Language Bank, language technology and donated speech

Donate Speech is a project by Yle (the Finnish Public Service Media Company), the Finnish Climate Fund (formerly the Finnish State Development Company Vake), Solita and the University of Helsinki, collecting free-form speech in Finnish for the development of AI and speech recognition. The aim of the project is to collect 10,000 hours of ordinary, everyday speech. The project has paid particular attention to the possibility of using the material for both scientific research and commercial purposes.

Donating speech has been made easy with mobile applications. The result is speech-controlled devices and services operating smoothly in Finnish. However, training artificial intelligence requires a lot of data representing different dialects and speech from people of different ages and speech styles, such as Finnish as a foreign language. The data also has to be easily accessible to researchers and service developers.

The material collected in the Donate Speech campaign is stored in the Language Bank where it can be offered to companies, researchers, higher education institutions and research institutes.

LUMI and corporate use

One of the new features brought by the LUMI supercomputer is the corporate use of CSC’s computing resources. The distribution of resources takes place through Business Finland, and even small growth companies get an equal footing with international giants, at least in terms of computing resources. This kind of corporate use is already being piloted.

One of the first pilots through Business Finland’s AI calculation support is Speechly, which specializes in speech interfaces. For nine months, Speechly’s speech recognition technology based on deep neural networks was developed with CSC’s Puhti-AI supercomputer and Speechly’s data.

CSC’s resources provide early-stage growth companies such as Speechly with computing resources that only the world’s most well-known technology brands, such as Google, Apple and Amazon, have previously had access to. During the project, Speechly managed to improve the recognition accuracy of its speech recognition model by almost 60% measured by word error rate.

– Speechly’s product development work involves testing different machine learning algorithms to find the best speech recognition models. Modern machine learning models for deep neural networks require high computing capacity. How quickly we can train and test one model is one of the most important factors affecting product development results. Training one model can take anything from days to weeks. Thanks to CSC’s computing capacity, we were able to work much faster than before and with significantly larger data sets, explains Speechly’s leading researcher and co-founder Janne Pylkkönen in a Business Finland press release.

In the future, we can expect a lot from this combination: there will be a large amount of high-quality speech data and heavy computing by supercomputers available for university researchers, research institutes and businesses, and it will be easier to implement various joint projects.

The Language Bank of Finland is a service package built in cooperation by Finnish universities and research institutes, coordinated by the University of Helsinki and carried out with CSC’s technical implementation. Together, these make up the FIN-CLARIN consortium, which in turn represents Finland in the European CLARIN infrastructure. The Language Bank contains text and speech materials for diverse languages and a range of tools for researching them. In addition to language researchers, it serves a wide range of digital human sciences researchers and students.

• Donate Speech project (in Finnish)
• Grand One digital media gala(in Finnish)
• Business Finland: AI computing grant accelerates product development

Tommi Kutilainen
The author has a few decades of experience in scicomm at CSC and he is especially interrested in everything. Twitter: @TommiKutilainen

Tero Aalto
The author is a language technologist and works with the Language Bank of Finland.