The Linguist

Chartered Institute of Linguists SUMMER 2025 The Linguist 27 SECTION HEADER REVIEWS speeches before being developed to support every stage of content creation, including subtitling. Others, like Limecraft and Stellar, were made to support subtitle creation mainly for broadcast media. The features I tested were transcription and translation: transcription of English speech into English subtitles, and translation of English speech into Greek subtitles. I asked each tool to transcribe, translate and timecode five documentaries in the historical and true crime genres. Within a few minutes I was able to export the output in SRT (SubRip Subtitle) format. I then compared the subtitles created by the AI tools to subtitles created in the traditional way. It is important to note the limitations of my tests, which mainly concern time and language. I was only given access to the tools for a limited period, but I felt that the time I had was more than enough. I am able to test only two languages – English and Greek – as these are the languages I am fluent in. The results I gathered for Greek were very revealing, but it would be interesting to find out how the systems deal with languages from different language families. For instance, would they work better for Germanic languages as opposed to Afroasiatic languages? Transcription issues Although I focused mainly on Trint and Limecraft, the results for transcription were surprisingly similar across all the tools. There were a high number of mishearings, particularly in cases where there are sounds in the background and intense dialogue. For instance, in a scene where two women are walking in the mountains in stormy weather, one says "I'm cold", which Trint transcribed as "I'm calm". The software performed better when there was a narrator telling the story rather than action with dialogue. Line breaks are a very frequent issue; sometimes there are four-, five- or even six- line subtitle segments which cover the face of the speaker on screen. In this respect, Limecraft's output was closer to the standard guidelines (i.e. a two-line maximum). This was to be expected given that Limecraft is specialised in video production, whereas tools like Trint are more specialised in transcription and subtitling for other industries, such as newsroom workflow and content creation. A further problem is that dialogue is not presented in dialogue format – for instance, there is no dash at the beginning of a dialogue line. On the positive side, timecodes are remarkably accurate, but that doesn't mean that timecode adjustments that: can AI aid subtitlers? won't be needed after any line break errors are corrected by a human editor. Translation results Incorrect line breaks are to be anticipated, even with traditional subtitles. What I found more interesting was the translation of content. The AI is surprisingly accurate when translating simple phrases. For instance, Trint correctly rendered "Good morning, I hope you had a lovely time yesterday evening" as Καλημέρα, ελπίζω να πέρασες όμορφα χθες το βράδυ. However, when it comes to metaphors, local slang and cultural references the translation is, predictably, too literal and the sentences do not make sense in most cases. For instance, the word 'catfish' (i.e.

• Cover