Amberscript has now developed its own ‘engine’: the motor that ensures that Dutch audio is translated into Dutch text. Peter-Paul de Leeuw (Amberscript’s CEO) explains that work on the engine is in full swing, but his own product should already translate 80 to 90 percent well, depending on the audio quality.
Seeing is believing, so we took the test. An interview without microphones (the use of microphones is recommended) with complicated technical terms still produces quite a few errors. It’s different with a clearly spoken piece of text into a microphone. Here, Amberscript does an amazing job. A piece of sample text of about a hundred words on artificial intelligence is translated almost flawlessly. Only the term “master-slave relationship” from philosopher Hegel presents difficulties. That becomes “most slave relationship. Not crazy, but it doesn’t add up. In the small test, Apple’s Siri and Google Docs make the exact same mistake. The two superpowers otherwise make a mess of things. Of course: unlike Amberscript, they are free programs (for example, Apple includes the option in the Notes app; Google includes it on desktops in the Chrome version of Docs), but the difference is striking.
Take a sentence like, “artificial intelligence will soon imitate much of what humans do, both in the physical and intellectual realms. Google Docs brews this from it: ‘the artificial intelligence will soon bravely imitate much of what hatted and it does so on the physical as well as the intellectual plane.’ And Siri? That one keeps it too: ‘artificial English genesis will do profile sole imitate of that human does so on the physical as well as on the intellectual plane.’ And from Marx and Buber, Siri stupidly makes ‘Max de Boer’. You can do better than that, Siri.
De Leeuw is also far from satisfied with his own software, but already it offers significant time savings for people who need to transcribe speech, such as journalists, students, doctors, or court employees. Amberscript initially focuses on the transcription of interviews. Indeed, the software neatly distinguishes between the interviewer and interviewee. As more training material for the engine becomes available, the machinery should improve. With each improvement made by users, it will.