Martin Carsten Nielsen, Co-Founder and CEO of Alvenir. Photo: Hanne Kokkegård, DTU Compute

Software from DTU spinout recognizes even small languages

Monday 01 Aug 22

Contact

Rasmus Stig Beck Jensen
Special Consultant
DTU Compute

Languages in Europe - EU

Alvernir wants to ensure that Europe's small languages also live digitally in the future. It aligns well with the EU, which translates most of its documents into the 24 official languages of the EU. One of the fundamental principles of the EU is multilingualism.

The purpose of multilingualism is to:

  • communicate with citizens in their own language
  • protect Europe's rich linguistic diversity
  • promote language learning in Europe

Source: European Union

Alvenir believes that the computer technology of the future should be able to understand everyone. In many instances the company's speech recognition software is better at understanding Danish than that of the giant corporations.

When I hurried through the traffic the other day, I asked my smartphone's digital assistant to locate the opening hours of the library. Although I tried to speak very clearly, the assistant did not understand me as is often the case.

In general, digital assistants find it difficult to  understand Danish, because the language models of large IT companies are not sufficiently specialized that they can properly understand  a language as small as Danish. The applies to many other smaller languages.

This is where the Danish startup Alvenir can help Danish language speaker and later on speakers of other smaller languages. The company specializes in making industry-specific models, which basically get it right about 90 percent of the time and up to 100 percent if it is adapted to a company's technical terms and product names.

“We all have the right to be understood by the computer technology of the future. Whether we are old, mumble, or speak one of the small languages with a particular accent. Alvenir's goal is to democratize the whole of Europe. There are an incredible number of small languages in Europe, and for many of them, there is a big task in creating good general models that can act as a foundation for the industrial specialization we focus on in Danish. We are underway in Denmark, by the end of 2024 it is the plan that we will scale to the Nordic region, and in 10 years we will be heavily involved as a listed company in the European market,” says CEO and co-founder of Alvenir Martin Carsten Nielsen.

"We all have the right to be understood by the computer technology of the future. Whether we are old, mumble, or speak one of the small languages with a particular accent. Alvenir's goal is to democratize the whole of Europe."
CEO and Co-Founder of Alvenir Martin Carsten Nielsen

Alvenir is based on knowledge from a DTU Compute thesis, where Martin Carsten Nielsen and Rasmus Arpe Fogh Egebæk under the leadership of Professor Lars Kai Hansen in the research section Cognitive Systems have developed AI-based models for speech recognition (Danspeech). With the help of DTU's supporting programme for entrepreneurship, the two entrepreneurs have shaped their business idea. On 1 April 2022, DTU signed an agreement on technology transfer, which has been paid for in an ownership model. Since then, Alvenir has been able to call itself a spinout, characterized by the fact that technology has shifted from basic research to something that creates value for society.

Digital Danish is at risk of disappearing

Alvenir's goal of democratizing Europe with its language models can be seen in the company's 'open core' business model, where large parts of the company's machine learning models are exhibited open source on the developer portals GitHub and HugginFace, while the underlying software is developed and maintained behind closed doors.

"Right now we have a lot of languages in Europe that are underrepresented in digital speech recognition, and we think that is a real shame. Because we speak English well in Denmark, we have come to accept that our devices speak English, and it propagates down to our children and grandchildren. At some point, ‘digital Danish’ is at risk of disappearing. The same will happen to many other digital languages if there is no one who will help ensure progress,” says Martin Carsten Nielsen.

From the beginning, Alvenir's dream has been to develop and exhibit Danish speech recognition models, which can constitute a real alternative to the services offered by the major IT giants.

“This dream is still an important motivation for our work with Alvenir. Today, that mission has also become a competitive advantage, because we have identified a number of usage scenarios where we can actually outperform the same solutions by a not insignificant margin,” says Martin Carsten Nielsen.

"But fortunately, the tech giants do not rest on their laurels. Since we do not have a fleet of talented people to research and develop for us, it is therefore also an important core value for us to help develop the community around open source speech technology to stay relevant in the market."

In front Kasper Schjødt-Hansen, Lead AI Engineer at Alvenir (left) & Rasmus Stig Beck Jensen, Business Unit Manager at DTU Compute. Photo: Hanne Kokkegård, DTU Compute
Alvenir has an office at CPH Fintech in Christianshavn in Copenhagen. In front Kasper Schjødt-Hansen, Lead AI Engineer at Alvenir (left) & Rasmus Stig Beck Jensen, Business Unit Manager at DTU Compute. Photo: Hanne Kokkegård, DTU Compute. 

Language models are spreading

Speech recognition is becoming increasingly important for companies because it is an easy way to document what is being said in conversations for both internal use and customer follow-up. E.g., in insurance cases where there is a requirement to document customer contact to ensure that people's cases are processed correctly. Here Alvenir can listen and transcribe conversations.

The Alvenir platform has a wide range of applications, e.g. in audio and podcast service, where the software makes it possible to search libraries and make recommendations to users. But a language model that listens could also improve communication and security in the health sector and the financial sector.

“If a customer services representative has to write a summary based on a four-minute-long conversation, it is quite likely that he or she will forget something. Here, a good speech recognition solution can help to ensure quality and in the long run focus on relevant parts of the conversation, which assists the employee in their work and leaves time for other tasks,” says Martin Carsten Nielsen.

Other companies want to do a back-end analysis e.g., to see how a market develops and what customers are interested in. Here speech recognition can also play a valuable role by analyzing and identifying trends in large amounts of spoken data that companies can subsequently use in decisions making.

Data management and machine learning

On the surface, Alvenir's speech recognition is similar to what is already on the market, but the technology behind it is different, and according to Martin Carsten Nielsen, it is one of Alvenir's strengths:

"Alvenir does not automatically store data. Instead, we prefer to build our language model into the customer's system so that data never has to leave the company. It provides increased ownership for the individual user, which is a marked difference from many of our competitors."

Alvenir's technology uses signal processing and machine learning modules. The technology looks at 20 milliseconds of sound and learns what the sound wave looks like for the letters.

In this way, the technology models acoustic images, where an 'a' looks different depending on whether a dialect is spoken, whether the microphone promotes treble, etc., and where the model learns to see patterns and is doing a noise filtering to extract the sound itself.

Important to quantify uncertainty

Alvenir's technology also allows the end user to change the database of the speech recognition model to e.g., add a new product with a unique name that has not been seen before. If the tool in its interception hears an unknown word, it sends an audio clip back to the customer and says that it is unsure of what is being said. In this way, the user can introduce new training examples that the machine learning model can learn from.

“To make a general model, you need 100,000 hours of data; we do that. For domain-specific models, you need 2000 hours of data and a lot of training. We do that too. The last 2-3 percent, which make the model understand the customer's own technical terms, product portfolio, and company lingo, the customers manage themselves by updating the models' database continuously," says Martin Carsten Nielsen and points out, that it is important that Alvenir quantifies when the language models are uncertain.

It could have serious consequences in some companies if a language domain is not introduced well enough, and therefore Alvenir's models are 'assisted solutions' so that a person is involved in the process of approving an analysis or decision that has been made on the basis of a transcript.

"This is exactly the discussion with voice bots, and there are many different degrees of seriousness. It is e.g., not serious if a citizen is advised incorrectly about the public swimming pool’ opening hours. But it is important to decide when we feel safe enough to let a machine make a decision - and when we should send it on to 'warm hands', if it is, for example, a decision in an insurance case," says Martin Carsten Nielsen.

DTU's ecosystem for entrepreneurship shapes people

Rasmus Stig Beck Jensen, who is employed as Business Unit Manager with a focus on spinout and innovation at DTU Compute, has helped Alvenir founders Martin Carsten Nielsen and Martin Arpe Fogh Egebæk during their journey through DTU's ecosystem for entrepreneurship:

"The challenge is always to get something that is basically basic research - here basic sciences in computer science and cognitive systems - to fit into a business world, where things must be able to be used immediately and create value for society, and where there are also competitors. The goal of Alvenir was to build a spinout that could make money from day one. It is a difficult task to start a deep tech business primarily funded by customers. But, Alvenir has become a distinguished example of that.”

“We often see that researchers do not see themselves as entrepreneurs. So, in general we must work much more with culture, and at DTU we are well on our way to creating a definite ecosystem for entrepreneurship, which can also inspire other research institutions. Entrepreneurship can contribute to more value creation and even scaling of research results, and we often see this as a motivation for researchers.”

Right now, Alvenir has three employees, but the company expects to have six employees by next year.

"I think that few people would have pointed to us (Rasmus Arpe Fogh Egebæk and Martin Carsten Nielsen, eds.) a few years ago and thought of us as the entrepreneurs of the future. We had probably been placed as developers somewhere, but DTU’s setup for entrepreneurship has helped to drive this forward and shown that we are capable of doing that,” says Martin Carsten Nielsen.

Martin Carsten Nielsen, Co-Founder and CEO of Alvenir. Photo: Hanne Kokkegård, DTU Compute
Martin Carsten Nielsen, Co-Founder and CEO of Alvenir. Photo: Hanne Kokkegård, DTU Compute

Alvenir in brief

  • Alvenir is based on knowledge from a DTU Compute thesis, where Martin Carsten Nielsen and Rasmus Arpe Fogh Egebæk under leadership of Professor Lars Kai Hansen in the research section Cognitive Systems developed an AI-based model for speech recognition (Danspeech).
  • Martin and Rasmus were then hired as research assistants at DTU Compute to create a software package, but the thesis idea turned out to contain a really good business opportunity so that basic research could create value for society.
  • With the help of DTU's setup for entrepreneurship, and the programme Open Entrepreneurship, the researchers received support from the DTU Discovery Grant, which supports DTU employees' work with early technical and commercial maturation and the risk of technology. Then money from the DTU Proof of Concept Grant (POC), which provides grants for e.g. development of prototypes for user validation and demonstration of scalability, but also for salaries, external consultants, reports, and more. Finally, the InnoExplorer Grant at the Innovation Fund Denmark has enabled Alvenir to continue working on product development and business plans.
  • On 1 April 2022, DTU signed an agreement on technology transfer, which has been paid in an ownership model, so that Alvenir is officially a spinout.