Base: ai4bharat/IndicF5 (MIT license)
Dataset: SPRINGLab/IndicTTS_Kannada + synthetic bedtime clips (800 clips, ~6h)
Method: Full fine-tune (CFM component only)
Steps: 500 (best checkpoint at step 500, loss 0.4043)
GPU: A100-80GB via Modal

Evaluation (10 sentences × 2 references)

Checkpoint	Silence	Pitch Std	Syll/s	MOS	Flatness	Rank
stock	49.3%	46.4Hz	1.9	3.9	0.629	7th
v3_step0200	52.0%	51.0Hz	2.2	4.2	0.582	5th
v3_step0400	43.8%	41.1Hz	1.7	3.7	0.532	6th
v3_step0600	49.6%	47.8Hz	1.6	3.9	0.578	4th
v3_step0800	47.1%	47.2Hz	2.1	4.0	0.487	3rd
v3_final	43.8%	43.7Hz	2.5	4.2	0.538	2nd
v2_step0500	45.1%	45.4Hz	3.0	4.2	0.491	1st

Demo Audio

Kannada bedtime stories (fine-tuned)

Mood	Audio
Magical	kannada_magical.wav
Funny	kannada_funny.wav
Calming	kannada_calming.wav
Dreamy	kannada_dreamy.wav

Full pipeline (translate + narrate)

pipeline_test_kannada.wav — English story → IndicTrans2 → IndicF5 Kannada narration
reference_voice.wav — Reference voice used for cloning

How to use

python

from transformers import AutoModel

model = AutoModel.from_pretrained("sush0401/IndicF5-Kannada-Bedtime-v2", trust_remote_code=True)
audio = model("ಕನ್ನಡ ಪಠ್ಯ", ref_audio_path="reference.wav", ref_text="Reference transcript")

Part of DreamVoice

DreamVoice — bedtime stories in a parent's cloned voice.