Successful cases
TTS in Basque for Tokimedia:
We developed a new text-to-speech (TTS) system in Basque for the Tokimedia digital network. It allows any digital text to be converted into podcast format, sounds more like the writing style of a journalist, and enables users to listen to the texts. The project provided two main innovations: the ability to create custom voices (requiring few recordings) and greater expressiveness when reading news articles. This was achieved by combining language models (BERT, POS tagging) with the TTS model.
TTS in Basque for Tokimedia:
Overview
We developed a new text-to-speech (TTS) system in Basque for the Tokimedia digital network. It allows any digital text to be converted into podcast format, sounds more like the writing style of a journalist, and enables users to listen to the texts. The project provided two main innovations: the ability to create custom voices (requiring few recordings) and greater expressiveness when reading news articles. This was achieved by combining language models (BERT, POS tagging) with the TTS model.
Challenge
The main challenge was to develop a neural network that could read any text with the proper intonation and be trained with only a few sentences. It also meets the challenges inherent in journalistic texts (long sentences, numerous entities), while allowing specific voices to be chosen for each entity.
Collaboration
The project was developed in collaboration with media group TOKIKOM (Tokimedia network).
Result
A total of 150 sentences per speaker were compiled and a multi-speaker model, FastPitch + HiFi-GAN, was trained on these recordings. To improve expressiveness, versions incorporating POS tagging and BERT vectors were also trained and evaluated. However, as the original FastPitch model obtained the highest quality results, this was the one that went into production. It is currently running on Orai servers, and TOKIKOM uses it as a service via an API.
