Imbatmann/matcha-tts-nepali-multispeaker-v5p1

Matcha-TTS Nepali — v5.1, multi-speaker (denoised + split 8s)

Self-contained release of the 2-speaker Nepali Matcha-TTS model: acoustic checkpoint, vocoder, exact training configs, the matcha package, a seeded synthesis script and reference audio.

• Model: https://huggingface.co/Imbatmann/matcha-tts-nepali-multispeaker-v5p1 (private)
• Source: https://github.com/45Harry/matcha-tts-code (branch fix/male-voice-pacing, bundle release/version5.1-denoised-split/)

This Hugging Face repo is the runnable one (GitHub gitignores checkpoint+vocoder). Previous: v3 pitch-filtered (private), v2, v1.

What changed vs v5: v5 used raw combined_Rajesh 1665 3023 train (16kHz podcast, 39dB SNR, 11s avg, 76ms/phone) — male had glitter/unclear & mid-utterance voice drift due to noise + long spans + host leakage. v5.1 denoised 1665 via noisereduce (39→69dB floor) and split >8s at silences via prepare_male_v4.py (821 split → 2596 segs 2466/130 218m), merged with female 1516 → 3907/205 train mel -6.19/2.47, 22050Hz peak 0.707. Result: reduced glitter, clearer phones, stable male register vs v5 22.95s long → 13.08s.

Run	logs/train/male_female_2spk_v5p1/runs/2026-09-08_10-48-19
Trained	2026-09-08 10:48→14:24 (3h36m), single NVIDIA GB10, 200 epochs 20.9M
Sample rate	22050 Hz, 80-band mel, hop 256
Speakers	2
Vocabulary	union n_vocab 264
Init	checkpoints/male_female_2spk_v5_adapted.ckpt (English LJSpeech + Nepali union, 192→256)

Speaker ids

spk	Voice	Training data
0	male (Rajesh, denoised+split)	2596 segs (2466 train/130 val)
1	female (chatterbox)	1516 (1441/75)

Keep ordering — API voice → spk: "male"→0, "female"→1.

Contents

version5.1-denoised-split/
├── README.md
├── synthesize.py
├── requirements.txt
├── MANIFEST.sha256
├── checkpoints/matcha_male_female_2spk_v5p1.ckpt  last.ckpt epoch 200
├── vocoder/g_02500000
├── configs/
│   ├── data_male_female_2spk_v5p1.yaml
│   ├── experiment_male_female_2spk_v5p1.yaml
│   └── train_config_resolved.yaml
├── matcha/  # model/text/vocoder package
└── samples/
    ├── v5p1_male_short.wav (0.74s)
    ├── v5p1_male_medium.wav (5.76s)
    ├── v5p1_male_long.wav (13.08s)
    └── v5p1_female_*.wav

Synthesis

.venv/bin/python synthesize.py --ckpt checkpoints/matcha_male_female_2spk_v5p1.ckpt --vocoder vocoder/g_02500000 --spk 0 --symbols union --steps 10 --text "नमस्ते।"