Basque

TL;DR I’ve created an experiment called NongoEuskara. To do this, I gathered data, trained a fastText classification/embedding model, and built this little demo website. Using WebAssembly, these models run directly in the browser. The long explanation We know that large language models (LLMs) are capable of doing pretty much everything. Even though they weren’t originally designed for it, today they can program, translate, transcribe audio, summarize long texts, and a long etcetera. Sometimes, however, the best solution isn’t always using an LLM. For example, for quick and simple translations, there are lighter neural translators that require far fewer resources than LLMs. ...

The Latxa model was an important milestone for artificial intelligence in the Basque language. Until then, we had no open-weight model that handled Basque decently. Since then, several others have been created specifically fine-tuned for Basque: Kimu (Orai) and the Latxa Qwen3 VL family of models recently released by HiTZ. At the same time, new top-tier local models have been emerging — Qwen, Gemma, etc. — which can be especially interesting for use in agents with external tools. These models are also improving their Basque language capabilities noticeably. But do they match the Latxa model? ...

In recent years, text-to-speech (TTS) models have advanced enormously. Month after month, new releases improved pronunciation, prosody, and overall audio quality. Unfortunately, most of those models were designed for English, or at best compatible with only a few languages. The tools available for generating Basque voices used to be quite limited: paid proprietary systems (such as Elhuyar’s neural TTS, developed by Orai), or older robotic voices that had clearly fallen behind. ...

Over the last few days, I have been fine-tuning Nvidia Parakeet for Basque, and I have published the result here. The goal was simple: to have a lightweight Basque speech-to-text model that runs fast and is practical on modest hardware. Accuracy vs speed To be clear, this model is not as accurate as my best Basque Whisper model: xezpeleta/whisper-large-v3-eu However, it has a major advantage: it is very fast and can run on CPU-only setups. ...

Orai has just released a language model called Kimu. Based on Gemma 2, they have created models with 2B and 9B parameters. They have successfully injected the knowledge required to speak and understand Basque into the base model. I have converted both Kimu models to the GGUF format and published them on Hugging Face. This makes it possible to use the Kimu model with applications like Llama.cpp or Ollama. ...