[ Robert Kausch ]
- build: honor an explicit b_lto=false; drop __stdcall on x86 Windows
[ Nils Schimmelmann ]
- Bump the ABI to 2: faac_params_init() takes the caller's struct size
- HE-AAC auto range 8-32 kbps per channel at 44.1 and 48 kHz (was
12-28), narrowing to 8-12 at 32 kHz; VBR up to -q 75 (was 60);
HE-AAC v1 for multichannel layouts up to 8 channels, with one SBR
fill element per SCE/CPE instead of one per frame - Rate control: signed bit account for ABR; a real bit reservoir for CBR
behind--cbr - Refit the ABR bandwidth curve from a sweep: a wider core at 24-56 kbps
stereo and every cutoff on the band grid reachable, where a double
snap had left seven - VBR codes at the top of the bandwidth curve (18750 Hz) instead of the
full band;-qand-bare exclusive;-qdocumented as a constant
quantizer quality, 1..5000, without bitrate anchors that drift - SBR: reconstruction toward 20 kHz; noise floor -18 dB; log-spaced band
table by rate; bs_stop_freq signals a k2 decoders actually accept;
envelopes measured over the slots the decoder applies them to, coded
in 3 dB steps, with two per frame without a transient from 17 kbps
per channel - Stereo: joint-stereo policy by rate; intensity on long windows below
40 kbps per channel, L/R on short windows below the crossover from 80;
short-window grouping and masking targets reworked; the sparsity mute
that zeroed a quiet unjoined channel's bands is gone; M/S codes the
side channel instead of dropping it and quantizes both halves against
the L/R level; long windows take it on every band from 32 kbps per
channel (24 on HE-AAC), short windows and HE-AAC at 20-24 kbps per
channel only below 1-2 kHz - A right-channel intensity band over a left PNS band is coded as PNS,
so decoders that apply intensity first don't play it silent - TNS on by default; frames screened before LPC analysis
- Block switching: the HE core runs its own transient detector instead
of following SBR's, and both cores judge a sub-block's attack against
a running level instead of only its neighbour, so dense stationary
material stops defaulting to short windows - Quantizer: short blocks take the same loudness target as long ones
instead of 0.45x coarser, so transient frames stop losing bands to
PNS; from 56 kbps per channel the treble de-emphasis is steeper, so
the low bands get the margin; bands far below the frame's energy
are allowed coarser targets; kernels take band widths in quads and
vectorize fully, and grouping band energy is summed in four lanes - HUFF2: choose section codebooks jointly across each window group,
accounting for both codeword and section-header costs - Enforce the ISO/IEC 14496-3 per-channel frame ceiling and bitrate
limits; map channel counts to the standard AudioSpecificConfig/ADTS
channel configurations and reject non-standard counts - PNS and the MPEG version are now resolved by the encoder instead of
the caller:faac_paramsdropspns_levelfor ause_pnsflag,
faac_encoder_inforeports the resolvedmpeg_version, and an
explicit MPEG-2 request silently upgrades to MPEG-4 for HE-AAC;
faactakes--no-pnsin place of--pns <0..10> - frontend: MP4 output by default with detection from the extension;
24-bit packed PCM input; 64-bit mdat and duration; track language;
UTF-8 and wide-character paths on CLI and Win32 GUI; iTunes gapless
metadata (iTunSMPB); shared encoding engine and MP4 output in the GUI;
GUI opens on ABR at 128 kbps; optional end-of-encode diagnostics;
a PCM WAV could be read as float from an uninitialised header field;
endian conversion utilities unified, fixing a WORDS_BIGENDIAN check
that only worked because the macro happened to be undefined rather
than 0; help text and the man page corrected throughout - mp4write: container timing, HE-AAC sample rate declaration and gapless
fixes; GUI PCM clipping fix - HE-AAC: pad the pipeline delay to an even multiple of the core frame,
so gapless-aware decoders trim the right sample count at both ends - SBR: parameter resolution in VBR, tail drained on flush, psPresentFlag
for mono, less per-frame work in the writer - Pipeline draining during flush on short streams
- Fix a 64 kHz short-window scalefactor band width table that summed
to 124 spectral lines instead of 128 - Build the lookup tables and FFT twiddle factors shared between
encoder handles once per process instead of per open, so independent
handles can be opened concurrently without racing each other's tables - fft: replace the bit-reversed DIF kernel and its reorder pass with a
self-sorting Stockham kernel that lands in natural order directly - filtbank/channels: inline the window dispatch and simplify grouping
bit accounting for vectorization - Drop the never-selected KBD window from the filter bank
- Drop dead count mode from bitstream writer
- build: force-include config.h instead of per-file guards; drop the
cold/noinline hints on the HE-AAC frame path so PGO and -Os/-O3
make the call instead of a fixed attribute - frontend: show the git commit hash in
-h, the GUI, verbose logging,
and the MP4 tag;-hreports the runtime libfaac version, not the
compile-time one