kleintts

Klein-TTS text-to-speech models in plain PyTorch: Pro (13M; English, Spanish, French, Italian), Flash (1.7M) and Flash Lite (0.4M) (English, French, Italian). One small package: the model, the text front-end (the full misaki + spaCy one used for training, installed for you), the voices (17 in Pro) and the synthesis loop.

pip install kleintts

from kleintts import KleinTTS

tts = KleinTTS.from_pretrained("pro")      # "pro", "flash", "flash-lite", a Hugging Face repo id or a local folder
out = tts.tts("Hello from Klein TTS.", voice="Sarah", speed=1.0, seed=0)
tts.save_wav("hello.wav", out.waveform, out.sample_rate)

kleintts "Ciao, questa è una prova." --voice Giulia -o ciao.wav
kleintts "Hello there." --model flash -o hello.wav
kleintts --list-voices

tts(text, voice, voice2, mix, speed, seed, pause, phonemes, use_m): voice2/mix blend two voices of one language, phonemes skips the text front-end, use_m switches on the optional evaluation stream (off by default).

No library? Use Transformers

The model repo also carries Transformers remote code, so this works without installing kleintts. It uses a compact copy of the English front-end (bundled dictionaries, no spaCy), which agrees with the full one on about 97% of English sentences:

from transformers import AutoModel
model = AutoModel.from_pretrained("klein-ml/Klein-TTS-X1.5-Pro", trust_remote_code=True)
out = model.tts("Hello from Klein TTS.", voice="Sarah")
model.save_wav("hello.wav", out.waveform, out.sample_rate)

Layout

• src/kleintts/core.py: everything (generator, phonemizer, voices, synthesis). It is the first half of modeling_klein_tts.py in the model repo, with one block swapped: the package front-end uses misaki + spaCy, the remote code uses vendor/frontend_vendored.py.
• hf/: the Transformers wrapper and config class, and an example script.
• build_release.py: turns a training checkpoint into the upload folder (safetensors, config.json, remote code).
• vendor/: make_g2p.py builds the compact English front-end and lexicons from misaki; test_g2p.py compares it with the real one.
• test_package.py: the package against the original training code, and every language.
• test_remote_code.py: the Transformers path with misaki, spaCy and num2words blocked.

License

Apache-2.0 (see LICENSE).