Google’s Euphonia project is an AI research initiative that seeks to equip its voice recognition systems with the ability to recognize the utterances of people with speech difficulties.
Today, in the framework of World Accessibility Awareness Day, the expansion of this research, which initially covered only English, to tests in French, Hindi, Japanese and Spanish, was announced.
Google wants its voice recognition system to be friendly to people who have speech difficulties
The Euphonia Project began in 2019 as a research initiative that seeks to make speech recognition technologies more accessible to people affected by speech impairments.
The algorithms that give life to any assistant controlled by voice commands are trained to recognize utterances emitted with normal diction, after having been trained precisely with vocal samples that meet this criteria.
For health reasons, many people cannot modulate easily. Voice samples of this type, until now, had not been considered in these developments, generating as a consequence the impossibility for the affected people to interact with the aforementioned virtual assistants.
The generation of a widely varied set of oral expressions, including samples of people with speech difficulties, is the purpose pursued by the Euphonia Project, to bring these technologies closer to more people. Since the start of this research, project volunteers have contributed more than 1,600 hours of speech samples, creating the largest known dataset of disordered speech in the world, according to affirm from Google.
These contributions have enabled Google’s Speech and Research teams to conduct cutting-edge research on machine learning applied to speech recognition. This includes the possibility of creating personalized models, capable of understanding people individually and speech-to-speech recognition, which allows the repetition of words with a clear synthesized voice.
“We are excited about the success of our research in English and hope to make similar progress in other languages, starting with French, Hindi, Japanese and Spanish, but we need your help. Through our work we have learned that the more speech samples we expose to our speech recognition models, the more people can potentially be understood.” Pan-Pan Jiang, technical manager of the program, commented through the company’s statement.
For the Spanish-speaking public, a interest form, through which people who have difficulty making themselves understood speaking are invited to participate in this project. The requested collaboration consists of the recording of phrases, which can be provided in basic sets of 30 phrases or in a complete set of approximately 1,300.
This initiative seeks to bring technology closer to a broader public, additionally providing practical resources for those people who face diction problems on a daily basis.
