kokoro_tts_eu_pt — European Portuguese (pt-PT) voice

A small, permissively licensed (Apache-2.0) European Portuguese text-to-speech voice, fine-tuned from Kokoro-82M. It runs on CPU and is driven by an Apache-2.0 grapheme-to-phoneme front-end, so nothing GPL is required to install or run it.

Use it through the tts_eu_pt package (source on GitHub):

pip install tts_eu_pt

from tts_eu_pt import TTS

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

tts-eu-pt "D. Afonso I reinou de 1139 a 1185." reis.wav

Files

File	Size	What it is
tuga_kokoro.pth	327 MB	Acoustic model weights — encodes the language (pt-PT phonology and accent)
tuga_voicepack.pt	524 kB	Voicepack — encodes the speaker (timbre), a [510, 1, 256] float32 tensor

Weights and voicepack are deliberately separate: the same weights can drive a different speaker by swapping the voicepack. The voicepack is indexed by phoneme-string length (pack[len(phonemes) - 1]), which is why its first dimension is 510.

Why it exists

The hard part of pt-PT TTS is not the acoustics — it is reading text the way a Portuguese speaker actually does, and doing it without GPL dependencies. Most open TTS stacks reach for espeak-ng (GPLv3) for G2P, which forces copyleft on anything built on top. This voice uses TugaPhone (Apache-2.0) instead, on the Lisbon lect (pt-PT-x-lisbon).

Handled out of the box by the tts_eu_pt front-end:

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 pt-PT speakers say them

Note the European-Portuguese quirk in the twenties: 21–24 keep the connector e ("vinte e quatro") while 25–29 drop it ("vinte cinco"), at every scale.

Model details

• Architecture: Kokoro (StyleTTS2-derived) — ALBERT-based text encoder, prosody predictor, iSTFTNet decoder. Five modules: bert, bert_encoder, predictor, text_encoder, decoder.
• Parameters: 81.8 M
• Output: 24 kHz mono, float32
• Phoneme vocabulary: 178 tokens, 510-token input cap (longer text is chunked by sentence upstream)
• G2P: TugaPhone, lect pt-PT-x-lisbon, plus a 774-entry English-loanword pronunciation table
• Precision: float32; runs on CPU or CUDA

Performance

Measured on CPU (no GPU), synthesising with tts_eu_pt, median real-time factor ≈ 0.14 — roughly 7× faster than real time, about 1 s of compute per 7 s of audio. Model load (download excluded) is ~2.3 s.

RTF varies with utterance length, because per-call overhead dominates short inputs:

Audio produced	Compute	RTF
3.5 s	0.93 s	0.27
5.9 s	0.89 s	0.15
7.1 s	0.95 s	0.13
44.2 s	5.88 s	0.13

So expect ~0.13–0.15 for sentences and paragraphs, and closer to 0.27 for very short phrases. Figures are from one commodity CPU and are indicative, not a benchmark.

Intended use

General-purpose European Portuguese speech synthesis: voice assistants, screen readers, audio for accessibility, prototyping, and offline/on-device use where a GPL-free stack matters.

Limitations

• European Portuguese only. Brazilian Portuguese is not supported and will sound wrong.
• Single speaker. One voicepack ships here.
• Honorific D. is expanded to the masculine Dom; D. Maria → Dom Maria is wrong. Regnal numerals likewise map to masculine ordinals (Primeiro, Segundo, …), so queens are mis-gendered.
• Regnal numeral mapping covers I–XX only.
• An unrecognised lect string silently falls back to generic pt-PT rather than raising.
• Four TugaPhone symbols fall outside Kokoro's vocabulary and are mapped or dropped: uvular trill ʀ→ʁ, dark ɫ→l, ASCII g→ɡ, and the apico-alveolar diacritic (dropped — a genuine pt-PT feature the vocabulary cannot represent).
• No formal listener evaluation has been run yet. There is no MOS score for this voice; quality claims here are architectural and objective-metric based, not perceptual. Judge it by listening before relying on it in production.

Out-of-scope / misuse

Do not use it to impersonate a real person or to generate speech presented as a genuine recording of someone. Synthetic speech should be disclosed as synthetic.

Training

Fine-tuned from Kokoro-82M on European Portuguese, with the voicepack extracted separately from a single-speaker corpus, so that language and timbre stay disjoint. Training used the same TugaPhone G2P as inference — a mismatch between training-time and inference-time phonemisation is the classic way to wreck a fine-tune of this kind.

Licensing

Apache-2.0, and everything required to run it is permissively licensed.

Component	License
These weights + voicepack	Apache-2.0 (fine-tuned from Apache-2.0 Kokoro-82M)
Kokoro acoustic model code (vendored in tts_eu_pt)	Apache-2.0
TugaPhone (pt-PT G2P)	Apache-2.0
PyTorch	BSD-3-Clause

The upstream Kokoro KPipeline is not used, because it pulls in GPLv3 espeak-ng via misaki. Only the acoustic KModel is. The English-loanword pronunciation table was generated offline with espeak-ng at data-prep time; only the resulting phoneme strings are distributed, as data — espeak-ng is not a runtime dependency and is not redistributed.

The Kokoro architecture derives from StyleTTS2 (MIT).

Citation

@misc{kokoro_tts_eu_pt,
  title  = {kokoro_tts_eu_pt: a permissively-licensed European Portuguese TTS voice},
  author = {Cruz, Antonio},
  year   = {2026},
  url    = {https://huggingface.co/logus2k/kokoro_tts_eu_pt}
}

Acknowledgements

hexgrad for Kokoro-82M, and TigreGotico for TugaPhone.