LocalLLaMA

3 readers

1 users here now

Community to discuss about Llama, the family of large language models created by Meta AI.

founded 1 year ago

MODERATORS

communick@poweruser.forum

LLMs for learning a foreign language? (alien.top)

submitted 11 months ago by ethertype@alien.top to c/localllama@poweruser.forum

7 comments fedilink hide all child comments

Hi.

Anyone got any experience with using (a set of) local LLMs for practicing a new language? (Spanish, not Python). Curious about experiences and knowledge gained.

And, in the extension of that thought, what would be required 'scaffolding' around a set of LLMs to be able to:

assess a student's current proficiency
set up some kind of study guide
provide assignments (vocab training, writing prompts, reading comprehension, speaking exercises, listening exercises)
evaluate responses to assignments
give feedback on responses
keep track of progress over time and adjust assignments accordingly

I *assume* something like this would require multiple LLMs, in order to handle Text To Speech and Automatic Speech Recognition. Is whisper (for example) useful for evaluating (and give feedback on) pronunciation?

you are viewing a single comment's thread
view the rest of the comments

[–] tinykidtoo@alien.top 1 points 11 months ago (2 children)

I had some success making a workflow to use whisper to speak my target language. A llm made in the target language. And a tts capable of of producing my target language.

This allowed me to practice conversations. Some were hit or miss, I suspect this is because I am very new to my target language. But it was useful and allowed me to practice things like order food.

The language app memrise has a simlar system, of course for a price.

[–] ethertype@alien.top 1 points 11 months ago (1 children)

Would you care to describe your setup in more detail? Do you have any notes suitable for publishing on github or similar?

[–] tinykidtoo@alien.top 1 points 11 months ago

I am using the text-generation-webui by oobabooga https://github.com/oobabooga/text-generation-webui

One of the built-in plugins is the whisper_stt, you will need to enable it in the settings of the webui. https://github.com/oobabooga/text-generation-webui/tree/main/extensions/whisper_stt

I have been using the Elyza-japanese-llama-2-7B. Other models specific to your target language should work. https://huggingface.co/elyza/ELYZA-japanese-Llama-2-7b

Lastly, I created my own plugin, that is no longer maintained, unfortunately. Using a similar python script to the silero_tts extension, I swapped in calls to bark tts. I only chose bark because it had a Japanese model.
https://github.com/suno-ai/bark

But, you might have some luck with the new coqui_tts, which is under development. Hopefully they will fix the error I have been having with multi-language support. But it's built in, you would just need to install the requirements.txt https://github.com/oobabooga/text-generation-webui/tree/main/extensions/coqui_tts