Imbatmann/matcha-tts-nepali-single-v1

Matcha-TTS Nepali — v1, single-speaker (fresh)

Self-contained release of a single-speaker Nepali Matcha-TTS model: acoustic checkpoint, vocoder, the exact training configs, the matcha package, a seeded synthesis script and reference audio. Everything needed to reproduce the shipped samples bit-for-bit is in this directory.

Access: this repo is gated — submit an access request from the file listing and it will be reviewed. For questions or to be whitelisted, contact smartharry499@gmail.com.

Run	logs/train/nepali_real_fresh/runs/2026-08-11_15-33-57
Repo commit	e154b4b7f23c25010f3951b594ae310d7c5d1b9b (branch feature/finetune-english-male-voice)
Trained	2026-08-11, single NVIDIA GB10, 1000 epochs
Sample rate	22 050 Hz, 80-band mel, hop 256
Speakers	1 (single voice)
Vocabulary	union symbol set, n_vocab = 264
Params	20.9 M
Training data	chatterbox-nepali corpus (~1441 utterances), Devanagari only

Which checkpoint to use

Two checkpoints are shipped so they can be compared by ear.

file	epoch	step	note
matcha_nepali_fresh.ckpt (default)	249	5750	best loss/val_epoch = 2.7284
matcha_nepali_fresh_epoch999.ckpt	999	23000	final / last.ckpt

synthesize.py loads the epoch 249 checkpoint by default. The run trained the full 1000 epochs (its original config had max_epochs: 1000), but the best validation loss is epoch 249 and the experiment config now notes that longer runs overfit and degrade prosody. The shipped samples are all rendered from epoch 249.

Contents

version1-nepali-fresh/
├── README.md                                  this file
├── synthesize.py                              seeded, standalone synthesis (no speaker id)
├── synth_samples.sh                           regenerates samples/{short,medium,long}.wav
├── requirements.txt                           exact pinned runtime
├── MANIFEST.sha256                            checksums for every shipped file
├── checkpoints/
│   ├── matcha_nepali_fresh.ckpt               epoch 249 (best) — default
│   └── matcha_nepali_fresh_epoch999.ckpt      epoch 999 (final)
├── vocoder/
│   └── g_02500000                             HiFi-GAN universal v1 generator
├── configs/
│   ├── experiment_nepali_real_fresh.yaml      hydra experiment config
│   ├── data_nepali_real.yaml                  hydra data config
│   ├── train_config_resolved.yaml             fully-resolved config as trained
│   ├── train_overrides.yaml                   CLI overrides for the run
│   └── train.log                              training log for the run
├── matcha/                                    the model/text/vocoder package
└── samples/
    ├── texts.txt                              the 3 source sentences
    ├── phonemes.txt                           their G2P output + token counts
    ├── short.wav / medium.wav / long.wav      epoch 249, seeded

Synthesis

pip install -r requirements.txt          # see the note about `real_nepali`

# single sentence
python synthesize.py --text "नमस्कार, तपाईंलाई नमस्ते।" --out out.wav

# the two-length reference set
bash synth_samples.sh

# the final (epoch 999) checkpoint instead
python synthesize.py --ckpt checkpoints/matcha_nepali_fresh_epoch999.ckpt \
    --text "नमस्कार, तपाईंलाई नमस्ते।" --out out.wav

Defaults: --steps 25 --temperature 0.667 --denoiser_strength 0.02 --symbols union, HiFi-GAN universal v1. Useful extras: --speaking_rate 1.1 (>1 is slower), --lowpass_khz 9 (tames HF hiss), --print-phonemes.

This is a single-speaker model — there is no --spk flag; speaker conditioning is not part of the encoder, so spks is always None.

Text frontend

Devanagari is phonemized by the hand-written real_nepali G2P (profile real_nepali_v0.2), not espeak-ng — it comes from the nepali_frontend repo, pinned at commit b7395f28, installed editable. samples/phonemes.txt records the exact token string and id count for each sample sentence — diff against it (or --print-phonemes) to confirm your frontend matches this release before blaming the acoustic model.

Reproducibility

synthesize.py seeds the RNG per utterance (--seed, default 1234). This matters: the flow-matching prior is torch.randn_like(mu) * temperature, and unseeded runs give slightly different audio every time. Seeding is why the shipped wavs have stable checksums.

PYTHON=python bash synth_samples.sh      # re-render the three sample wavs
sha256sum -c MANIFEST.sha256             # confirm byte-for-byte match

Verified bit-exact on the training machine (NVIDIA GB10, CUDA 13.0, torch 2.13.0+cu130). On different GPU/CUDA/library versions the audio is perceptually identical but checksums will differ. Pass --seed -1 for varied output.

Training recipe

Warm-started from the English LJSpeech base checkpoint after its vocabulary was extended to the union set (scripts/extend_vocab.py → checkpoints/matcha_tts_english_plus_nepali.ckpt), then trained on the Nepali (chatterbox) phone stream.

Init	checkpoints/matcha_tts_english_plus_nepali.ckpt
Optimizer	Adam, lr 1e-4, no scheduler, weight decay 0
Epochs	1000 (ran to completion at training time)
Batch	32, accumulate_grad_batches: 2 (effective 64)
Precision	16-mixed, gradient_clip_val: 5.0
Encoder	RoPE, 6 layers, 192 ch, 2 heads (single-speaker: spk path not wired)
Decoder	CFM/euler, sigma_min 1e-4, channels [256, 256], snakebeta
Mel stats	mel_mean -5.980818, mel_std 2.582703
Seed	1234

Full settings in configs/train_config_resolved.yaml.

Known limitations

• Trained on a single voice from the chatterbox-nepali corpus; there is no speaker conditioning, so you cannot pick between voices with this model.
• Training data is Devanagari-only; code-switched Nepali+English is not covered by the training set.
• The run was originally allowed to train 1000 epochs; validation loss bottoms out around epoch 249 and the config note warns later epochs overfit and add noise. Prefer the shipped epoch 249 checkpoint.