The latest articles on DEV Community by N D (@n_d_87a2e0ec24f0923d167cb). https://dev.to/n_d_87a2e0ec24f0923d167cb https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F2697676%2F43f824ec-3da7-4762-8464-93d2cf510ae0.png https://dev.to/n_d_87a2e0ec24f0923d167cb en N D Fri, 27 Mar 2026 14:39:37 +0000 https://dev.to/n_d_87a2e0ec24f0923d167cb/building-an-ai-profanity-filter-with-vocal-separation-3m24 https://dev.to/n_d_87a2e0ec24f0923d167cb/building-an-ai-profanity-filter-with-vocal-separation-3m24 <p>I built an online tool that automatically detects and bleeps profanity in video and audio files. Here's the high-level architecture.</p> <h2> The problem </h2> <p>Manual profanity censoring takes 45+ minutes for a 10-minute video. You have to listen through, find each word, razor the audio, drop a beep effect. For songs, it's nearly impossible without destroying the music.</p> <h2> The solution </h2> <p>AI speech recognition + neural vocal separation.</p> <h2> How it works </h2> <ol> <li>User uploads a file or pastes a YouTube URL</li> <li>Audio is extracted with FFmpeg</li> <li>AI speech-to-text transcribes the audio (AssemblyAI / Deepgram)</li> <li>Profanity is detected using morphological analysis (lemmatization)</li> <li>Each word is replaced with beep/silence/custom sound via FFmpeg</li> <li>For songs: Demucs AI separates vocals from instruments first</li> </ol> <h2> Song mode — the hard part </h2> <p><a href="https://github.com/facebookresearch/demucs" rel="noopener noreferrer">Demucs</a> by Meta AI does the heavy lifting — splitting audio into vocal and instrumental tracks. Profanity detection runs only on the vocal track, then the censored vocals are mixed back with the original instruments. The music stays untouched.</p> <h2> Stack </h2> <ul> <li> <strong>Frontend</strong>: Next.js (React)</li> <li> <strong>Backend</strong>: NestJS (Node.js), BullMQ queues</li> <li> <strong>Audio processing</strong>: Python (FastAPI), Demucs, FFmpeg</li> <li> <strong>Infrastructure</strong>: Docker Compose, PostgreSQL, Redis</li> </ul> <h2> Results </h2> <p>12,000+ files processed. Three processing modes: standard (clean speech), precise (noisy audio), enhanced (songs with vocal separation).</p> <p>Free for up to 15 minutes per month at <a href="https://videocensor.net" rel="noopener noreferrer">videocensor.net</a>.</p> <p>Would love to hear your thoughts!</p> ai showdev python