Great strides in developing Finnish speech recognition
This video is a part of the FCAI success stories series. In the video series, we explain why fundamental research in AI is needed, and how research results create solutions to the needs of people, society and companies.
Speech technology is an integral part of our daily lives, but current language technologies were developed English first. Here, the approach is to first collect a comprehensive set of words, then learn how these words sound and behave with other words.
Languages such as Finnish, however, are structurally very different from English: words are often inflected, derived and combined to produce countless different forms.
“That's why it was necessary to develop our own approach for speech recognition, machine translation and other language technology tools to cover languages like Finnish”, says Mikko Kurimo, Associate Professor of Speech and Language Processing at Aalto University.
Data as bottleneck
Training statistical language models requires lots of speech and text data, and this is often not available. One solution is to use machine learning to adapt out of domain data from the same language, or extract and adapt information from larger related languages.
“We need all the time even more data, because the more accurate models are bigger and the novel training algorithms more data-hungry”, says Kurimo.
Kurimo leads the Aalto Speech Recognition Research Group which has developed speech recognizers and language models for Finnish, Swedish, Estonian and even Arabic and Northern Sami.
The models can be used not only in speech recognition, but also for translation, synthesis, adaptation and language generation in various AI tasks that involve language technology.
“It is very important to develop such AI applications in Finnish and other small languages because it helps to ensure that Finnish customers get customer service and speech-based applications in their own language also in the future”, says Kurimo.
Several companies use technology from the Aalto Speech Recognition Group
The language technology developed at Aalto University has been applied by several companies in Finland that operate with large amounts of speech and text data.
For example, with the help of Aalto's researchers, Finnish telecommunications company Elisa has been able to improve their Finnish speech recognition and analysis of conversations.
“As a consequence, their customer service is faster, as people can solve more issues with AI-based machines, instead of waiting for a long time in a phone queue” says Professor Kurimo.
Also other Finnish companies such as Lingsoft, Inscripta, Devoca, Speechly, Ääni Company and even Yle have benefitted from the open source speech and language modeling tools developed at Aalto.
The most significant ongoing research projects at the Aalto Speech Recognition Group include:
Studying methods for an efficient re-use and re-purpose of multilingual audiovisual content to revolutionize video management and digital storytelling in broadcasting and media production.
Adapting the methods to analyse audiovisual content presented in hundred years of Finnish fiction films 1907-2017 to illustrate the development of the Finnish society.
Developing computer-aided oral language testing and self-practising tools to enhance second language learning.
Collecting conversational speech spoken in Finland for studying the language and helping to develop AI in Finnish and Swedish in Finland in connection to Yle's Lahjoita Puhetta campaign.