tts-eu-pt 0.1.0

A permissively-licensed European Portuguese (pt-PT) text-to-speech voice, fine-tuned from Kokoro-82M.

tts_eu_pt - an European Portuguese text-to-speech based on Kokoro TTS

A small, permissively-licensed (Apache-2.0) European Portuguese (pt-PT) voice, fine-tuned from Kokoro-82M. It runs on CPU, ships as a tiny Python package.

Why it's different

The hard part of pt-PT TTS isn't the acoustics — it's reading text the way a Portuguese speaker actually does. tts_eu_pt handles, out of the box:

Input	Spoken as
25, 125, 1139	vinte cinco, cento e vinte cinco, mil cento e trinta e nove
16:54, 16:00	dezasseis e cinquenta e quatro, dezasseis horas
1139-1185	mil cento e trinta e nove a mil cento e oitenta e cinco
D. Afonso I	Dom Afonso Primeiro
IA, LLM, UTC	i-á, éle-éle-éme, u-tê-cê
online, software	kept in their English pronunciation (as Portuguese speakers say them)

The G2P (Grapheme-to-Phoneme) is built on the Apache-2.0 TugaPhone front-end (Lisbon lect).

Install

pip install tts_eu_pt

Quick start

from tts_eu_pt import TTS

tts = TTS()                                       # downloads the voice on first run
wav = tts.say("Olá! São dezasseis horas.")        # numpy float32 @ 24 kHz
tts.save("ola.wav", "Bem-vindo ao tts_eu_pt.")

Command line:

tts-eu-pt "D. Afonso I reinou de 1139 a 1185." reis.wav
tts-eu-pt --text "Bom dia." --speed 1.1 ola.wav      # or python -m tts_eu_pt.cli

The voice (~313 MB) downloads once from logus2k/kokoro_tts_eu_pt and is cached by huggingface_hub. To run offline or from a different copy, pass TTS(model_path=..., voicepack_path=...) or set TTS_EU_PT_REPO.

Web echo demo

A one-page demo: type a sentence, press Enter, the page speaks it back in European Portuguese.

pip install "tts_eu_pt[server]"
python -m tts_eu_pt.server            # open http://localhost:8000

A static page (tts_eu_pt/web/) plus a ~50-line FastAPI backend (tts_eu_pt/server.py) — browsers can't run the torch model directly. To run it against local weights, set TTS_EU_PT_MODEL and TTS_EU_PT_VOICEPACK; PORT overrides the default 8000.

How it works

• Weights encode the language (pt-PT phonology/accent); a small voicepack encodes the speaker. They are separate files.
• Text → tts_eu_pt.g2p (normalisation + TugaPhone) → phoneme ids → Kokoro acoustic model → 24 kHz audio.
• The Kokoro acoustic model is vendored (tts_eu_pt/kokoro/, Apache-2.0). Only the acoustic KModel is used; the upstream pipeline's GPL-dependent G2P is not.

Licensing

Apache-2.0.

Status

Early release: the runtime + SDK + examples.