Speech to Text Adds Norwegian, Swedish and Dutch

[TeresaDemel]

We have added Norwegian (Bokmal), Swedish and Dutch to the Transcription Panel in Premiere Pro. Since I don't speak any of these languages I would love to know how well they are working for you. Please try some of your footage and let us know:

• What kind of content are you transcribing? (zoom call, promotional video, YouTube DIY tutorial, Netflix documentary...) 
• Does it do better with certain dialects? If there are regional dialects that it doesn't handle as well, be specific about the names of the region, accent or dialect.
• How accurately does it transcribe numbers? (1984, February 1st)
• How accurately does it handle pop culture (movies, current events, famous people)?
• Will you use it in its current state?

[Kevin-Monahan]

Sweet. What do you think about this development @Ann Bens? 🙂

[Ann Bens]

Yeah, already gave a thumbs up.

Will test soon.

[Joost van der Hoeven]

Great! Will test this soon. 👍🏼

[R Neil Haugen]

If @Jarle Leirpoll  could hop on and test, det var snill ... I don't have my computer setup for Norsk language work.

Neil

[Jarle Leirpoll]

I tested it on a couple of projects now, and it works OK, but not great. Norwegian language have a lot of dialects, and it doesn't understand those. Voice-overs in plain mainstream Norwegian (bokmål) are pretty accurate - although there are some problems with words that are two words put together. We write those without a space between them so they're one word, but Premiere Pro likes to split them into two words.
It's not bad for a first try, but I really hope that Adobe collects all our edits in the text, so the AI can learn from them.

[Jarle Leirpoll]

Oh, and years, like 1890 (pronounced "atten nitti" - meaning eighteen ninety) was written 18 90. Some names of places and people were not correct, and it likes to put a full stop and a Capital letter in the middle of some sentences..

So to answer @TeresaDemel : Would I use it in it's current state? Well, I would, but I wouldn't be satisfied. Too many errors to fix. It's nowhere near the accuracy of the English version. I think it needs more training. 🙂

[Jarle Leirpoll]

I also get a crash every time I double-click a caption clip to open it in the Text panel, which is a hassle. I'm sure that bug will be fixed soon.

[R Neil Haugen]

I was really wondering how it would do past Bokmal and "Olso folk" ... as from my time in Norway and with Norwegians there and here, the pronunciations between valley-by-valley dialect changes are at times difficult to track.

Neil

[Kerstin Ebert]

Hi Jarle, are you able to catch a crash log when this crash is happening? I'm not seeing it on my end. Thanks!

[Jarle Leirpoll]

I just used the standard Windows "report a crash" thing that appears whenever a software crashes. I added my email address, so if you search for "leirpoll" you should find the log. If that's not enough, can you please remind me how I can grab a better crash log and send it to you?

[TeresaDemel]

Hi Jarle, We have updated the model. Numbers should be improved. We have also added more training data for Norwegian. Could you give it another spin? 

[Neckebba]

Is then Dutch. (BELGIUM) comming?  :-). VRT is hoping for it 🙂 

[TeresaDemel]

Hi Neckebba, Are you saying that Dutch doesn't do well with the Belgian dialect? Are there any kinds of words that it doesn't handle well? (slang, pop culture, numbers, etc)

[Neckebba]

Hello there,

is there any idea when it moves from beta to production? VRT is eager to test this out on a big scale here :-). For Dutch 🙂 

[Ann Bens]

Tested Dutch on two projects (travel videos).

It does make mistakes, some of them understandable. Due to the way we write the words, e.g. 's morgens (in the morning)

Has some trouble with our way of saying numbers: e.g. 23 is transcribe to twintig en drie (twenty and three). It is the way we say that in Dutch). But it should say drieëntwintig.

It also picks up audio (music) from another track, even though I selected one specific track, it transcribe to Uh, Uh.

Overall its not bad at all. I would use it.

[TeresaDemel]

Ann, could I get a clarification on the 23 issue? Did the person say "drieentwintig" and it transcribed as "twintig en drie?" 

[Ann Bens]

Ann, could I get a clarification on the 23 issue? Did the person say "drieentwintig" and it transcribed as "twintig en drie?" 

By @TeresaDemel

That is correct.

I also had some Spanish geographical words that came out rather funny.

[TeresaDemel]

Ann, could you try the latest beta? You should see improvements on how we show numbers in Dutch. 

[Ann Bens]

Ann, could you try the latest beta? You should see improvements on how we show numbers in Dutch. 

By @TeresaDemel

Much better.

Trips over nuances in the language, such as 124ste (English 124th) translated as 100/24.

[Denise24566295b6j5]

Where can i download the beta?

[Ann Bens]

It is in your CC desktop app.

[Denise24566295b6j5]

I tested the dutch language. It works OK. But if we use a english word or name for example harry styles it will give me ''harries titels'' Some words we also use in english like, cool or nice ore someting like that. 

[TeresaDemel]

Hi Denise, Thanks for sharing. "Loan words," especially from English, is something we haven't solved but we are investigating. I'm glad you brought it up as part of your testing. 

[Ann Bens]

Here is another one for numbers in Dutch.

Voice over say:  2450 meter

Transcribe to 24, 100, 50 meter which is exactly the way we say that but not the way we write those numbers.