Encoder: transient-driven window switching (long/start/short/stop, grouping),
sine/KBD MDCT, MDCT-domain psychoacoustic model (Bark spreading, SFM tonality,
ATH, PE), TNS, per-band M/S, lambda rate-distortion loop with measured-error
scalefactor search and DP sectioning over all codebooks, bit reservoir with
fill elements (exact CBR), VBR quality scale, 1-8 channels (configs 1-7 + PCE
for 7ch), ASC + optional ADTS, 1024-sample priming reported.
Decoder: full AAC-LC syntax incl. PNS (correlated M/S noise), intensity stereo,
TNS, M/S, pulse data, both window shapes, PCE/DSE/FIL, ADTS helpers; never
panics on malformed input.
Tests use ffmpeg only as an oracle: our streams decode cleanly in ffmpeg
(CBR within 0.1%), our decoder matches ffmpeg to ~1e-7 on ffmpeg-encoded
streams (8-96 kHz, mono..7.1, IS/MS/TNS), plus fuzzing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
c2872fbd8d