This blog was originally posted in the OpenVoiceOS blog
Asturian (ast) and Aragonese (an) are Romance languages spoken in northern Spain by communities who get essentially no coverage from commercial voice technology. Speakers who want a voice assistant use Spanish, or go without. We built text-to-speech models for both using VITS, an open text-to-speech architecture, and released them as open weights on HuggingFace.
The constraint: Mozilla Common Voice 23.0 for both languages is multi-speaker and relatively small. Training a high-quality single-speaker TTS model directly on that data isn’t viable. So we used a hybrid approach.
The gap: no G2P
Neither Asturian nor Aragonese has a usable open phonemizer, and that won’t change with a future model upgrade. These models are trained directly on graphemes. A phonemizer would add IPA input at runtime, forcing pronunciation for names, neologisms, or code-switches, and would generally improve consistency. If you know of existing lexicons or pronunciation data for either language, or want to collaborate on one, get in touch.
The hybrid pipeline
The full methodology is in our Whitepaper on Hybrid TTS Dataset Synthesis. The short version:
-
Source data: Common Voice 23.0 (Asturian) and Common Voice 23.0 (Aragonese). Multi-speaker, varied recording conditions.
-
Filtering: audio normalized, silence trimmed, outliers by words-per-minute removed. This step matters: transcripts from rushed or unclear speakers undermine alignment.
-
Zero-shot revoicing: a voice-conversion model takes the filtered multi-speaker audio and a short donor voice clip, and re-synthesizes all utterances in the donor voice. The result is a coherent single-speaker dataset in a consistent voice, sized from the original multi-speaker source.
-
VITS training via phoonnx: VITS trains fast, runs without a GPU at inference, and the resulting models drop into any phoonnx-compatible pipeline.
Results
“L’arcu la vieya ye un fenómenu ópticu y meteorolóxicu que produz l’apaición d’un espectru de lluz continu nel cielu cuando los rayos del sol trespasen pequeñes partícules de mugor conteníes n’atmósfera terrestre. La forma ye la d’un arcu multicolor col roxo hacia la parte esterior y el viola hacia la interior. El arco iris duble, ye menos avezau a vese, y tien los colores invertíos, esto ye, el roxo hacia dientro y el viola hacia l’esterior.”
Download: Asturian, dii (female) and Asturian, miro (male)
“L’arco de sant Chuan ye un fenomeno optico y meteorolochico que produce l’aparición d’un espectro de luz contino en o cielo cuan os rayos d’o sol trescruzan chicotas particlas d’humidat situatas por l’atmosfera terrestre. A forma suya ye a d’un arco multicolor con o royo en a parti exterior y o morau en a interior. No ye tan cutiano l’arco de sant Chuan dople, que incluye un segundo arco mas tenue con as colors chiratas, ye dicir o royo en l’interior y o morau en l’exterior.”
Download: Aragonese, dii (female) and Aragonese, miro (male)
The pipeline validates. Pronunciation is imperfect: grapheme training without a phonemizer means the model is approximating phonetics from spelling patterns alone. What exists now is a usable baseline and a repeatable process.
Download the models and listen above. If you have Asturian or Aragonese pronunciation data, a G2P lexicon, or want to contribute voice recordings under an open license, contact us. Those resources directly improve the next iteration.