knik0/faac faac-2.0

FAAC 2.0

on GitHub

[ Nils Schimmelmann ]

• Harden input parsing and guard against allocation failures.
• Replace the classic faacEnc* C API with a new faac_* API in
  faac.h: parameters supplied once to faac_encoder_open(), uniform
  faac_status error reporting, resolved-property queries, and
  library-owned AudioSpecificConfig. The legacy faacEnc* surface and
  faaccfg.h are removed; the encoder core they exposed is now internal.
  SONAME bumped to 1 (libfaac.so.1). faac.h now defines
  FAAC_VERSION_MAJOR/MINOR/PATCH, tracking the SONAME, so callers
  can detect the new API at compile time (FAAC_VERSION_MAJOR >= 1).
• frontend: Port faac and the Win32 GUI to the new API; this fixes the
  AudioSpecificConfig memory leak on the MP4 output path.
• Add HE-AAC v1 (AAC-LC core + Spectral Band Replication) encoding via
  FAAC_OBJ_HE_AAC_V1, or FAAC_OBJ_AUTO, which picks LC or HE-AAC from
  the bitrate. The SBR core runs dual-rate (half-rate core, 2048-sample
  frame); faac_encoder_get_info() reports the full output rate and frame
  size so callers need no SBR-specific handling. CLI:
  --object-type lc|he-aac-v1|auto.
• stereo: Add Mixed Mode joint stereo (--joint 3) as the new default:
  per scale-factor band, picks IS above 5.5 kHz, M/S where channels are
  correlated, or L/R otherwise. Fix a midside band-silencing bug that
  produced L=R=0 on decode, guard against negative rounding before
  sqrt, and keep IS available at low sample rates.
• quantize: Restore short-window masking penalty, floor masking
  targets in quiet bands, drop stale tonal scaling in bmask, and keep
  global_gain a valid regular scalefactor -- fixing energy drops,
  quiet-band artifacts, and a decoder out-of-range error.
• tns: Cap TNS filter order to 12 to fix AAC-LC decoder rejections and
  audio dropouts in FFmpeg/FAAD.
• Retire the never-implemented MAIN/SSR/LTP object types across
  filtbank, frame, and the frontend, and collapse the single-model
  psychoacoustic vtable into direct calls.
• Fix a lookahead regression that had truncated the encoder's
  lookahead to an effective 1-frame delay, causing pre-echo artifacts
  and broken stereo coherence; then simplify the lookahead into a
  unified per-channel FIFO and energy timeline, replacing the ad hoc
  sampleBuff/next*SampleBuff/overlapBuff/prevSamples buffers.
• Relicense: rewrite stereo, quantize, channels, bitstream, filtbank,
  tns, huffdata, huff2, and the MP4 output code as original work under
  LGPL-2.1-or-later, and drop the matching MPEG reference-code preamble
  from README.md, --license, and docs/. Every source file is now free
  software. Update the license header itself from GNU Library GPL to
  Lesser GPL wording.
• fft: Replace the radix-2 FFT with radix-4 decimation-in-frequency as
  the sole implementation and precompute MDCT twiddles, for ~18%
  faster encode versus the radix-2 baseline.
• Standardize libfaac on single-precision floating point: drop the
  floating-point Meson build option and faac_real.h, and use
  float/sinf/cosf/sqrtf directly throughout the per-sample hot
  path. One-time init-time table generators (FFT/MDCT/SBR twiddles,
  window shapes, quantizer limits) keep double precision, where it's
  free. Benchmarked MOS-neutral versus the prior double-precision path.
• blockswitch: Move short-window detection to the time domain.
• Performance passes: reuse a shared work buffer for filtbank's MDCT
  scratch, and optimize TNS and stereo IS-gate hot paths.
• frame: Buffer encoder input so any input chunk size works.
• bitstream: Stop embedding the libfaac version string.
• build: Fix compiler warnings and transition to C11.
• frontend: Add a --creation-time flag for configurable MP4
  timestamps, vendor mingw-w64 wingetopt as a minimal-diff getopt
  fallback for MSVC builds, and keep TNS off by default in the GUI to
  match the CLI.

[ Fabian Greffrath ]

• Remove obsolete pkg-config template
• Restrict LTO to the shared library only

[ Orbis Security AI ]

• filtbank: Fix an out-of-bounds memcpy in FilterBank() that used
  FRAME_LEN instead of BLOCK_LEN_LONG for the overlap buffer copies