SpeakerSplitter

This project is an audio processing script that separates speakers in a conversation into different audio channels.

Basic workflow

• Create session folder with session_name
• Ensure name of audio file (either wav or mp3) is also session_name.wav or session_name.mp3
• Place audio file session_name.wav into session folder sesesion_name
• Place audio diarization json file into session_name folder, titled output.json
• Run this tool in the parent directory, where all of the subdirectories are the names of the sessions

Diarization Model

The model is based on a pre-trained speaker diarization pipeline from the pyannote.audio package, with a post-processing layer that cleans up the output segments and computes input-wide speaker embeddings. Here is the model on replicate

Input Handling

Takes a session folder name as input Loads JSON metadata containing speaker segments Accepts either WAV or MP3 audio files

Audio Processing

Converts timestamp formats (H:M:S) to milliseconds Creates two separate channels (left and right) for different speakers Speaker A's audio goes to the left channel Speaker B's audio goes to the right channel Handles overlapping timestamps to prevent audio conflicts

Audio Enhancement

Normalizes the audio to maintain consistent volume levels Creates a stereo track by combining both channels Exports the final result as a WAV file with separated speakers