FSP Finder runs each track through a short pipeline of audio and speech-recognition models: it isolates the vocals, transcribes them word-by-word, checks each word against a filter list, and silences only the flagged moments before recombining the song. The goal is precise, predictable edits — and a human review step before anything is applied.
The uploaded file is split into a vocal track and an instrumental track using Demucs, a deep-learning source-separation model (the mdx_extra variant). Isolating the vocals is what makes accurate transcription possible — a loud beat can't mask a flagged word, and a clean instrumental line won't be mistaken for one.
The isolated vocals are transcribed word-by-word, with timestamps, using an adapted version of OpenAI's Whisper model (whisper-medium.en) fine-tuned with LoRA weights on timestamped lyric data from the DALI dataset — so it handles sung vocals better than a stock speech model. More detail on the fine-tuning is on the project's GitHub page.
Any stretch of vocals that comes back untranscribed but still contains audio (average level above −30 dB) is re-run through Whisper with more generous settings to try again. Anything that still can't be transcribed after the second pass is flagged with a visible warning on the review page, so you can check it by hand.
To keep the transcription focused, Silero VAD identifies which parts of the audio actually contain speech. This speeds up processing and improves how well quieter or overlapping vocals get picked up. VAD is on by default and can be turned off in advanced options.
Each transcribed word is checked against a filter list of known curse words and multi-word explicit phrases. You can add your own terms in advanced options, and you always get the final say on the review page — nothing is censored that you don't approve. Detection is filter-list based and predictable, not a guess.
Once the flagged segments are confirmed, only those moments are silenced within the isolated vocal track. The edited vocals are then recombined with the original instrumental, losslessly, and the original file's metadata is carried over. All processing is done on lossless .wav audio; you can export the result as WAV, FLAC, or 256 kbps MP3.