Sound class taxonomy
Every catalog entry has a sound_class field that informs validator
thresholds, mastering preset selection, and pipeline routing. There
are 7 classes:
synth-pure
Pure mathematical synthesis — binaural beats, isochronic pulses, solfeggio tones, single-frequency drones.
- What it sounds like: a sine, a square, a binaural beat. Frequency is the entire content; if you change it, the psychoacoustic effect breaks.
- Validator tuning: disable
intent-driftand loosenspectral-anomaly(anything outlier IS an artifact, but CLAP doesn’t understand pure tones). - Mastering:
studio-binaural-pureONLY. No EQ, no compression, no reverb. - A/B eligibility: yes — the studio chain often adds harmonic warmth that pure synthesis misses.
- Pipeline preference: DSP only. Never AI-generate.
synth-ambient
Pads, drones, evolving textures — synthesised but not single-frequency.
- What it sounds like: a warm pad, a slow drone, a synth bed.
- Validator tuning: loosen
intent-drift; standardspectral-anomaly - Mastering:
studio-pad-ambientorstudio-drone-deep(depends on register). - A/B eligibility: yes.
- Pipeline preference: DSP first; ElevenLabs SFX as fallback if DSP feels flat.
cultural-instrument
Singing bowls, gongs, bells, chimes — sounds made by physical objects with complex resonance.
- What it sounds like: that’s clearly a Tibetan singing bowl.
- Validator tuning: standard, but per-pattern null for known legitimate tail content.
- Mastering:
studio-bell(preserves transient + tail). - A/B eligibility: yes for the mastering chain only.
- Pipeline preference: Freesound CC0 strongly preferred. DSP attempts (additive synthesis) can land but rarely match a real recording.
animal-specific
Cat purrs, dog breathing, bird songs, owl hoots — animal vocalisations.
- What it sounds like: that’s clearly a real cat. (Or it’s wrong.)
- Validator tuning: loose. Animals do unpredictable things.
- Mastering: light only — preserve realism.
- A/B eligibility: no.
- Pipeline preference: Freesound CC0. ElevenLabs SFX as fallback if no good source.
nature-ambient
Rain, wind, streams, fire crackle, ocean waves — field-recording-style environmental textures.
- What it sounds like: I’m at the actual location.
- Validator tuning: loose
spectral-anomaly(legitimate transients abound). - Mastering: light — preserve naturalness.
- A/B eligibility: no.
- Pipeline preference: Freesound CC0 strongly preferred.
voice-narration
Affirmations, guided meditations — TTS or recorded human voice.
- What it sounds like: someone is calmly telling you something.
- Validator tuning: null
intent-drift(CLAP doesn’t understand TTS content); standard others. - Mastering: voice-specific chain (hi-pass 80 Hz, sibilance EQ, multi-band, LUFS -20).
- A/B eligibility: no.
- Pipeline preference:
tts-affirmationspipeline only.
mechanical-ambient
Train rumble, plane white noise, household appliances — mechanical sources that produce ambient texture.
- What it sounds like: I’m in a moving vehicle / near an appliance.
- Validator tuning: standard.
- Mastering:
studio-modularorstudio-drone-deepwork well. - A/B eligibility: yes.
- Pipeline preference: Freesound CC0 or ElevenLabs SFX.
Setting sound_class on a new track
The classify_sound_class.py --apply script (in tools/curate/)
uses Gemini Flash to suggest a sound_class for any catalog entry
without one. Re-run it after every catalog rebuild because
catalog.py overwrites the field.