Step 3 of 7

Audio Preprocessing

Vocal source separation with Demucs (htdemucs)

Last updated: 2026-06-03 00:43
4 Separated stems
28,121 Videos processed
Data Flow

Continuous Update Pipeline

Input Raw audio
Process Demucs htdemucs (GPU vocal separation)
Output Isolated vocal tracks
Methodology

How It Works

Each audio file is processed through the htdemucs model for source separation. The signal is decomposed into four stems (vocals, bass, drums, other). Only the vocal stem is retained for subsequent steps.

Vocal separation runs on the collaborative machine network with GPU acceleration. Processing is parallelized across available compute nodes.

Click each card above to expand details

Audio tracks extracted from videos are processed by the source separation model Demucs (htdemucs model). Demucs decomposes the audio signal into four components (vocals, bass, drums, other) and only the vocal stem is retained. This step removes background music, jingles and parasitic noise, significantly improving the quality of subsequent speech recognition.

Processing is performed continuously on the collaborators' machine network, with GPU acceleration. Each video is processed automatically upon detection by the scanner.

Technology Stack

Tools Used

Demucs (htdemucs)
PyTorch
Data Architecture

Database Schema

Six tables in a normalized relational schema, from raw metadata to sentence-level NLP annotations.

# Table Description Scale
1 videos One row per video: ID, channel metadata, views, likes, comments, tags, duration, upload date, political orientation, country, gender. 26,396 rows
2 comments All comments with author info, like counts, timestamps, nested reply structure, and JSONB analysis column. 9.6M+ rows
3 video_transcripts Full diarized transcripts with speaker labels and cleaned text versions. 28,121 rows
4 transcription_speakers Individual speaker segments from diarization, ordered by position within each video. 1,021,611 rows
5 comments_processed Sentence-level tokenized comments with NER entities (PER, ORG, LOC) and ML prediction columns. 15.3M+ rows
6 transcription_speakers_processed Sentence-level speaker segments with NER extraction and full annotation suite. 4.8M+ rows

Continuous Observatory

The database is continuously updated: channel scanning, video transcription and annotation, comment extraction, metadata updates (views, likes, subscribers). Each scan produces a longitudinal history accessible via the API.

Last updated: 2026-06-03 00:43
Today
videos transcribed
comments extracted
Since January
videos transcribed
comments extracted
videos detected
metadata updated
channels scanned
Technical Paper
Contact the Team
Have a question about the data, the API, or the project? Send us a message.
Suggest a Channel or Feature
Help us improve the YOUPOL corpus. Suggest a political YouTube or TikTok channel we should track, or a feature you'd like to see.