Canberra’s ASR gurus are hyper-tuning existing ASR models and teaching them to behave in a certain way that is customised for each client. Photo: Trellis Data.
A Canberra-based artificial intelligence (AI) company is optimising automatic speech recognition (ASR) technology for law enforcement, defence and government organisations the world over.
Trellis Data is a global leader in improving the accuracy of speech-to-text platforms. Its bespoke product, Trellis Secure Speech, also makes the technology accessible across a variety of languages.
Trellis Data Head of Communications Tim McLaren said when it comes to using ASR in a courtroom or police interview, the accuracy of the transcript is of utmost importance.
“We’re working with big players in Australia, South East Asia and Europe who demand really high accuracy from their transcriptions,” he said.
“Some existing platforms are about 80 per cent accurate, which is not bad, but our team works with clients to optimise the technology for their organisation and aim for an average accuracy of 95 per cent, which can be a significant improvement.”
Trellis Data AI and machine learning specialists are at the cutting edge when it comes to improving speech-to-text models.
“Knowing a translation from voice-to-text is close to accurate can save the client time when they don’t have to go back and check things quite so closely,” Tim said.
“Some of our clients record hundreds of hours of audio per day, so we are making ASR as efficient as possible for them.”
It takes humans about three to four hours to manually transcribe one hour of audio with 99 per cent accuracy. In contrast, Trellis Secure Chat can do it in real-time with 95 per cent accuracy, which is better than some of the best-known AI models, including Open AI (makers of Chat GPT).
Trellis Data IT specialists are creating and optimising machine learning and speech-to-text models, and Tim said they’re trying to “build our product to be as transparent and trusted as possible”.
They’re fine-tuning their software to recognise and transcribe languages, dialects and even strong accents.
“We’re working on ways to create, enhance, refine and improve the accuracy of ASR models without the need for hundreds and hundreds of hours of audio,” Tim explained.
“This will make them more accessible and reliable for people, which is a really big deal.”
The Trellis Intelligence Platform (TIP) has been built to improve accuracy across a variety of languages and also supports a universal language translator, which allows people to have a conversation in different languages.
“This sort of technology can be very useful if you’re talking to someone from another country because it will translate German to English, for example, in almost real-time,” Tim said.
“Having that immediate translation is a game-changer for many of our clients, particularly in the defence and law enforcement sectors.
“We’re closing the gap on AI when it comes to language.”
Law enforcement agencies are some of Trellis Data’s’ biggest customers and they’ve created a dynamic speaker identification feature that is clever enough to recognise if there are two or three people in a room and automatically assign segments of the conversation to each person.
“Not only can we assign speech to different speakers dynamically, we can also identify the language being spoken,” Tim said.
“The goal is to make it as easy as possible for our customers to accurately capture and understand the conversations in real time.
“This is a huge leap for AI, but of course, accuracy and reliability are the key, which is where we’re seeing incredible results.”
Trellis Data delivers customised, easy-to-use machine learning and AI technology solutions to business and government customers across a range of sectors in Canberra and around the world.
Original Article published by Katrina Condie on Riotact.
