Same v5 bitstream as 1.2.0 — files are fully interchangeable in both
directions (verified with MD5 checks: 1.2.0 code decodes 1.3.0 files
and vice versa).
The engine is rebuilt for speed and memory:
- FFT autocorrelation replaces the O(n^2) lag scan in LPC analysis
- Pruned LPC order search, windowed Rice parameter search
- Multithreaded segment coding (numba kernels run with nogil)
- Streaming encode: constant RAM regardless of file length
- Tight int16/int32 decode buffers; decode-to-WAV streams per segment
- Speed profiles:
--fast(pure LPC+Rice),--normal(light NLMS),
--max(full cascade, default — same compression as 1.2.0,
byte-identical size on the demo corpus)
Measured on the reference machine (numba, 8 threads):
| clip | encode max | encode fast | decode max | decode fast |
|---|---|---|---|---|
| 44.1 kHz / 16-bit | 21.6x RT (1.2.0: 10.8x) | 26.4x RT | 47x RT | 516x RT |
| 192 kHz / 24-bit | 6.3x RT (1.2.0: 5.5x) | 9.1x RT | 4.8x RT | 67.8x RT |
Fast profile costs ~2-3 pp of compression ratio vs max.
ALAFC_1.3.0_windows.zip — standalone Windows binaries (numba
included). For the .py version: pip install numpy numba.