Date: Wednesday, November 29th, 2017
Speechmatics put Automatic Linguist, their AI framework, to the test during Phase 1 of Project Omniglot to learn as many languages as possible
Speechmatics, a leading provider of automatic speech recognition technology, has completed Project Omniglot, the team’s innovation challenge. It stress-tests the recently launched unique AI framework, the Automatic Linguist (AL), and paves the way for the wider goal of building every language in the world. Having successfully built 46 new speech-to-text languages in just 6 weeks, the company has taken their total number to an industry-leading 72. While other solutions, including Google, focus their efforts on dialects rather than unique languages, Speechmatics have increased the scope to cater for a wider range of languages.
Traditionally, building a new language pack takes months and is a laborious affair, meaning only the most widely spoken languages in the world remain the focus. The challenge was to see how many languages AL could build in just 6 weeks. It exceeded the team’s expectations by learning a language a day. Due to the speed at which the languages were produced automatically, Speechmatics are offering for people to use them for free initially.
This initial phase of Project Omniglot has proven that the machine learning framework behind AL works extremely well. It can automatically learn the sounds (phonemes) of a language as well as the grammar and semantics in order to determine which sentences make sense. Speech-to-text technology is one of the most widely discussed topics right now and, as the world is becoming increasingly more connected, broad language coverage is also becoming essential. From broadcast subtitling and interview transcription to accessibility within the education sector, Speechmatics is hoping to open the door to speech-enabled future in as many languages as possible, for more countries than ever before.
Benedikt von Thüngen, CEO of Speechmatics said: “We are already seeing a shift to a speech-enabled future where voice is the primary form of communication. Transcription not only eases the lives of many people, but opens the door for new opportunities, especially in regions with lower literacy rates. As a company, we have focused on accuracy and deployability, whereas through AL and Project Omniglot we focused on increasing our language coverage. As a team we are very humbled and impressed by the results we have achieved and are excited by the potential opportunities we will now enable. We would love for people to try the new languages, give us feedback and see how we can develop them from there.”
Tom Ash, Speech Recognition Director at Speechmatics and recent winner of the ‘Speech Luminary’ award, commented: “The framework we’ve created works completely autonomously and so efficiently, that we can experiment with languages that would otherwise be uneconomic to build. Also, it now gives us the ability to iterate rapidly on any given language and improve them at an unprecedented pace. We have built ASR for languages spoken by communities in the Philippines, India or central Asia, all of which are often overlooked. There are over 7,000 languages in the world, and our ultimate goal is to make speech recognition technology available to as many as possible.”
• Automatic Linguist (AL) – a unique AI framework that learns a new language with minimal data in a short period of time
• Project Omniglot – Speechmatics’ goal to build as many – if not all – languages in the world
• Phase 1 of Project Omniglot – initial stress test of AL to build as many languages as possible in 6 weeks