This project implements a web-based interactive language learning platform using Gradio. I built this because it helps with my German learning process.
explanation_gradio_app.mp4
This is sequencial pipeline design based in STT->LLM->TTS. The workflow is as follows:
- Input: Accept either spoken input (mic) or typed text.
- Speech-to-Text (STT): If audio is provided, the cheap OpenAI
4o-miniis used (approx. $0.003/min). - LLM: The trascribed text (or typed input) is passed to a free LLM (with OpenRouter) with appropiate context, returning your question in the target language.
- Text-to-Speech (TTS): The response is spoken using a the same OpenAI
4o-mini(approx. 0.015/min).
I'm currently allowing for experimentation the following models:
deepseek-r1-distill-llama-70b:free(DeepSeek Llama 70B)meta-llama/llama-3.3-70b-instruct:free(Llama-3 70B)meta-llama/llama-4-scout:free(Llama-4 Scout - default)meta-llama/llama-4-maverick-17b-128e:free(Llama-4 Maverick)
As you can see, all my models are 70B+, that's because I want consistent and clear sentences in a variety of topics (and contextual information as languages are learned that way). The Llama-4 have a MoE architecture, so very interesting conversation can be obtained, which much better translations. *Personally, Llama-4 Scout works smoothly.
Important
Currently I'm enforcing max_completion_tokens to 1024. This is for testing.
It can be easily modified to a larger value. Change it accordingly!.
completion = client.chat.completions.create(
model=model_id,
messages=messages,
max_completion_tokens=1024 * 8 # CHANGE HERE
)You need Python 3.x and API keys for OpenAI and OpenRouter
- Clone the repository.
- Install packages
pip install openai gradio- Set API environment keys.
export OPENROUTER_API_KEY="your_openrouter_api_key"
export OPENAI_API_KEY="your_openai_api_key" # Optional, only for STT/TTS.Run in the terminal
python app.pyand locate the localhost port provided by gradio.