A Sound Idea
Voice for Purpose was born from an intuition of the Italian actor, voice actor and TV presenter, Pino Insegno, who wanted to donate his voice to people who could not speak due to Motor Neuron disease (ALS- Amyotrophic Lateral Sclerosis). The condition causes a progressive loss of control of the movements, including those involved in speech. It is estimated that approximately 32,000 people are affected by ALS in Europe, and most of those affected will lose the ability to speak, swallow, walk, and breathe.
Translated, a company specialised in languages and AI-based communication services, picked up the challenge in 2022 and created the Voice for Purpose platform to allow anyone to donate their voice to non-verbal people and to deliver expressive, personalised voices. In the past, the technologies that provided such applications for voice loss sounded robotic rather than human. Voice loss is not just a communication problem it also has a serious impact on the individual's emotional well-being and sense of self. Voice for Purpose grew from the belief that people would prefer to speak with a real human-sounding voice that was close to their own in terms of gender, language, identity, and tone. It would allow them to express how they felt in a voice that matched the emotional register they wanted to communicate in and not in the flat, emotionless delivery of a synthetic voice.
Voice for Purpose invites people to donate their voice to help people who suffer from voice loss. Working with Translated, the project has now developed into a fully functional service with a community of 6000+ donors in 5 languages. It provides voice support services to hundreds of people with the possibility to assist across a range of medical conditions including stroke, autism, SMA (Spinal Muscular Atrophy), and cerebral palsy.
Supercomputing Solutions
After two years of operations, the project identified key areas where new, specific solutions would have to be developed to better serve people with speech and movement impairment. One issue highlighted was that the input of text using eye-gaze technologies was very slow. Creating personalised expressive voices for people who have limited control of their larynx and vocal cords was more challenging because the sample size was very small. As so many people in Europe speak more than one language, the team also wanted to work on speech models that could be natively multilingual.
For this, the researchers at Translated turned to larger models for text prediction and speech synthesis. They were delighted to discover that EuroHPC JU could provide access to the infrastructure they needed to host the size of experiments they wanted to perform. The team were awarded access to the EuroHPC supercompter Leonardo and from this allocation they selected and trained two Large Language Models (8B and 70B) to predict sentences from word initials in the 24 official European languages. They also trained a zero-shot text to speech model, which can leverage even small snippets of reference data (in this case open conversational data), to produce multilingual audio.
The research team delivered very encouraging results in LLM prediction. They were able to match previous work performed by Google on the English language and extend it to 23 more languages with comparable performance, using a model that was 8 times smaller.
The aim, once in production, is to double or triple the text input speed, a great improvement for movement-impaired people who use eye-gazing to slowly type text to generate speech. In relation to text-to-speech (TTS) the researchers proved that with focused collection and selection of open data sources it is possible to train from scratch multilingual voice models that can deliver speech that faithfully reproduces human voices in 24 languages.
Based on these positive results, the team are planning new experiments that utilise EuroHPC JU supercomputers to refine the models to the stage of production ready. Translated are also looking to work with the new AI Factories resources made available to European startups and SME and would encourage other companies to engage with the access and services provided by EuroHPC JU. For now, they continue their excellent work to give a voice to those suffering from degenerative diseases like ALS.