Chatbots & Voice Assistants Meetup Jan – what3words, Speechmatics & BBC

The team headed over to the what3words HQ for the first meetup of 2020

John Campbell

Managing Director & Co founder

///month.spit.others …. OK, a strange start to a blog post but ///month.spit.others is the what3words address for the Rabbit & Pork office – it takes you to the front door of our building.

The first meetup of the Chatbots & Voice Assistants group was hosted by what3words. If you don’t know what they do, the video below explains. They split the globe up into 3×3 meter squares and gave each one an address using.three.words.

Tom Blaksley, Head of Voice Partnerships at what3words 

First up, Tom explained the use cases of what3words and voice. A simple example was a user asking for “Church Street in London”. It turns out there are 17+ Church Streets which is one of the reason why what3words works so well.

We already know that voice input in cars is a great use case when you can’t be looking or interacting with a voice device.

Tom went on to explain how, as the team developed the combination of words to be used for a language, they would need to conduct a number of checks, regardless of the language. First of all, checking if any words wouldn’t be appropriate to be used.  For example “turtle” in Tamil would be unlucky as having a “turtle” in your house can be associated with tsunamis.

A more voice-related aspect is that the teams would take out any homophones (words that are pronounced the same, e.g. bear, bare, etc.), which would cause problems with voice input. But most importantly, using three words for addresses actually reduces the number of words required. For example,  around 100,000 words are used in all of the standard addresses in Germany, whereas only 25,000 words would be required for the what3words platform – which helps with voice, as we would hear later on.

Tom explained then how what3words wanted to start to create their own voice api, so developers could send over a mp3 audio file and be returned with a what3words address. To do this, they partnered with Speechmatics. There is a good podcast over at VUX world which explain what problems what3words has with the built-in voice recognition and NLP with Google/Alexa.

James Page – Product Owner at Speechmatics

James picked up from Tom and explained how Speechmatics build custom voice recognition and NLP for companies, for a wide range of use cases, which can include “on the device” (no need to go to the cloud), and explained their approach to languages. One of the main differences between themselves and Amazon/Google is they build for a central language e.g. English, rather than a specific country variation e.g. South African English.

He explained that for the what3words product, they had a headstart as the combination of words to be understood had already been whittled down as part of the processes explained by Tom in his presentation.

More details over at https://what3words.com/news/navigation/voice-api-powered-by-speechmatics/

Tallulah Berry – Senior Journalist and News Producer with BBC Voice + AI + panel

Tallulah Berry from the BBC explained about a recent project in India with a Google Assistant action for the national elections.

Starting from 15 April the BBC will launch the first ever voice activated interactive bulletin in Hindi. This special election coverage will be available on smartphones and allow users to instruct their voice device to navigate through the news according to their interest.  To access it, just open Google Assistant on your phone or smart speaker and say: “Ok Google, Talk to BBC Elections or BBC Elections से बात कराओ”. The service will be updated each afternoon so users will be able to connect to BBC Election News coverage and get the latest on what is going on.

A quick insight from Tallulah was that it was essential to have a version in English and a version in Hindi to support both language speakers in the country.

The panel went on to cover their learnings from each of their projects and recommended approaches for working in multiple languages. The common thread was the requirement to work with local language experts in your projects. Not only from a translation point of view, but also getting those individuals to feedback on the design of your voice experience. The flow of your skill might require changes to break from the English model in order to work in different languages.

 

Big shoutout to the three speakers and to Tom Hewitson, Marton and Carl Aylett as ever for organising.

Head over to https://www.meetup.com/Messaging-Bots-London to find out about upcoming events.

John Campbell

Managing Director & Co founder

John founded Rabbit & Pork as the third agency and Voice Experience arm of TIPi Group. John has developed and launched several Google Actions and Skills, presented at several voice events around the UK, spoken on the VUX podcast and authored multiple whitepapers.