Hindi TTS — distilled IndicF5 student (~190M)
Paper: arXiv:2607.18662 — Staged Depth-Pruning Distillation of a Flow-Matching Text-to-Speech Teacher: A Compact Hindi Speech Synthesizer
A distilled IndicF5 "student" (DiT depth-12 flow-matching, ~190M params) for Hindi
text-to-speech, plus a Vocos mel→waveform vocoder (24 kHz). Voice-cloning: it speaks
in the bundled reference voice (ref_audio.wav + ref_text.txt).
Live demo: https://huggingface.co/spaces/5ivatej/hindi-tts-190M
Files
| File | Description |
|---|---|
model.pt |
distilled d12 student weights (fp32, EMA) |
mel_basis.pt |
fixed mel filterbank (matches training) |
vocab.txt |
IndicF5 character vocab (2545 tokens) |
vocos/ |
Vocos vocoder (config.yaml + pytorch_model.bin) |
ref_audio.wav, ref_text.txt |
reference voice + its transcript |
Usage
The inference code lives in the demo Space repo. Point it at this model repo (default) and it downloads the weights automatically:
from infer_core import HindiTTS # from the Space repo
import soundfile as sf
tts = HindiTTS() # pulls weights from this repo on first run
sr, wave = tts.synth("नमस्ते, यह एक परीक्षण है।")
sf.write("out.wav", wave, sr)
License
Released under the MIT License. Note this model is derived from IndicF5 / F5-TTS — please also review and respect the upstream model and code licenses for your use case.
Benchmark
50 unique unseen sentences from the FLEURS hi_in test split, identical inputs for all models
(NFE 32, CFG 2.0, fixed seed for flow models). Full methodology, per-sentence scores, and scripts: eval/.
| Model | Params | WER ↓ | CER ↓ | Speaker SIM ↑ | UTMOS ↑ | RTF (median) ↓ |
|---|---|---|---|---|---|---|
| IndicF5 (teacher) | 337M | 0.098 | 0.030 | 0.784 | 3.79 | 5.56 (M4 MPS) |
| hindi-tts-190M (this model) | 190M | 0.170 | 0.077 | 0.750 | 3.65 | 3.13 (M4 MPS) |
| MMS-TTS-hin | 36M | 0.195 | 0.066 | 0.278* | 3.54 | 0.31 (CPU) |
* MMS-TTS is a fixed-voice model (no voice cloning); similarity to this model's reference speaker is shown for completeness.
- WER/CER: ASR judge
Harveenchadha/vakyansh-wav2vec2-hindi-him-4200(greedy CTC), punctuation + nukta normalized. This open judge has a higher intrinsic error floor than the gated IndicWav2Vec-Hindi used in the paper — compare rows to each other, not to the paper's WER table. - Speaker SIM: ECAPA-TDNN cosine similarity vs
ref_audio.wav. - UTMOS: predicted MOS (utmos22_strong).
- RTF: synthesis time ÷ audio duration, fp32, median over 50 sentences after warmup. On a CUDA GPU (paper setting) synthesis is real-time.
Takeaway: the 190M student retains ~96% of the 337M teacher's naturalness and speaker similarity at 56% of the parameters and 1.8× faster synthesis, and is the only compact Hindi model here with voice cloning.