[tap]

What this changes

PR 1b of the comparison plan (after #73; bridge and rational follow). SpeexDSP's resampler joins the async comparison in both of its arithmetic builds, and r8brain moves from 7.5 to the upstream tip (7.6).

• SpeexDSP, pinned at the SpeexDSP-1.2.1 commit, compiled from resample.c twice (cmake/speexdsp.cmake: FIXED_POINT and FLOATING_POINT under separate RANDOM_PREFIXes) for the host bench, the embedded counts (four new comparison binaries: quality 10 and quality 4 per build, in compare.yml) and the notebook (a two-function shim over both builds, tools/compare_shim/).
• Notebook: two SpeexDSP quality rows, a SampleRateTap Q15 row through the new Q15 C ABI, and a sweep over Speex's quality knob (latency and 20 kHz gain per quality, both builds).
• update_compare_docs.py derives the SpeexDSP rows when the record carries them, the fixed-point rows ratioed against balanced Q15; COMPARISON.md gets the quality rows, a Speex latency section, a per-target reading paragraph, the landscape row, and a refreshed host table; README.md names the new dependency.
• The record (bench/compare_counts.json) is re-harvested from compare.yml run 37970395757 on this branch's first commit, the first run with the Speex binaries; the second commit carries it and the regenerated block.
• r8brain 7.6: upstream's one change is faster filter calculation. Measured as construction only: every steady-state count is identical to 7.5's; construction moves by −0.1 M of 38 M on the M33.

Why

SpeexDSP is the one resampler library with a fixed-point build, so it is the first competitor the Q15 rows can be measured against on their own terms: until now the Q15 datapath's cost tables had no row to be read against, and the quality table had no 16-bit competitor beside it.

Verification

• Quality (notebooks/asrc_comparison.ipynb, committed executed): every existing row reproduces to the displayed precision. SpeexDSP float at quality 10 measures −133.7 dB THD+N / 149.1 dB DR, where balanced sits, not at the format ceiling the other three libraries reach; its fixed-point build −80.2 dB / 97.4 dB beside SampleRateTap Q15 at −76.6 dB / 97.3 dB (the README's "77 dB, format-limited"). The sweep's assertions pin the knee: no quality is flat to 20 kHz (−0.42 dB from quality 4 up), so quality 4 (32 frames) is the matched row beside quality 10 (128 frames).
• Embedded counts: CI-measured (run 22, above). The record reproduces the local M55/M33/Hexagon derivation cell for cell (pinned toolchains, the tree's plugin). SpeexDSP fixed-point at quality 4 costs 1.9× / 2.0× / 1.3× the Q15 datapath (M55 / M33 / Hexagon), quality 10 6.9× / 7.3× / 4.1×. Its float build is single-precision throughout, so on the two soft-double targets it undercuts SampleRateTap's float row; the doc says which row to read there. Every pre-existing steady-state cell is unchanged from Comparison freshness gate, construct-only ratchet legs, async comparison refresh #73's record.
• Host table: re-measured in one session on this host (all 25 subjects, median of 5); the pre-existing ratios held within ~15 % of the 2026-09-25 session, and the reading guide carries the new session's numbers. SpeexDSP is the slowest subject on x86 at every setting (4× to 14× behind the matching SampleRateTap row).
• Gates run locally: retired-words clean, clang-format and clang-tidy clean on the three changed C++ units, update_compare_docs.py idempotent on the committed record, the workflow YAML parses, the book builds, native runs of all four Speex comparison binaries report ok.
• First on CI: the compare.yml harvest with the Speex binaries ran end to end on this branch (green, record above); the Comparison build smoke job is the first CI build of the Speex shim on the hosted compilers.

Notes for the reviewer

• Notebooks re-executed, because the subject set grew. Committed executed.
• New third-party fetch (BSD-3) under the same rule as r8brain: opt-in comparison builds only, never linked into the library or its tests. The two libraries' compile definitions are PRIVATE by necessity, because a translation unit that reads the header for both builds has to define them itself; speexdsp.cmake says so.
• Speex's speex_resampler_* names are macros over RANDOM_PREFIX, expanded at the use site, so the shim and the bench call the prefixed symbols directly and undefine the include guard between the two reads of the header. It looks odd; the comment at each site explains why.
• The "0.1×" cells for Speex's float build on the M33 and Hexagon are honest: SampleRateTap's float datapath accumulates in double and is soft-float there, which is why the README directs those targets to Q15. The derived table keeps the one rule (float rows vs. balanced float, fixed rows vs. Q15) and the prose carries the reading.
• Provenance toolchain string now comes from the runner's package versions (15:13.2.rel1-2 is arm-none-eabi-gcc 13.2.1); the prose above the block says the same thing in the human form.

🤖 Generated with Claude Code

https://claude.ai/code/session_01G3HxzEGiZp7jStoMuYNuBK