Speech Condenser

Speech condenser is a tool for reducing the size of a dialogue.

Pipeline

It combines several tools to achieve the goal of reducing the size of a dialogue. Each step of the above pipleine runs inside a container.

Steps:

1. Audio extraction - Extracts the audio from the video file.
2. Speaker diarization - Identifies the speakers in the audio file.
3. Split audio - Splits the audio file into smaller chunks based on the speaker diarization.
4. Speech to text - Transcribes the audio chunks into text.
5. Combine ASR and diarization - Combines the results of the ASR and diarization to get the text for each speaker as a dialogue.
6. Summarization - Summarizes the dialogue.

Installation

The setup uses docker or podman to run the containers. A set of local scripts are provided to run the pipeline.

• build.sh - Builds the containers.
• pipeline.sh - Runs the pipeline.
• yt-pipeline.sh - Runs the pipeline on a youtube video.

Videos needs to be provided in the data/input directory. yt-pipeline.sh will use this directory to download to cache the video. The output will be in the data/output directory.

Make sure to create a .env based on the .env.example file and privide the required values:

• SC_RUNTIME - The runtime to use for the containers. Either docker or podman.
• HF_TOKEN - The Hugging Face token to use for the summarization step.

Make sure to visit hf.co/pyannote/speaker-diarization and hf.co/pyannote/segmentation and accept user conditions. This required in order to be able to run the speaker diarization.

Usage

Run agains a local video file:

./pipeline.sh "data/input/video.mp4"

Run against a youtube video:

./yt-pipeline.sh "https://www.youtube.com/watch?v=video_id"