How Amazon Trained Alexa to Communicate with an Irish Accent

How Amazon Trained Alexa to Communicate with an Irish Accent

Amazon’s researchers faced the challenge of teaching Alexa to speak in an Irish brogue, a task that involved solving a long-standing issue in data science known as voice disentanglement. Similar to Henry Higgins in George Bernard Shaw’s play “Pygmalion,” data scientists Marius Cotescu and Georgi Tinchev worked on helping Alexa overcome pronunciation difficulties and adopt an Irish-accented English with the help of artificial intelligence and recordings from native speakers.

The process of voice disentanglement required more than just understanding vocabulary and syntax. It involved capturing a speaker’s pitch, timbre, and accent, which give words nuanced meanings and emotional weight. This aspect, called “prosody,” has been challenging for machines to master. However, recent advancements in AI, computer chips, and hardware have led to progress in this area, making computer-generated speech more pleasing to the ear.

The ultimate goal is to have voice assistants like Alexa, Siri, and Google Assistant capable of speaking multiple languages and various regional accents. Achieving this has been a costly and time-consuming process, involving the hiring of voice actors to record extensive speech data. Advanced AI systems known as “text-to-speech models” are starting to streamline this process by generating synthetic speech that sounds natural and can replicate different accents and dialects.

Amazon aimed to catch up with competitors like Microsoft and Google in the AI race, planning to make Alexa more proactive and conversational with sophisticated generative AI. Irish Alexa, which took nine months of training to understand and speak with an Irish accent, marked a significant step in this direction.

The challenges in building Irish Alexa lay in dealing with the specific linguistic features of Irish English, such as dropping the “h” in “th” and overpronouncing the “r.” The researchers relied heavily on an existing speech model of British-English accents and trained it to speak Irish English. This approach, using AI rather than voice actors, allowed Alexa to learn a new accent, but it required fine-tuning to correct pronunciation errors.

The positive feedback received for Irish Alexa’s accent has given the researchers hope that they can replicate accents in other languages as well, extending their methodology beyond English.