Sound for travel films

Picture got most of my complaints. Sound got the most expensive ones.

A score written to the cut

The background music is generated per film from the edit plan, not laid on top of it. Chord changes land on the beat nearest to a cut, parts of the film can run at different tempos, and the plugins from creative-music-with-claude mix it with a sidechain on the voice track, so the music steps back under speech by itself.

Two things surprised me:

• Mood is the scale degrees, not the tempo. My first major score came out sad. The tempo was fine; the chords were the problem. Walks alone now get major sevenths and a minor sixth, films with my daughter get triads with an added sixth and no minor chord at all.
• Where something is played live, there is no score. Not quieter: a hole. Inside the window the score sits at −91 dB against −13.6 at full level. You hear music ducking. You do not hear a hole.

If someone plays in the footage, the harmony can be taken from the recording and fed to the score (the audio-to-midi skill from project-superduper-dsp). On a noisy street recording three models agreed on the key and disagreed on the melody, so it recovers chords, not a tune.

Live music in the frame stays one continuous piece under the whole scene, with the picture cuts on top. Cut it per shot and the beat restarts at every cut; I called that version garbage the moment I heard it.

Other people’s music

Car radio, shop speakers, a festival stage. One California episode got 10 Content ID claims, about four minutes of other people’s songs, and four of the claims blocked the video in two countries, 21% of my audience. A claim is not a strike, but you do not get to pick the block.

The fix runs before upload:

demucs -o stems -d mps film_audio.wav   # 4 stems, ~4 min per 20 min of audio on Apple Silicon

-d mps matters: without it Demucs stays on the CPU, about half the speed. The solo:stem-split skill has the install line and the variants.

• Detect radio by drums and bass, not “everything but the voice”. Demucs puts the engine, wind and city into other; a radio has a rhythm section. The first detector read the no-vocals stem and was blind.
• Keep the vocal stem whole. The first version passed my voice only inside subtitle windows. I noticed the same day: the emotions on a bridge were gone. Speech recognition does not hear exclamations; 31 seconds of loud speech lived outside the subtitles. The singer’s ghost stays in the stem 5 to 10 dB under my voice and disappears under the score.
• The picture is not touched. Only the audio is rebuilt, the video goes -c:v copy.
• YouTube’s claims are a check, not the map. Content ID sees a subset. Clean everything the detector found, then confirm it also covers the windows YouTube flagged.

An already published video with a claim that changes nothing stays as it is. Re-uploading pays only when there is a block. Instagram trailers keep the radio: no monetisation there to lose.

Picture side: creative-video-agent-cut. Platforms: creative-video-shipping. Transcripts: creative-video-transcribe. The rules and ffmpeg recipes are in the solo:video-edit skill (solo-factory).