Data scientists describe the arduous, AI-powered task of getting Amazon’s Alexa to speak in an Irish lilt

• Two scientists at Amazon used artificial intelligence to help them make Alexa Irish. 
• The team trained Alexa on accents generated by text-to-speech models. 
• This then helped them identify distinctive aspects of the Irish accent to train Irish Alexa.

Two data scientists who created an Irish accent for Amazon’s Alexa say that text-to-speech systems powered by artificial intelligence shifted how the team approaches the creation of different voice patterns and accents, according to a new feature from The New York Times.

The scientists use of text-to-speech is just one example of how the technology — in research, commercial, and personal settings — has exploded over the past year due to advances that have helped make these generated voices sound eerily similar to that of humans. Users online are now using AI audio to narrate audiobooks, scam people out of money, and create voiceovers for advertisements.

Georgi Tinchev and Marius Cotescu, who described the Irish brogue to the Times as “a hard one” to master, were tasked with creating the accent for Amazon using a technique known as voice disentanglement. 

Voice disentanglement is the pinpointing of aspects of speech like tone and intonation. Last year, Amazon scientists began to promote the technique as a way to bring more diversity and inclusion in voice assistants.

In creating an Alexa that speaks Irish English, released in November, data scientists told the Times that hours of language training using the help of voice actors had been condensed because of AI, which now has the ability to mimic regional dialects and accents in numerous languages. This would be used to help disentangle — or identify — certain distinctive aspects of the accent.

Using an already existing Alexa model that speaks in a British accent as a baseline, the data scientists trained it on the Irish accents provided by text-to-speech models as well as voice recordings, the Times reported. 

They later zeroed in and modified certain sounds — for example, emphasizing the “r” sounds in words much like an Irish person would — to create the voice that eventually accompanied Irish Alexa.