itzune/pocket-tts-basque-24l

Pocket TTS Basque — 24-layer model (euskarazko ahots-sintesia)

A Basque (euskara) text-to-speech model based on Pocket TTS by Kyutai. This is the 24-layer "teacher" model (~340 M parameters), obtained by fine-tuning Kyutai's 24-layer English model for Basque with the official new-language recipe, on the BSSD speech dataset. It supports zero-shot voice cloning from a short reference clip.

For real-time use on CPU, see the lighter 6-layer model distilled from this one: itzune/pocket-tts-basque.

Usage

With uv installed, no setup is needed:

uvx pocket-tts generate \
  --config hf://itzune/pocket-tts-basque-24l/basque_24l.yaml \
  --voice my_voice.wav \
  --text "Kaixo! Hau euskarazko ahots sintetiko bat da." \
  --output kaixo.wav

Python:

from pocket_tts import TTSModel
import soundfile as sf

model = TTSModel.load_model(config="hf://itzune/pocket-tts-basque-24l/basque_24l.yaml")
voice = model.get_state_for_audio_prompt("my_voice.wav")      # 5-10 s of clean Basque speech
audio = model.generate_audio(voice, "Kaixo! Hau euskarazko ahots sintetiko bat da.")
sf.write("kaixo.wav", audio.squeeze().numpy(), model.sample_rate)  # 24 kHz

Notes:

• Write the text as it should be read: numbers, dates, abbreviations and symbols must be spelled out in words (e.g. hogeita lau rather than 24). The dataset's normaliser (bssd.text_norm) does this for standard Basque.
• The reference clip defines the voice. Only clone voices you have the right to use.
• --config also accepts a local copy of basque_24l.yaml; the weights and tokenizer are fetched from this repository.

Evaluation

Held-out BSSD test sentences, 40 sentences per voice, scored with two independent Basque ASR systems (xezpeleta/whisper-large-v3-eu and HiTZ/stt_eu_conformer_transducer_large_v2), UTMOS naturalness and WavLM speaker similarity to the reference clip.

voices	WER (Whisper)	WER (Conformer)	failure rate	WER excl. failures	UTMOS	speaker sim.
8 named prompts (dataset default voices)	0.095	0.091	5.3 %	0.042	3.59	0.958
5 unseen voices	0.073	0.069	3.5 %	0.033	3.61	0.961

The median sentence is read with no errors. "Failures" are sentences with WER > 0.5: in nearly all of them the model continues the reference clip's own speech instead of reading the input text, a known failure mode of prompt-conditioned models. It occurs more often with some reference clips than others (from 0 % to about 20 % per voice in our tests); if a voice misbehaves, try a different or shorter clip.

Training

• Base: Kyutai's 24-layer English Pocket TTS model (text embedding re-initialised for the new tokenizer).
• Tokenizer: SentencePiece, 4000 tokens, trained on the BSSD transcripts.
• Data: BSSD train split (about 350 k utterances, 450 h of synthetic Basque speech, 2,400+ voices, word-level alignments from deepdml/wav2vec2-large-xls-r-300m-basque).
• Recipe: official finetune_language configuration (learning rate 2e-4, effective batch 64, classifier-free-guidance dropout as upstream); the released weights are the EMA weights at step 40 k, chosen among the 40 k-80 k checkpoints by the evaluation above (later checkpoints had higher failure rates on unseen voices).

Training code, configs, evaluation scripts and a demo page with audio samples and a comparison with Piper and OmniVoice: https://github.com/itzune/pocket-tts-basque (demo: https://itzune.eus/pocket-tts-basque/)

Limitations

• Basque only; code-switched Spanish/French/English words are read with Basque phonetics.
• Prompt-continuation failures as described above; always check generated audio.
• Trained on synthetic speech: the voices in the dataset are themselves synthetic or re-synthesised, so the model may inherit artefacts of the dataset's teacher system. See the BSSD dataset card for the provenance of the voices and texts.

License and attribution

Model weights: CC-BY-4.0 (as the upstream Pocket TTS weights). Please cite Kyutai's Pocket TTS and the BSSD dataset. Dataset: itzune/bssd (CC-BY-SA-4.0).