Sign up for our newsletter ⟩
News |

Language-based AI leaves Finland behind--can a new initiative change that?

Finland launches a project to collect speech samples from all Finnish speakers to develop voice-recognition technology.

Nainen tuijottaa älypuhelintaan.
To improve Finnsh voice-recognition services, "Donate Your Speech" wants to collect Finnish from speakers in and outside of the country, from those speaking it as a second language as well as from people with speech impediments. Image: Tiina Jutila / Yle

Using voice-activated commands to make phone calls, dictate shopping lists or dim the lights is becoming increasingly common in many countries, but Finland has fallen behind on anything related to voice-activated technology. The country doesn’t have a smart speaker market, for example, and this is partly because of the limited selection of Finnish speech available for artificial intelligence (AI) to work with.

It's a common issue for languages with fewer speakers. The EU Commission has expressed concern that natural language technology built primarily for English speakers is leaving other European languages under-represented.

If English speakers can casually ask their map service to direct them to a certain location, speakers of less common languages should be able to do something similar--but the market is so small that big tech firms are not investing in providing those services.

Now Finland is trying to do something about it.

To help establish a foundation for voice-activated services in Finnish, broadcaster Yle, Helsinki University and state development company VAKE launched "Donate Your Speech" (Lahjoita puhetta in Finnish), a publicly funded venture.

The project aims to draw speech samples from a large and diverse pool of Finnish speakers, including those learning it as a foreign language.

Finnish from foreigners

"As more and more people speak Finnish as a foreign language, it’s important for speech recognition AI to learn to detect non-native accents," Katja Solla, a journalist working with the project at Yle, explained.

With some six million speakers worldwide, Finnish ranks 160th among the world’s 7,000 spoken languages. While not an endangered language, according to Lotta Jalava, a Finno-Ugric specialist at the Institute for the Languages of Finland, the role of English in academic circles in Finland is growing.

The city of Helsinki predicts that foreign-language-speaking people in the capital area will climb from the current level of 14 percent to around 25 percent by 2035. This means any future voice recognition services in Finnish will need to be able to deal with a wide range of accents.

"Getting non-native Finnish speakers to share their speech is really crucial--especially when you consider how well voice recognition systems understand heavily accented English," said Pia Erkinheimo of VAKE.

Volunteers are asked to speak naturally and not read anything out loud.

"People pause, stumble and mumble a lot more than they think," Erkinheimo added. "And it's important AI learns to recognise that."

A common interest

The idea is that the more speech people donate, the more material companies and linguists have for developing Finnish-language voice recognition services.

"Our goal is to record 10,000 hours of human speech, and so far we’ve got about 2,000 hours," Solla explained. "This is a project in the common interest--no one else is doing this."

The Language Bank of Finland, an entity within Helsinki University, will manage the accumulated speech. For a nominal fee, the material will be available to anyone, provided they can prove compliance with EU GDPR privacy rules.

Major tech companies welcomed the initiative, according to Erkinheimo, who said she doesn’t see an ethical dilemma in foreign companies accessing volunteers' speech.

"We don’t think of this as large Chinese or US companies benefiting from the project. The point is that we are all in this same ecosystem--including many small Finnish companies--so everyone’s benefiting from it," Erkinheimo said.

"We’re open about it"

Not everyone is so relaxed about the privacy issues around voice recognition.

American tech giants Apple and Google have faced criticism for eavesdropping on customers to develop their own speech recognition technologies. This includes storing audio command files of users talking to their virtual assistants.

In contrast, "Donate Your Speech" is a voluntary and transparent initiative for AI speech collection, according to Sami Köykkä, a consultant with tech firm Solita which developed the app for the project.

"We are doing this in a way that’s not scary. We are telling people what we’re doing and they’re giving us this information voluntarily," he said.

Köykkä said applications for voice-enabled services in Finnish could range from automatic subtitling of TV shows to medical transcription.

Privacy concerns

"I don’t see any problems with this from a security standpoint," said Leena Romppainen, who chairs data privacy hawk Electronic Frontier Finland (Effi).

But ethical implications regarding this type of data can surface further down the road, according to Dave Sayers, a lecturer at Jyväskylä University who researches how emerging technologies are shaping language.

"There’s a lot of ways in which innocently released data can be misused--a risk that this kind of information can fall into the wrong hands and be manipulated to create deep fakes," Sayers said, referring to the possibility of fraudsters fabricating audio samples to yield speech appearing to be authentic.

"In the foreseeable future, AI could create recordings used to call up the bank or phone a relative asking for money," he said of the possibility of AI mimicking someone's voice, a field known as automated voice mimicry. "It’s not difficult for computers to match up anonymous audio samples with real people if their voices are available on the internet."

While AI scams are still far from an everyday occurrence, last year thieves used voice-mimicking software to impersonate a chief executive’s voice to make a fraudulent transfer of 220,000 euros.

AI companies have also shown how small snippets of audio can not only mimic the speech of a real person but also display shifts in emotional cadence, such as in these clips of US presidents Obama and Trump.

Android-puhelimissa on nykyisin puheenohjaus. Puhelinta voi pyytää soittamaan vaikkapa äidille. Kaikki puheentunnistukselle puhutut jutut päätyvät kuitenkin tallenteina Googlen tietokantaan.
Speech recognition technology is becoming increasingly embedded in our everyday lives. Image: Tuomo Björksten / Yle

"It isn’t until someone misuses data that people wake up and ask, 'isn’t anybody watching over this information,'" said Jyrki Kasvi, who became known as a digital-savvy politician and now works with digital advocacy non-profit TIEKE, the Finnish Information Society Development Center.

Kasvi called on Finland to establish a new agency to oversee matters related to AI as this is an area beyond the scope of Finland’s Data Protection Ombudsman. He said the current rush among states to develop AI capabilities was akin to countries competing to build factories during the industrial revolution. "No one stopped to consider byproducts like pollution."

Aleksi Rossi, who’s been heading up "Donate Your Speech" for Yle, agreed that the potential for abuse exists as new technology emerges in the realms of face and voice recognition.

"If someone really wanted to manipulate your voice for criminal purposes they could already do so without your sample," Rossi explained. "Bringing the threat of misuse into the world actually makes things safer because it prompts the development of new safeguards."

Sanakirjan sivu
Finnish has some six million speakers globally, including diasporas in Sweden and North America. Image: Ville Tiirikainen / Yle

Rossi said the "Donate Your Speech" data will be treated with utmost caution and interested parties wanting to access the material will face multiple levels of scrutiny.

Every five years the programme will re-evaluate whether to discard or maintain the collected data. Volunteers can also request their voices get deleted from the system.

"As a language, we stand to gain so immensely from this project that its importance outweighs the risks which we are very conscious of and working against," Rossi said.

Yle and Helsinki University’s language bank have provided their time and resources to the project, with the brunt of the 100,000-euro budget used to pay technology partner Solita. Launched in June, "Donate Your Speech" will continue through the end of the year, with possible add-ons for Swedish and Sami included at a later stage.

For more information and to participate, head to lahjoitapuhetta.fi (in Finnish).

Latest in: News

Headlines

Our picks

Latest

Muualla Yle.fi:ssä