AudioSlicer (Editor) for ai-voice-cloning by mrq

Requirements (just use the same venv as for the ui)

• ffmpeg (ffprobe)
• PyQt6
• ffprobe-python

Clone this repo to "modules". To install the requirements, activate your venv (venv\Scripts\activate on Windows) and run pip install ffmpeg-python PyQt6 ffprobe-python.

Usage

Run main.py (with the activated venv)

Select either a generated audio_list.json (by this tool to continue working on the same project) or a generated whisper.json (by the ui with whisperx) in the traning folder of the voice.

Note: It is adviced to give a -1s trim offset for the start, whisperx cuts often to late, and you can cut it later with this tool.

Buttons

• prev / next (alt+left / alt+right): go to the previous / next audio file
• delete (alt+down): delete the current audio file (not really delete but flag as not to export for training)
• play (alt+up) palys the segment
• trim (alt+t): trim the current audio file (respecting the 2 fields above, cuts x seconds from start/end - floats are allowed), after cutting, the length can be longer since ffprobe reports the actual length and not the weird whisper timestamps
• trim_undo: undos the previous trim: ONLY WORKS ONE TIME (could be improved)
• ffprobe returns the actual new length
• save (ctrl+s): save the current progress to audio_list.json
• export: writes the train.txt and vaidation.txt (2% of the "good" samples taken)

Notes

• To add new audio files, just transcribe them via the UI and then select the audio_list.json, it will add the new files to the list (takes a bit) DONT SELECT THE WHISPER.JSON AGAIN, IT WILL OVERWRITE THE WHOLE LIST, AND THEREFORE YOUR PROGESS. Also make sure to check the skip existing in the webui, otherwise the trimmed audio files will be overwritten, and the trim is gone.
• First import will be slow (or adding new files) - each file will be probed by ffmpeg for the real length
• The word quality is taken from whisperX, therefore its only an indicator nothing more.
• Every sentence including ... is auto-excluded from traning (can be changed again - but thats why the progressbars aren't at 0 from the beginning and the "good/deleted segments" bar wont represent the actual outcome until a good amount of segments are flagged
• After hearing a segement, it gets auto flagged as "okey" and therefore marked as good to train

Todo / Improvements

Waveform preview with two handles to trim visually - idk how yet