Korva-TTS v1 — ONNX bundle

Pre-trained multilingual (Vietnamese + English) text-to-speech weights for korva-tts.

Quick start

from korva_tts import KorvaTTS

tts = KorvaTTS()                                # downloads this repo on first run
wav, sr = tts.synthesize("Xin chào", voice="female1", language="vi")

import soundfile as sf
sf.write("hello.wav", wav, sr)                  # 44.1 kHz mono

Bundle contents

File	Size	Purpose
`text_encoder.onnx`	~27 MB	IPA text ids → text embeddings (style-conditioned)
`vector_estimator.onnx`	~126 MB	Flow-matching vector field for latent denoising
`vocoder.onnx`	~97 MB	LatentDecoder1D: 24-ch latent → 44.1 kHz waveform
`length_pred_style.onnx`	~1.5 MB	Style-token-conditioned duration prediction
`stats.npz`	<1 MB	`mean`, `std`, `normalizer_scale` for latent denormalization

Total ≈ 252 MB.

Voice-cloning paths (reference_encoder.onnx, duration_predictor.onnx) are intentionally NOT included — Korva v0.1 ships pre-baked voices only.

Architecture

Flow-matching latent TTS:

text → text_encoder → text_emb
      ↓                  ↓
length_pred_style → T_lat   vector_estimator(noise, text_emb, style, …)
                                 (N=8 flow-matching steps)
                                    ↓
                               denormalize + reshape
                                    ↓
                                 vocoder → waveform (44.1 kHz)

See onnx_export_v1/README.md for the full I/O contract (input/output names, dtypes, dynamic axes).

Languages

Vietnamese (vi) — via viphoneme G2P
English (en) — via espeak-ng (phonemizer-fork)

Inline mixing supported with <en>...</en> tags inside Vietnamese text.

Intended use & limitations

Intended: on-device speech synthesis for vi/en text. CPU runs in real time on a modern laptop; GPU optional via the [gpu] extras.
Out of scope: voice cloning (not shipped), CJK languages, streaming.
Not suitable for: medical, legal, or safety-critical applications. Synthesis quality has not been formally evaluated for any production use.

License

MIT for the code and these ONNX weights.

Note: voices/male1.json in the korva-tts package is derived from the LibriTTS corpus (CC-BY-4.0); see the package voices/ directory for full attribution.

Repo info

GitHub: https://github.com/nghiakvnvsd/Korva-TTS
HF repo: https://huggingface.co/nghiakvnvsd/korva-tts-v1
Export bundle revision: see commit hash on this repo.

Downloads last month: -; Downloads are not tracked for this model. How to track