whisper
  - thai
  - asr
  - speech-recognition
  - lora
  - peft
  - lotusdis
datasets:
  - custom
base_model: nectec/Pathumma-whisper-th-large-v3
pipeline_tag: automatic-speech-recognition
---

# 🇹🇭 Whisper Thai - LOTUSDIS Fine-tuned (LoRA)

Fine-tuned **nectec/Pathumma-whisper-th-large-v3** for Thai distant meeting transcription.

## Training Details
- **Task:** LOTUSDIS Distant Meeting Transcription Challenge
- **Method:** LoRA (rank=64, alpha=128)
- **Training data:** Multi-microphone (6 mic types) ≈ 95K samples
- **Epochs:** 1
- **Effective batch size:** 48
- **Precision:** BF16
- **Trainable params:** ~2.4% of total

## Usage

```python
from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="Kanompung/whisper-th-lotusdis-lora",
    device="cuda",
    chunk_length_s=30,
)

result = asr("audio.mp3", generate_kwargs={"language": "thai", "task": "transcribe"})
print(result["text"])
```

## LoRA Adapter Only
If you want just the adapter (~150 MB):
```python
from peft import PeftModel
from transformers import WhisperForConditionalGeneration

base = WhisperForConditionalGeneration.from_pretrained("nectec/Pathumma-whisper-th-large-v3")
model = PeftModel.from_pretrained(base, "Kanompung/whisper-th-lotusdis-lora-adapter")
```