Vovo

Vovo is an English text-to-speech model built from scratch in Swift with hand-written Metal kernels — its own tensor engine, autograd, optimizer, data pipeline and trainer. This 20 M-parameter single-speaker voice was trained on LJSpeech in about 13 minutes on an M2 Max.

Run it from Python on Apple silicon with vovo-mlx:

pip install vovo-mlx                                        # from PyPI
pip install git+https://github.com/franckverrot/vovo-mlx    # or straight from GitHub
vovo-mlx say "The quick brown fox jumps over the lazy dog." -o fox.wav --play

from vovo_mlx import VovoTTS, SAMPLE_RATE
from vovo_mlx.audio import write_wav

tts = VovoTTS.from_pretrained("franckverrot/vovo")
write_wav("hello.wav", tts.say("Hello from Vovo.", seed=0), SAMPLE_RATE)

Files

file	what
model.safetensors	acoustic model (phone encoder → mel prior + durations → flow-matching DiT decoder), EMA weights, config in the safetensors metadata
vocoder.safetensors	Vocos mel-24kHz fine-tuned on Vovo's predicted mels with adversarial + mel losses (PyTorch key layout, also loadable by the vocos package)
vocoder_base.safetensors	the unmodified Vocos mel-24kHz weights, for comparison

Mel convention: 24 kHz, n_fft 1024, hop 256, 100 HTK mel bands, log(clamp(x, 1e-7)).

Model

Matcha-shaped: text → normalizer → ipa-dict lexicon G2P → 67-symbol phone inventory → conv prenet + 6-layer RoPE transformer (d=192) → per-phone mel prior μ and log-durations → length regulation → 6-layer DiT flow-matching decoder (d=384, adaLN-zero) sampled with 16 Euler steps and classifier-free guidance 2 by default → Vocos. Alignment during training comes from monotonic alignment search; the optimizer is Muon for the weight matrices. Word error rate on Vovo's 20-sentence test set with Apple's on-device recognizer: 2.1 %.

Limitations

One voice, English only, no streaming; out-of-vocabulary words are spelled by rules. Intelligible, not studio-clean — it is a 13-minute training run.

License

MIT (code and weights). Lexicon: ipa-dict (MIT). Vocoder derived from Vocos (MIT). Data: LJSpeech (public domain).