A Japanese transcription/diarization pipeline with Pyannote and Whisper large-v2 that uses a custom-tuned segmentation model and custom filtering on the audio (low-pass filter, equalizer, etc.) for improved performance. Can be given a video file or mp3/wav file.
Performance is considerably better than default JP whisper for most tasks involving Japanese content, with the exception of singing/karaoke (Where performance is below the original due to the training dataset.)
Requires ffmpeg, openai-whisper, pyannote and facebookresearch's demux model. Cuda is also strongly encouraged. Pyannote requies a Huggingface API key, which it will currently look for under the environment variable "HF_TOKEN_NOT_LOGIN" (At the time of this writing, naming your HF token "HF_TOKEN" causes bugs.)
Originally intended as a solo project, but I'm upping it here in the hopes it will be useful to practicioners. If you're doing work in this space please feel free to reach out.