Cookie Notice
Diskusija uses cookies to improve your user experience and our site. By using our website, you consent to all tracking cookies in accordance with our Website Privacy Policy. You may withdraw your consent at any time.
Project summary
Client type:
a language service provider operating as a contractor for the end-client
Duration:
2 months
Languages, technology:
Russian, Polish and English; software similar to transcription software
Goals achieved:
creation of audio corpus for testing voice recognition software, timely deliveries, approval for a second project phase.
Introduction
The client wanted to develop and test tools that could recognise and interpret human speech in various audio formats, regardless of audio quality.
Challenge
The project required teams of linguists who could meticulously identify and label sound activities in bad audio conditions. It was important to accurately complete as much data as possible within a few weeks before project funding ran out.
Solution
Initially, Diskusija was assigned to handle Russian audio recordings, but soon enough the client faced repetitive issues with the Polish and English teams and asked us to take over these languages, too.
We tested and evaluated the linguists immediately.
The client’s aim was to be able to distinguish music from speech, as well as from noises that might “seem” musical (machinery, birdsong, etc.). Capturing vocalised human sounds like laughing, grunting, humming and sighing was also involved. If executed poorly, the client would have ended up with sloppy data.
Part of the project was to train the team by giving them feedback on the completed files. We had to closely monitor deliveries and answer questions on a daily basis. Fortunately, we enjoyed the full support of and great involvement from the client who was enthusiastic about the project, which led to the production of quality data output
Results
Audio recordings segmented:
20.7 hours of Russian
10 hours of Polish
8.4 hours of English
It takes between 6 to 8 hours of work to complete 1 audio hour (20-30 clips).
All in all, over 250 hours were spent on all three languages.
The success of the first phase meant the client could re-apply to receive funding for a second project phase. The progress achieved was a strong argument in favour of its continuation and resulted in approved financing for the further development of voice recognition software.
