Pocket TTS Basque Euskarazko ahots-sintesia, denbora errealean zure ordenagailuan · Open Basque text-to-speech that runs in real time on a CPU

~100 Mparameters, 6-layer backbone
3.5×real time on 2 CPU threads (fp32, this page's clips)
5bundled voices + zero-shot cloning
CC BY 4.0model weights

Pocket TTS Basque is a Basque (euskara) adaptation of Kyutai's Pocket TTS, trained on the synthetic BSSD dataset (470 h, CC BY-SA 4.0). A 24-layer teacher was fine-tuned first and then distilled into this 6-layer model, which keeps the teacher's pronunciation and prosody while running several times faster than real time on an ordinary laptop CPU. All clips on this page were generated by the released model on 2 CPU threads; nothing was hand-picked.

Try it

# with uv installed, no setup needed
uvx pocket-tts generate --config hf://itzune/pocket-tts-basque/basque.yaml \
  --voice hf://itzune/pocket-tts-basque/voices/Maider.safetensors \
  --text "Kaixo! Hau euskarazko ahots sintetiko bat da." --output kaixo.wav

# your own voice: 5-10 s of clean Basque speech as the prompt
uvx pocket-tts generate --config hf://itzune/pocket-tts-basque/basque.yaml \
  --voice my_voice.wav --text "Kaixo!" --output kaixo.wav

Write the text as it should be read (numbers, dates and abbreviations in words). Python usage, evaluation details and licences are in the model card.

Voices

Maider and Antton are the two professional speakers of the HiTZ-Aholab studio corpus (CC BY 4.0). Nahia, Ane and Aitor are synthetic voices from the BSSD voice bank. Each reads the same three sentences: a news item, a narrative line and a question with numbers.

Maider female · HiTZ-Aholab studio speaker

Use it: --voice hf://itzune/pocket-tts-basque/voices/Maider.safetensors

Antton male · HiTZ-Aholab studio speaker

Use it: --voice hf://itzune/pocket-tts-basque/voices/Antton.safetensors

Nahia female · synthetic voice

Use it: --voice hf://itzune/pocket-tts-basque/voices/Nahia.safetensors

Ane female · synthetic voice

Use it: --voice hf://itzune/pocket-tts-basque/voices/Ane.safetensors

Aitor male · synthetic voice

Use it: --voice hf://itzune/pocket-tts-basque/voices/Aitor.safetensors

Comparison with other Basque TTS

The same sentences and the same two voices, synthesised with three systems. Piper uses the community Maider and Antton voices trained on the HiTZ-Aholab recordings; OmniVoice clones the voice from the very same 5-7 s prompt clip that our voice states were built from (it is also the system that generated the BSSD training data, i.e. our teacher's teacher).

Pocket TTS Basque (this model)~100 M parameters, 438 MB fp32 (int8 available) · CPU, several times faster than real time · zero-shot voice cloning · CC BY 4.0
Piper (VITS)~20 M parameters per voice, ~60 MB ONNX · CPU, very fast · fixed voices, espeak-ng phonemes · voices Apache-2.0
OmniVoice (k2-fsa)~0.6 B parameters · needs a GPU for practical use · zero-shot cloning, 100+ languages · CC BY-NC (weights)

Maider Bihar, goizeko hamarretan, Donostiako liburutegi berria inauguratuko dute, eta sarrera doakoa izango da.

Pocket TTS Basque
Piper
OmniVoice

Maider Mendian galdu ginenean, ilargiaren argiak lagundu zigun bidea aurkitzen; isiltasuna erabatekoa zen.

Pocket TTS Basque
Piper
OmniVoice

Maider Zenbat balio du txartel horrek? Hogeita bost euro, baina ikasleek beherapena dute.

Pocket TTS Basque
Piper
OmniVoice

Antton Bihar, goizeko hamarretan, Donostiako liburutegi berria inauguratuko dute, eta sarrera doakoa izango da.

Pocket TTS Basque
Piper
OmniVoice

Antton Mendian galdu ginenean, ilargiaren argiak lagundu zigun bidea aurkitzen; isiltasuna erabatekoa zen.

Pocket TTS Basque
Piper
OmniVoice

Antton Zenbat balio du txartel horrek? Hogeita bost euro, baina ikasleek beherapena dute.

Pocket TTS Basque
Piper
OmniVoice

How it was built

  1. Data. BSSD: 470 h of synthetic Basque speech generated with OmniVoice from curated Basque text, 2,400+ voices (a synthetic voice bank plus consented real prompts), filtered with two Basque ASR systems, UTMOS and speaker checks. Word alignments from a Basque wav2vec2 model.
  2. Teacher. Kyutai's 24-layer English Pocket TTS fine-tuned for Basque with a new 4000-token SentencePiece tokenizer (official new-language recipe). Released as pocket-tts-basque-24l.
  3. Student. Depth distillation of the teacher into a 6-layer backbone with classifier-free guidance baked in (150 k steps). On held-out sentences and unseen voices it reaches WER 0.03 with both ASR systems, below the teacher, at a slightly lower naturalness score.

Everything needed to reproduce the training and evaluation is in the GitHub repository; the dataset pipeline is documented with the dataset.