mitvho09/IndicF5-Kannada-Bedtime

IndicF5 Kannada Bedtime Story Fine-Tune

Fine-tuned ai4bharat/IndicF5 on Kannada bedtime-story speech.

Training Details

• Base model: ai4bharat/IndicF5 (0.4B params)
• Training data: Rasa Kannada (18 clips, ~2min)
• Training steps: 300
• Learning rate: 1e-5
• Final loss: 0.4553
• GPU: A100-80GB (Modal)

How to Use

from transformers import AutoModel
import soundfile as sf

model = AutoModel.from_pretrained(
    "mitvho09/IndicF5-Kannada-Bedtime",
    trust_remote_code=True,
)

audio = model(
    "ಮಕ್ಕಳೇ, ನಿದ್ರೆ ಮಾಡಿ.",
    ref_audio_path="reference.wav",
    ref_text="transcript of reference",
)

sf.write("output.wav", audio, 24000)

Evaluation

Compare stock vs fine-tuned audio in the comparison/ directory.

Fine-Tuning Details

This is an experimental fine-tune for the DreamVoice competition badge. Stock IndicF5 may perform better for general Kannada TTS.