bol-tts-marathi v0.1-preview — Kokoro-82M fine-tuned for Marathi

⚠️ Preview release. This checkpoint is from Stage 2 epoch 2 (pre-joint-training) of an ongoing 10-epoch Stage 2 run. Phoneme accuracy is good (ɭ for ळ is correctly distinguished from ɖ), but prosody is still being refined. A full v0.1 release will follow once adversarial training completes.

Marathi (मराठी) fine-tune of hexgrad/Kokoro-82M, built with the semidark/kokoro-deutsch training recipe.

• Model type: StyleTTS2 acoustic model + ISTFTNet decoder (Kokoro-82M architecture, unchanged)
• Parameters: 81.76 M
• Sample rate: 24 kHz
• Voices in this preview: 2 (Asha, Vivek). IV-R voices (Mukta, Dnyanesh) to follow.

Recipe repo: github.com/shreyask/bol-tts-marathi.

Voices

Voice ID	Display	Source	Meaning
mf_asha	Asha (आशा)	Rasa marathi_female	hope
mm_vivek	Vivek (विवेक)	Rasa marathi_male	wisdom
mf_mukta	Mukta (मुक्ता)	IV-R top female speaker	pearl (pending)
mm_dnyanesh	Dnyanesh (ज्ञानेश)	IV-R top male speaker	knowledge (pending)

Usage

import torch, soundfile as sf
from kokoro import KModel, KPipeline
import kokoro.pipeline as _kp

_kp.LANG_CODES["m"] = "mr"  # monkey-patch Marathi lang code

kmodel = KModel(
    repo_id="shreyask/bol-tts-marathi",
    config="config.json",
    model="kokoro-mr-v1_0.pth",
)
kmodel.train(False)

pipeline = KPipeline(lang_code="m", repo_id="shreyask/bol-tts-marathi", model=kmodel)
voice = torch.load("voices/mf_asha.pt", map_location="cpu", weights_only=True)

text = "नमस्कार, मी मराठी बोलतो."
chunks = []
for _gs, _ps, audio in pipeline(text, voice=voice, speed=0.85):
    chunks.append(audio)

sf.write("out.wav", chunks[0].numpy() if len(chunks) == 1 else torch.cat(chunks).numpy(), 24000)

Per-voice speed

Rasa voices prefer a slight slowdown; IV-R voices prefer a slight speedup. Defaults in voice_speeds.json:

{"mf_asha": 0.85, "mm_vivek": 0.90, "mf_mukta": 1.15, "mm_dnyanesh": 1.15}

Timestamps

Kokoro predicts per-phoneme durations. KModel.forward_with_tokens returns (audio, pred_dur). pred_dur is in predictor frames where 1 frame = 600 audio samples at 24 kHz (the prosody predictor runs at half the mel-frame rate; the decoder upsamples 2× before iSTFT):

audio, pred_dur = kmodel.forward_with_tokens(input_ids, ref_s, speed=1.0)
durations_sec = pred_dur.squeeze().cpu().numpy() * 600 / 24000
starts = durations_sec.cumsum() - durations_sec
# (starts[i], starts[i]+durations_sec[i]) is the time span of phoneme[i]

Training

Phase	Details
Base	hexgrad/Kokoro-82M
Stage 1	10 epochs, bs=12, fp32, ~9h on A100 SXM 80GB. Final val_loss ≈ 0.23
Stage 2 (this preview)	2 of 10 epochs, bs=8, PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
Train utts	24,676 (95/5 split)
Speakers	331 (2 Rasa + 329 IndicVoices-R)
Vocab change	ɭ (U+026D, retroflex lateral) at Kokoro slot 144

Full methodology: TRAINING_GUIDE.md.

Datasets

• AI4Bharat/Rasa (CC-BY-4.0) — Marathi, 13,900 studio-quality utts, 2 speakers.
• AI4Bharat/IndicVoices-R (CC-BY-4.0, gated) — Marathi, ~11,910 utts, 329 speakers after filtering.

Limitations

• Preview: Stage 2 adversarial training not yet complete. Expect prosody improvement in the final release.
• Minglish (Marathi + English code-switch) sounds like Indian-English accent — usable but not polished.
• IV-R-derived voices (Mukta, Dnyanesh) pending speaker selection.
• Single language (Marathi only).

License

Apache 2.0. Training data under CC-BY-4.0.

Citation

@software{bol_tts_marathi_2026, title={bol-tts-marathi: Kokoro-82M fine-tuned for Marathi}, author={Karnik, Shreyas}, year={2026}, url={https://github.com/shreyask/bol-tts-marathi}, license={Apache-2.0}}
@software{kokoro_2025, title={Kokoro-82M}, author={hexgrad}, year={2025}, url={https://github.com/hexgrad/kokoro}}
@software{kokoro_deutsch_2026, title={kokoro-deutsch}, author={semidark}, year={2026}, url={https://github.com/semidark/kokoro-deutsch}}
@inproceedings{li2024styletts2, title={StyleTTS 2}, author={Li, Yinghao Aaron and others}, booktitle={NeurIPS}, year={2024}}