Whisper Meeting Transcription and Translation

中文版

Local subtitle generation for audio and video, plus Chinese-English translation. The project uses mlx-whisper for MLX-accelerated transcription on Apple Silicon and runs gpt-oss:120b through local Ollama for fully offline translation.

• Apple Silicon Mac (M1 / M2 / M3 / M4)
• Python 3.11+
• ffmpeg for media decoding
• Ollama for translation (optional)

# Install ffmpeg
brew install ffmpeg

# Install Python dependencies
pip install -r requirements.txt
# Or install from pyproject.toml
pip install -e .

# Pull the translation model on first use
ollama pull gpt-oss:120b

Put your audio or video files into the media/ folder, then run:

python main.py

Subtitles are written to subtitles/ and keep the original media extension in the filename:

media/lecture.mp4   ->   subtitles/lecture.mp4.zh.srt
                    ->   subtitles/lecture.mp4.en.srt

By default, Whisper detects the language automatically and only generates subtitles in that detected language:

python main.py

If auto detection is wrong, you can force the source language:

python main.py --language zh   # force Chinese transcription
python main.py --language en   # force English transcription

More examples are in QUICKSTART.md.

python main.py --force
python main.py --no-translate
python main.py --language zh
python main.py --language en
python main.py --target-language zh
python main.py --target-language en
python main.py --input ./my_videos
python main.py --output ./subs
python main.py --whisper-model mlx-community/whisper-large-v3-mlx
python main.py --ollama-model gpt-oss:120b
python main.py --ollama-url http://localhost:11434

Chinese -> Chinese, generate only Chinese subtitles:

python main.py --language zh

Chinese -> English, generate Chinese subtitles plus English translated subtitles:

python main.py --language zh --target-language en

English -> Chinese, generate English subtitles plus Chinese translated subtitles:

python main.py --language en --target-language zh

English -> English, generate only English subtitles:

python main.py --language en

The project probes audio streams with ffprobe, so any format that ffmpeg can decode is supported, including but not limited to:

.mp4 .mov .mkv .avi .webm .m4v .mp3 .wav .m4a .flac .aac .ogg .ts

The model is downloaded from Hugging Face and cached locally on first use.

• media/ and subtitles/ are already in .gitignore
• The Docker image does not support MLX acceleration and is only intended for Ollama-related deployment work

Run the small srt_summarizer unit test suite with:

python3 -m unittest discover -s test -p 'test_srt_summarizer.py'

These tests use very short inputs and mocks, so they do not trigger expensive real summary computation.