whisper-th-lotusdis-lora

🇹🇭 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

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):

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")