Min Local English Tutor

English|繁體中文

This project is a minimal implementation of a local English tutor designed to help users improve their English language skills.

Model Used

• Whisper-medium: Speech to Text Model (Hugging Face)
• Gemma3 12B: Local LLM that runs by ollama (Ollama) (Google Blog)
• Kokoro-82M: A light weight model for text to speech (Hugging Face)

Usage

main.py

The main.py script allows you to interact with the chatbot, save conversations, and load previous conversations.

Command-Line Arguments:

• -l: Specify the file to load a previous conversation from (default: message.json).
• -s: Specify the file to save the current conversation to (default: message.json).

Example Commands:

Load and save using custom files:

python main.py -l previous_conversation.json -s new_conversation.json

main_audio.py

The main_audio.py script allows you to interact with the chatbot using voice input and output. It supports saving and loading conversations to retain memory across sessions.

Run without memory:

python main_audio.py

Enabling Memory:

To enable memory, uncomment the following lines in the script:

# try:
#     mesg = load_conversation(load_file)
#     english_tutor = chatbot(mesg)
# except FileNotFoundError:
#     english_tutor = chatbot()

And comment out this line:

# english_tutor = chatbot()

This allows the chatbot to load previous conversations from the file specified with the -l argument and save the updated conversation to the file specified with the -s argument.

Example Commands:

Load and save using custom files:

python main_audio.py -l previous_conversation.json -s new_conversation.json

Commands:

• /r: Start recording your voice.
• /s: Stop recording.

Notes:

• Memory is disabled by default because the Whisper model may not always transcribe speech accurately, which could lead to issues when loading or saving conversations.

Dependency

Ollama Server:

• Use latest version to be compatible with Gemma3. LINK

• Download the model LINK

# using 12B in my case
# it takes some time when run the first time 
# only needed to run when download the model
# use /bye to exit when it is done
ollama run gemma3:12b

# use ollama server to host the model
# always run it before main.py or main_audio.py
ollama serve

Python Library

• python version: 3.11.10
• library (see the attach links and check installation details is the best way to make sure it works properly)

pip install pyaudio
pip install numpy
pip install soundfile
pip install playsound
pip install ollama

# reference https://pytorch.org/get-started/locally/ for GPU or CPU
# pip3 install torch torchvision torchaudio

pip install transformers #https://huggingface.co/docs/transformers/installation

My Hardware:

• RAM: 48GB

• GPU: RTX3060

• VRAM: 12GB

• try smaller model if your hardware has less VRAM or RAM

Remark

Upon seeing Gemma3's performance, I just can't wait to see how it performs on this task. For a 12B model, it does remarkably well on what I want with a simple instruction. It's crazy!!