Google Crowdsourcing Accents to Improve Voice Recognition

google speech recognition voice commandGoogle has turned to social new aggregator Reddit in order to find accents that it can use to test its Google Now voice recognition software.

Google Now, which uses a combination of machine learning and crowdsourced data to improve its ability to understand natural language, is already fairly efficient at cutting through murmurs, mumbling and other tics of language. Last year, Google ran a project involving Google Voice and Project Fi to cut transcription errors by 49 per cent.

However, accents have remained a sticking point for the software, with Google Now only recently gaining official support for Indian and Australian accents. In an effort to improve quality for these and other accents, Google has decided to recruit help.

The tech giant has hired a third-party firm, Appen, to source users with accents, which in turn has begun hiring Reddit users, posting gig listings on various subreddits (the individual communities within Reddits ecosystem) in order to find them.

The job listings are split fairly evenly between users who are searching for short-term, part-time work on subreddits like WorkOnline and Beermoney, and those who live in cities and areas with distinctive accents, like Edinburgh. All the listings state that they are after users with particular accents.

According to one listing, users must own an Android phone, and are required to record a series of voice prompts and phrases, each around three to five seconds long, for around three hours (totalling around 2,000 phrases). Adults earn $36 (£27), with children under 16 working from a smaller list for $26.

Accents that Google is attempting to capture include a variety of local intonations from around England and the US, as well as Indian and Chinese-accented English, but the most tricky accents appear to be Scottish dialects, which frequently trip up voice assistants from Google Now to Apples Siri.

The phrases vary from simple commands to brand names, idioms and trivia questions, with the voice data collected and processed by Appen before it is passed Google. While the machine learning tools require massive amounts of data to boost recognition, this word-farming approach will hopefully push Google Nows ability to interact with natural language higher than ever before.

Got a piece of work you want to show off to the world? Enter the Effective Mobile Marketing Awards. The final deadline for submissions is this Friday, 19 August. Get the full details here.