Transcribe any audio file and convert to a Microsoft Word document. Simple wrapper around ffmpeg, whisperx (whisper pyannote) and pandoc, initially intended for interview transcription. Assembled by @4yn.
Quickstart colab notebook (make sure to pick the GPU runtime and to shutdown once done)
./wxscribe.py audio_file_1 audio_file_2 audio_file_3 ...
You will need to accept the pyannote model
ToC and generate a
huggingface model hub token for the
script to work. Provide the token using command line option --hf-token hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx or using the HUGGING_FACE_HUB_TOKEN
environment variable.
An audio file my_audio.mp3 will be transcribed to my_audio.docx. If some
my_audio.docx already exists it will be skipped; delete the old document if
you want to repeat the transcription.
pip install git https://github.com/m-bain/whisperx.git ffmpeg-python pandoc
Also requires pandoc=2.x and ffmpeg installed on system. Alternatively, use
this script on Google
Colab
as it already has these dependencies installed.
$ yt-dlp https://www.youtube.com/watch?v=FAPeFmWmYQo -x -o "audio.%(ext)s"
[youtube] Extracting URL: https://www.youtube.com/watch?v=FAPeFmWmYQo
[youtube] FAPeFmWmYQo: Downloading webpage
[youtube] FAPeFmWmYQo: Downloading ios player API JSON
[youtube] FAPeFmWmYQo: Downloading android player API JSON
[youtube] FAPeFmWmYQo: Downloading m3u8 information
[info] FAPeFmWmYQo: Downloading 1 format(s): 251
[download] Destination: audio.webm
[download] 100% of 8.73MiB in 00:00:00 at 11.48MiB/s
[ExtractAudio] Destination: audio.opus
Deleting original file audio.webm (pass -k to keep)
$ wxscribe.py audio.opus --hf-token hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
Converting audio.opus to audio.mp3
Running recognition on audio.mp3
Lightning automatically upgraded your loaded checkpoint from v1.5.4 to v2.0.7. To apply the upgrade to your files permanently, run `python -m pytorch_lightning.utilities.upgrade_checkpoint --file ../../.cache/torch/whisperx-vad-segmentation.bin`
Model was trained with pyannote.audio 0.0.1, yours is 2.1.1. Bad things might happen unless you revert pyannote.audio to 0.x.
Model was trained with torch 1.10.0 cu102, yours is 2.0.1. Bad things might happen unless you revert torch to 1.x.
Lightning automatically upgraded your loaded checkpoint from v1.5.4 to v2.0.7. To apply the upgrade to your files permanently, run `python -m pytorch_lightning.utilities.upgrade_checkpoint --file ../../.cache/torch/pyannote/models--pyannote--segmentation/snapshots/c4c8ceafcbb3a7a280c2d357aee9fbc9b0be7f9b/pytorch_model.bin`
Model was trained with pyannote.audio 0.0.1, yours is 2.1.1. Bad things might happen unless you revert pyannote.audio to 0.x.
Model was trained with torch 1.10.0 cu102, yours is 2.0.1. Bad things might happen unless you revert torch to 1.x.
Saving to audio.docx
--model-size/--batch-size/--language: Options for the whisper model. Adjust to change how much GPU memory is used and what language the transcribed text should be in.--show-timestamps: Print timestamps before each paragraph--seperate-paragraphs: Show speaker names before all paragraphs even if the same speaker continues speaking.--no-speaker: Do not show speaker names