Since we have many folks here working with OER in multiple languages. Let’s say I have a hypothetical re cording where my voice asked questions in English, and two respondents answered in 2 other native languages, so I have an audio in 3 different languages.
I have now a fully multilingual transcript as a text file, what is a good way I can generate the full transcript in each language?
Alan, this isn’t an OER question per se. It’s a question about which tools to use to do translation. And, it’s important to know what your intended purpose is (always), with what languages are you working, what file format is the target, are you providing only a text file and no audio. Then, you must decide which license to use and why.
There will probably be other questions to answer before you’re done. And, because it’s OER, it will never be done.
Well, then I have to give away my element of surprise.
This is for an episode of our OE Global Voices Podcast, all shared under CC BY. This episode includes two segments, in the first, I asked questions in English to our guest, who responded in Spanish. The second segment is in English, but at one point, I asked the guest to respond in French.
So I have my audio already edited that includes spoken English, Spanish, and French. The software I use for audio editing has already transcribed into a text file the full content, in mixed languages (I can output in PDF, or docx, but I usually do transcripts in .txt
As a recent episode example, I provide the audio and the transcript.
What I would ,like is to be able to generate a version of the 3 language text transcript, into ones that are translated fully into English, French, and Spanish. So I am looking for something to translate a text file of English, French, and Spanish and get a version say all in Spanish.
Hi Alan. Have you looked at https://omegat.org/. Or perhaps if you are on Wikipedia, you might use the translation tool. I’ve got some basic instructions but Bobby’s Video is better.
Thanks Derek, OmegaT is interesting and a commendable one but that looks like to me for people who are working as translators, like a transcriptionist tool. I guess it looks like you can plug in some AI services.
I was able to finish this episode editing, first of all, here is the end result for a new OE Global Voices podcast
This conversation was with two of our OE Global Board members, as an introduction to them as people, leaders, and thinkers. Maria asked me at recording time if she.could respond in her native Spanish to my questions asked in English. Why not? The second half is another recording with Perrine- I made the same offer since her first language is French, but she wanted us to chat in English. But I did ask her one question to respond in French to speak to her Francophone colleagues.
Again, I have source audio with three spoken languages. Since January 2023, I have switched from using a traditional wave form editor, many years using Audacity, to Descript which not only generates transcripts, but turns audio editing to an experience of doing it mostly the text of the content – it’s been a game changer for someone who started podcasting in the early 2000s.
You can find more details of how the podcasts are made at https://podcast.oeglobal.org/about/.
I have used Descript previously to transcribe a episodes that were recorded completely in Spanish; I had gotten access to the beta versions of the newest “Underlord” features of the tool (just released this week) which included ones for translating transcripts, but uit was not quite able to handle the multlingual aspects. Still, it provided the main transcript of this episode with three languages represented - or follow along with the version you can listen to and see the transcript displayed.
But what I really wanted to do is have versions of the transcript with all content in Spanish, and another in French, and also in English, all the content. Plus I wanted to have the podcast published web page include translations of the episode description, some of the quoted, in the 3 languages.
I ended up doing all of this with back and forth copy pasting from the original transcript into Google Translate, using the language detection mode, I could used, say mixed text of English and Spanish to translate all into Spanish or French. It was a bit back and forth, and of course, is far from ideal, but seems to me good enough.
The other problem it solved for me is fine tuning the editing of languages I am not fluent in. For Marisol’s spoken Spanish, the Descrpt transcription often gets the places it marks as sentence endings wrong, and then as I do not understand Spanish, I cannot do much editing of the audio. But having an English transcription, I was able to edit it by matching the flow from English to Spanish.
Again, some of this is helped by tools, but also took quite a bit of manual processing. There likely are more elegant or practical solutions (e.g. “Alan should learn Spanish!”) but to be honest, I like the hand crafted parts, even if it took a bit longer than ideal.
Mostly, I want to be able to give more people we interact with the opportunity to speak in their natural language.
Beyond all this blather on the technology and please enjoy the conversations we are sharing with our board members, and get a chance to know them as both leaders and regular people, their interests in say, running and family and art and the places they live.
Did you hear something of interest? Please reply to the episodes posting here in OEG Connect.
Folks might be interested in this report from Intento on machine translation
Turns out it varies greatly with language pairs and materials. Here’re some key findings:
LLMs reshaped the MT landscape and now account for 55% of top-performing models
Quality varies by language and domain, with LLMs excelling in Colloquial, Education, and Entertainment
Traditional Machine Translation (MT) still outperforms in specific language pairs and domains
LLMs are less expensive but slower than MT engines
Open-source LLMs lag behind commercial offerings
Customization through translation memories, glossaries, and prompt engineering is key to eliminating errors