Useful Finnish transcripts and readable SRT subtitles from specialised dental education videos
Company
QAdental (Suomen Kotilääkäripalvelu Oy)
Industry
- Healthcare & Life Sciences
- Education & E-learning
Problem
QAdental provides Finnish-language continuing education content for oral healthcare professionals, including webinars, podcasts, video lectures, and procedure videos. The content is directly accessible to Finnish-speaking professionals, but its international reach is restricted by language.
The Finnish market for specialised dental education is limited, while potential audiences in Sweden, the Baltic countries, and other markets cannot use videos available only in Finnish. Accurate Finnish transcripts enable reliable translation and help QAdental extend its existing video library to international audiences. They also improve accessibility for viewers who are deaf or hard of hearing and provide a foundation for services such as video search within the library.
General-purpose speech-recognition systems have difficulties with Finnish spoken language and specialised dental terminology. Recognition errors involving clinical terms, proper nouns, omitted words, or short but meaningful expressions can change the meaning of a sentence. The transcription quality also varies depending on the recording and how clearly the lecturer speaks. Manually transcribing and correcting long educational videos is time-consuming, while the large size of the original video files creates additional processing challenges.
The main challenge was therefore to develop a practical process for augmenting Finnish dental education videos with sufficiently accurate, readable, and editable transcripts and SRT subtitles. Reliable Finnish transcription is a prerequisite for translating the educational materials, making the existing video library accessible to international audiences, and using the transcripts in other digital services, for example, for video search within the library.
Solution
The AI-assisted solution extracts audio from large educational video files using FFmpeg and uses speech-recognition models to generate Finnish transcripts and time-coded SRT subtitle files. An AI enhancement stage improves the structure and readability of the initial transcript while limiting the amount of new text that the language model can introduce. The transcript remains available for human review and correction, which is particularly important for specialised dental terminology, proper nouns, omissions, and errors that could affect meaning.
Various AI models were compared during the project, including WhisperX with Finnish-NLP/whisper-large-finnish-v3-ct2 and alignment, OpenAI Whisper Large, OpenAI GPT-4o Transcribe, and Google Gemini 3. This comparison supported the systematic evaluation of different approaches to Finnish-language dental transcription rather than relying on a single model.
The pilot also supports Finnish-to-English translation after the Finnish transcript has been reviewed. However, development focused primarily on producing a reliable Finnish source transcript because testing indicated that translation works adequately when the source transcript is correct.
The solution processes large video files and produces subtitles designed to follow established readability practices for Finnish audiovisual content. The intended next step is an API-based workflow through which QAdental can submit videos or extracted audio directly from its own content library and process multiple recordings with limited manual intervention.
The solution was co-developed by QAdental and Haaga-Helia University of Applied Sciences as part of the development of the GAIK GenAI Toolkit. The company pilot contributed practical requirements, testing feedback, and evaluation results to the co-development of the toolkit. At the same time, reusable components produced for the GAIK GenAI Toolkit were configured and combined to create the company-specific transcription and translation solution.
The GAIK GenAI Toolkit provides reusable technical components and implementation guidance for creating company-specific generative AI solutions. This use case is linked to the toolkit’s knowledge capture process, as it extracts information from video and audio content in the form of transcripts and subtitles. Reusable toolkit components were configured and combined rather than developing the QAdental solution entirely from scratch.
The University of Helsinki supported testing and evaluation of the transcription output, including ground-truth preparation. Tampere University conducted an AI adoption readiness assessment. Cooperation among the universities provided a structured foundation for development through model comparison, systematic testing, and evaluation.
Outcome
The pilot demonstrated that AI-assisted transcription can produce useful Finnish transcripts and readable SRT subtitles from specialised dental education videos. During the pilot, 50 videos from the QAdental library were processed and complemented with transcripts and subtitles. We found the overall transcription quality promising, although the results varied depending on the recording and speaker. Some transcripts required substantial correction, while others required little or practically no correction. Once sufficiently accurate Finnish transcripts with correct dental terminology had been produced, they were translated into English subtitles. This made the existing Finnish-language videos accessible to international audiences without requiring the original audio to be re-recorded.
We considered the resulting subtitles clear and easy to read. The enhanced subtitles followed conventions commonly used in Finnish audiovisual content and converted spoken Finnish into a more readable form without making the language unnecessarily formal.
The pilot also confirmed the importance of retaining a human-in-the-loop review stage. Dental terms, proper nouns, omissions, and short words can still be transcribed incorrectly. In some cases, the omission of a word such as “no” could reverse the meaning of a sentence. Human validation therefore remains necessary before specialised educational subtitles are published.
We found the caption-correction functionality efficient and convenient compared with external subtitle-editing tools we had previously tested. Reviewing an entire transcript still requires concentration and professional knowledge, but the correction interface effectively supports the necessary validation work without unnecessary functionality.
We consider integration into our future workflow highly feasible and support continuing development from the pilot toward a production system. The intended future workflow would use an API to import videos or audio from the QAdental library, produce transcripts and subtitles, and return the results to our content-management process.
ROI
Quantitative ROI has not yet been established at the pilot stage.
The expected return is based on reducing the manual work required to create initial Finnish transcripts and subtitles, making subsequent human correction more efficient, and reducing reliance on fully manual transcription and external subtitle-editing tools.
Accurate Finnish transcripts also enable more efficient translation and reuse of existing educational videos for audiences outside Finland. This supports QAdental’s ability to enhance its continuing education library and extend the availability of its content beyond the Finnish-language market.
The pilot also provides a technical and organisational foundation for batch processing QAdental’s video library and integrating transcription into the company’s content-production workflow. Quantitative ROI can be evaluated after the integrated workflow has been tested with a larger number of videos and compared with the previous manual process.
Technologies
- AI: AI for Speech Recognition & Synthesis, Natural Language Processing (NLP)
- Closed selection: Azure
- Free-text field: GAIK GenAI Toolkit; CSC Rahti; FFmpeg; Various AI models and speech-recognition approaches, including WhisperX with the Finnish-NLP/whisper-large-finnish-v3-ct2 model and word-level alignment, OpenAI Whisper Large, OpenAI GPT-4o Transcribe, and Google Gemini 3.
Partners
The solution was co-developed with Haaga-Helia University of Applied Sciences as the primary development partner. The University of Helsinki supported testing and evaluation of the transcription output, including ground-truth preparation, while Tampere University conducted an AI adoption readiness assessment. The collaboration provided a structured foundation for development through model comparison, systematic testing, and evaluation.
Process Impact
- Product Development & Innovation
- Operations Management
Languages
- Finnish
- English