Expand description
Learned audio heads — loaded via a common P64 loader that fails closed (abstains /
NeedsWeights) when weights are absent. No Candle/Burn runtime, no Python. Real weights are
gated on the principal (HA1 corpus, HA6 speech model). See ADR 007 + delivery plan Waves 7–8.
Modules§
- aed_p64
- Acoustic-event detection head (P64, fail-closed). Re-exports only (AU-LEARNED).
- audio_
tokens - RVQ / FSQ discrete audio token codec (quantizer ABI). Re-exports only (AU-GEN).
- crepe
- CREPE pitch head (P64, fail-closed). Re-exports only (AU-LEARNED).
- loader
- Common P64 learned-head loader — fails closed (
BackendUnavailable) without weights. Re-exports only (AU-LEARNED). See ADR 007. - separation
- Source separation (classical masking now; learned = NeedsWeights). Re-exports only (AU-GEN).
- speech_
p64 - Speech CTC encoder head (P64, fail-closed). Re-exports only (AU-LEARNED).