From the Aholab Group and the HiTZ center we present to the public the new speech recognition system in Basque. This technological advance has the potential to transform the interaction between people and technology, especially in the field of the Basque language.
The system has been trained with 548 hours of Basque voices from different public sources (Mozilla Common Voice 16.1, Basque Parliament, OpenSLR), which allows it to accurately recognize the words and phrases spoken by users, reaching quality levels of WER less than 5%.
Two different models have been created based on NVIDIA pre-trained models. One of them using a language model with more classic techniques, and the other using more emerging technologies such as transducers. The training of the models was carried out on the Hyperion system from the DIPC servers.
The system can potentially be integrated into virtual assistants to perform tasks such as sending messages, searching for information or setting reminders. It could also enable the automation of responses to telephone calls, improving efficiency and customer service. And it will facilitate the transcription of audio recordings in Basque.
A demo of the speech recognition system in Basque is available at this link and the models are available at Gaitu-Data. The team invites the community to use it and provide feedback to continue improving the technology. We hope that it will be a valuable tool for the Basque community and contribute to the strengthening of our language.