audio: find libsoxr on Apple Silicon, and stop the unfiltered decimation

[CryptVenture]

Depends on #358. This branch is stacked on it, so the diff shown here includes that commit. Review the second commit only, or merge #358 first. The test writes its probe as float32 — a 16-bit container would put a −96 dBc floor under a −117 dBc measurement.

What this changes

Two resampling defects that compound, and both cost quality silently.

1. libsoxr is never found on Apple Silicon

SoxrApi loads libsoxr with dlopen("libsoxr.dylib") — a bare leaf name. On Apple Silicon dyld's default search path covers /usr/local/lib and /usr/lib but not /opt/homebrew/lib, so a Homebrew libsoxr is never found and resample_mono_soxr_or_linear falls back to linear interpolation. The fallback warning goes through the debug logger, which defaults to disabled, so the degradation is invisible.

Adds absolute-path candidates for the common install prefixes plus an AUDIOCPP_SOXR_LIBRARY override, and a one-time stderr notice on fallback so it can no longer happen unnoticed.

2. The enhancement-model input path has no anti-alias filter at all

read_mono_resampled — the input path for every denoise and super-resolution entry point — called resample_mono_linear unconditionally.

Measured against the real sources, a 12 kHz tone decimated 48 → 16 kHz for a 16 kHz model:

Path	Folded 4 kHz alias
resample_mono_linear (previous)	0.00 dBc
sinc width 6	−55.06 dBc
sinc width 64	−117.08 dBc
soxr	−288.17 dBc

0.00 dBc is not a typo. 48 → 16 kHz is an exact 3:1 ratio, so the interpolation fraction is identically zero and linear interpolation degenerates into plain sample-dropping — the alias arrives unattenuated. FlashSR is then asked to re-synthesise the band that was just destroyed.

The default that was deliberately not changed

TorchaudioSincHannResampleOptions::lowpass_filter_width stays at 6. That is torchaudio.transforms.Resample's own default, and the ~36 call sites taking it are feature-extraction paths whose contract is bit-parity with a Python reference — widening it would change model inputs everywhere. Per CONTRIBUTING's framework-module rule, the capability was added as torchaudio_sinc_hann_playback_options() (width 64) and only paths producing audio a listener hears were switched over: the audio-utility input path and the mix bus.

Width 64 chosen from the measured knee on a 44.1 → 48 → 44.1 round trip:

Width	SNR	Time (4 s mono)
6	60.54 dB	0.56 ms
64	122.96 dB	6.51 ms (615× realtime)
128	141.42 dB	15.81 ms

Also adds resample_mono_soxr_or_sinc — soxr when available, else the in-tree windowed sinc, with output_length_policy honoured on the fallback (the linear path ignored it).

Validation

Build

cmake -S . -B build -DENGINE_BUILD_TESTS=ON
cmake --build build

Test

ctest -R audio_resample_quality_test

tests/unittests/test_audio_resample_quality.cpp asserts the folded 4 kHz alias is at or below −60 dBc by coherent single-bin DFT. That threshold fails the old linear path by 60 dB and fails a regression to the width-6 default, while passing both the sinc fallback and soxr — so it does not depend on whether libsoxr is installed on the build machine. It also pins the shared defaults so a future change to them is deliberate.

Backend tested: CPU (host DSP). Full suite 40/40 on macOS/Metal, Apple M4 Max.

Affects

Output changes for deepfilternet2, rnnoise, zipenhancer and flashsr whenever the source rate differs from the model rate — that is the fix. Same-rate input is a no-op fast path in both old and new code, so 16 kHz → 16 kHz and 48 kHz → 48 kHz are untouched. Output length may differ by ±1 sample (sinc sizes with ceil, the old linear path with llround).

Any process that hits a soxr fallback now prints one line to stderr, once per process.

🤖 Generated with Claude Code

https://claude.ai/code/session_01ATa5YkLUPMDPRL7w1gCo9p

[CryptVenture]

Closing for now to stay inside the 3-concurrent-PR policy (see the discussion on #422). This one has also gone stale against main; I will rebase it and resubmit once a review slot is free.