Language translation feature from Nepali to English not functional

Pratik

Dear Alex/Paco, Please note that even after updating Dedalo to the latest version, the automatic translation feature CTA for our native language Nepali to English is actionable but not functional at all. As you can see in the attached screenshot after pressing translate only a few words are translated to English while most words are copied as it is.
Kindly suggest, any workaround for this, or if this issue will be addressed in near update.

Thank you.

Alex

Hi Pratik

We are working to change the translator, right now, Dédalo is using Apertium technology that is not working well from Nepali... I will back soon... (we are working hard about it)

best

Alex

Hi Pratik

We added a new AI translation engine into the version 6.9.0, it can translate Nepali to 54 languages (English, Spanish, French, etc.)

Update to obtain it!
To run it, you'll need a modern computer with a browser that supports WebGPU. The latest versions of Chrome, Edge, and Safari work well. Please note that the result is not deterministic (unlike Babel) because it's an AI process that can hallucinate during the process. As a result, some tags may be re-positioned. Review the result

The process takes time. Please refrain from closing the browser window, as the process is currently running in the tool_lang window on your computer. You can continue with other tasks, but ensure that you do not close the tool.

I hope it helps.

Best

Pratik

Hello again team. Thank you for integrating the Local AI translator, we were excited to test it out but unfortunately the local AI translator keeps getting stuck in an infinite process and produces no result. We have tried logging in from high end devices too to see if it was a memory or processor issue but same problem there. Is there something we can do from our end to fix it or shall we await update in Dedalo itself.
Pls find attached screenshot for reference:

Alex

Hi Pratik.

Can you tell me the browser that you are using? we test it in Chrome, Edge and Safari (in Safari you need to able it, is not by default)
and its version of the browser are you using? (only works in latest versions)
I need more information about the error, so, can you send me the console error showed in the browser? (press F12 to cmd+opt+j or ctrl+opt+j )

Note: the AI translation uses the graphic GPU of your computer, (not the typical CPU RAM) the process needs a good GPU card with at leats 8GB of VRAM.

If you are using a Mac, all M series (M1. M2. M3. M4. M5. in normal, Pro, Max, Ultra version ) works super fine. (Apple silicon is the best local architecture to use local AI right now)
If you are using a PC, you will need a Nvidia or AMD GPU card with at least 8GB VRAM
If you are using a PC with Linux, you will need to enable it.

Note2: see the WebGPU implementation here

Local AI is nice, but requires a correct hardware.

Best

Pratik

Thank you for the quick response Alex.
The issue seemed to be with the browser on our part.
We are using Google Chrome and the hardware specifications match with the requirements listed by you.

The translation feature is now working partially, Our issue was with http protocol where the browser was not running VRAM usage on http, removing restrictions on dedalo http IP on the browser got the translation feature to work with few inaccuracies.

Trying to get it working on the 'compatible but slower' option however produced the following error - Cant create a session ERROR_CODE 9 (screenshot attached). Kindly let me know if you have any suggestions on this, we are trying some workarounds and multiple hardware options in the meantime.

Thank you again.

Alex

Hi!

The more compatible, slower option removes the WebGPU support, and the AI will work with your CPU instead of the GPU. This option is not a good idea to be used for huge transcriptions; it takes over 20 times slower than the GPU and it could close the browser session. This option is implemented for shortest texts. (and I will review the error code carefully.)

If you can use your GPU, it is the way to go.

I tested Nepali texts, and the model translated almost correctly, but with large transcriptions, it could make hallucinations and some parts became untranslated.

How can you deal with it? If you need quick translations, you can change the untranslated parts manually.
And we need some cases to test how we can harness the model to be more accurate, so if you can send me text fragments of the interview to reproduce the issue, we can optimise the context of the AI, but this will take time.

Note: Don't use this chat for the examples. If they have sensible information (the àgora is a public site), use my email.

Best

Pratik

Alex Thank you Alex, sure we shall email you some of the Nepali Transcripts along with Translated texts to train the AI model or reproduce the issue. I shall heed your suggestion on the sharing too. Thanks.

Alex

Hi Pratik

We've updated the AI translation engine in v6.9.2, providing you with more options and models. To access the new version, please update your installation. Now, you can choose from models other than translategemma. This model has some issues and doesn't work properly for large translations. I recommend trying Qwen; in my tests, it performs much better at similar speeds.

Best
Alex