First open Basque speech language model built on LLaMA-Omni
Original title: Developing a Speech Language Model for Basque
Fine-tuning a Latxa backbone on translated, synthesized VoiceAssistant-400K data scored 3.86/5 content quality and 9.94% WER.
Researchers present what they call the first open speech language model for Basque, aimed at low-latency end-to-end spoken interaction. It uses the LLaMA-Omni architecture with a Latxa backbone and fine-tunes several modules to carry Basque text ability into speech. To offset scarce data, the English VoiceAssistant-400K corpus was translated into Basque and the audio was synthesized with a Basque text-to-speech system. Reported scores are 3.86/5 for content quality, 9.94% word error rate for spoken intelligibility, and 3.07 UTMOS for audio quality.
Unconfirmed: The claim that this is the first open Basque speech language model is the authors' own.
- Architecture is LLaMA-Omni with a Latxa backbone.
- Training audio comes from translating VoiceAssistant-400K and synthesizing it with Basque TTS.
- Content quality is 3.86/5, WER is 9.94%, and UTMOS is 3.07.
- Authors say modular design plus synthetic data can support low-resource spoken systems.
From the original article
1 min read at arxiv.orgAlthough the recent development of LLMs has paved the way for multi-modal interaction, minority languages face difficulties keeping pace. This work presents the first open Speech Language Model for Basque, designed for low-latency, end-to-end spoken interaction.
By Andoni Sudupe · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.