ITU compliance Stage 1: the calibrated signal layer

.gitignore

@@ -7,3 +7,7 @@ CMakeUserPresets.json
 .claude/
 book/book/
 __pycache__/
+
+# ITU real-speech test vectors: local test use only, never committed
+# (download from the ITU test-signal database; see docs/itu-compliance.md)
+tests/data/itu/

HANDOFF.md

@@ -263,13 +263,29 @@ stand-ins marked method-equivalent; (7) P.502 still unobtained —
 analysis methods reconstructed from the P.340/P.111x test descriptions.
 Rec PDFs live in the session scratchpad only (never committed).*
 
-**Stage 1 — Calibrated ITU signal layer** (`tests/support/`). Level
-calibration conventions (dBov/activity-gated levels), a deterministic
-P.501 CSS generator, the P.501 double-talk sequences, per-bandwidth
-band-limiting, driving-noise generator + car-cabin echo paths for the
-P.1100 series (modeled first; Tim's measured cabin RIRs can join via the
-existing fixture pipeline). Pass: generated signals match published
-spectral/temporal properties.
+**Stage 1 — Calibrated ITU signal layer** (`tests/support/`). *DONE —
+`itu_levels.h` (dBov/dBm0/dBPa conventions, prewarped-bilinear
+A-weighting, 35/5 ms integrators, sliding peak, P.56-style active-level
+meter) + `itu_signals.h` (CSS single/double-talk with the P.501
+Table 7-1/7-2 voiced segments transcribed from the PDF, 8192-pt PN
+adaptive-systems variant, shaping + band-limit filters, AM-FM
+orthogonal pair + comb analysis, activation sequence, Hoth + synthetic
+driving noise, moving-reflector time-variant path) +
+`tests/fixtures/rir_cabin.h` (image-source cabin, RT60 66.6 ms) +
+`test_itu_signals.cpp` (12 tests, measured-first). KEY DECISION made in
+stage: the suite runs natively at 44.1 kHz (CSS is sample-exact only
+there; matrix updated). Measured instrument floor worth knowing: the
+AM-FM comb separation is 94.3 dB at the spec's exact band edges — and
+collapses to 19 dB if any guard band is added (recorded in the test).
+Existing fixtures verified bit-reproducible with pyroomacoustics 0.10.1
+before generating the cabin (--only flag added to the generator). Still
+open from this stage: the ITU real-speech attachment WAVs
+(tests/data/itu/, git-ignored) for the three signal-exact rows.
+AMENDED (Tim's directive): 48 kHz and 16 kHz are REQUIRED operating
+rates (44.1 kHz stays the P.501-native reference) — the NOTE 2
+resampler was built (ripple <= 0.014 dB, alias rejection 101 dB), the
+formula generators take fs directly, and make_css_at() emits
+sample-exact periods at every required rate; matrix policy updated.*
 
 **Stage 2 — Residual-echo post-filter + comfort noise.** The big DSP
 item. New header (working name `mutap/postfilter.h`): coherence-based

docs/itu-compliance.md

@@ -65,9 +65,25 @@ approach:
   detector-free double-talk adaptation is MuTap's measured strength.
 - **Bandwidth: wideband (P.1110) and SWB/FB (P.1120) are primary**;
   narrowband (P.1100) variants run as band-limited configurations of the
-  same tests. Everything executes at 48 kHz with P.501-specified
-  band-limiting of the receive-direction signals (NB 3.6/4 kHz, WB
-  7.2/8 kHz, SWB 14.4/16 kHz, FB 20 kHz low-pass).
+  same tests, with P.501-specified band-limiting of the receive-direction
+  signals (NB 3.6/4 kHz, WB 7.2/8 kHz, SWB 14.4/16 kHz).
+- **Sample rates (revised twice; current policy per Tim's directive):
+  the compliance suite RUNS AT — and every Tier A/B row must pass at —
+  48 kHz and 16 kHz, with 44.1 kHz kept as the P.501-native reference
+  rate.** Rationale per rate:
+  - **48 kHz (required)**: MuTap's operating rate, the automotive recs'
+    specified analysis rate (8k FFT @ 48 kHz), and the rate at which
+    the RIR fixtures are exact.
+  - **16 kHz (required)**: the wideband telephony rate (AMR-WB/P.1110
+    world). SWB/FB-only clauses are exempt at 16 kHz (they exceed its
+    8 kHz Nyquist); the AM-FM plans (<= 7.04 kHz) fit.
+  - **44.1 kHz (reference)**: P.501's native rate, where the CSS
+    framing is sample-exact without conversion; used to validate the
+    signal layer itself.
+  The bridge is the P.501 NOTE 2 resampler (Stage 1): measured passband
+  ripple <= 0.014 dB (spec < 0.2) and alias rejection 101 dB (spec
+  > 60); the CSS period stays sample-exact at every required rate
+  (15435 / 16800 / 5600 samples per 350 ms at 44.1 / 48 / 16 kHz).
 - **The simulated echo paths**: the three committed image-source rooms
   (fixtures) + a new car-cabin family (small volume ~2.5 m^3, RT ~60 ms
   per G.167 5.2.3.1's car figures) + the P.1110/P.1120 time-variant-path
@@ -238,7 +254,37 @@ implementation budget.
 
 ---
 
-## Simulation assets required (feeds Stage 1)
+## Simulation assets delivered by Stage 1
+
+`tests/support/itu_levels.h` + `tests/support/itu_signals.h` +
+`tests/fixtures/rir_cabin.h`, validated by `tests/test_itu_signals.cpp`
+(measured-first thresholds; the file header carries every number).
+Instrument floors that bound what the compliance suite can measure:
+
+- CSS single/double-talk: sample-exact framing, PN crest 10.97 dB
+  (spec 11 +- 1), bin-flat PN, level calibration exact, voiced segments
+  transcribed from P.501 Tables 7-1/7-2. Available at every required
+  rate via `make_css_at` (PN crest 11.86 dB at 48 kHz / 11.44 at
+  16 kHz; AM-FM separation 94.0 / 92.6 dB at 48 / 16 kHz).
+- AM-FM orthogonal pair: **94.3 dB comb separation** at the spec's exact
+  band edges — the double-talk echo-loss measurement floor sits far
+  beyond the >= 33 dB margin targets. (Analysis discipline recorded in
+  the test: any added guard band collapses the floor.)
+- A-weighting: -19.18 / 0.00 / -1.76 dB at 100 / 1k / 10 kHz
+  (prewarped bilinear; IEC class-1 envelope).
+- NB band-limiter: -85 dB at 4 kHz. Cabin fixture: RT60 66.6 ms.
+- P.56 active-level meter: definitionally consistent on sparse speech
+  (activity 0.312, delta = 10 log10(1/activity) exactly); on CSS the
+  101 ms pause sits inside the 200 ms hangover, so the G.168
+  active-part constants (+1.49 / +1.66 dB) are applied arithmetically.
+
+**ITU real-speech attachments** (P.501 7.3.2/7.3.3/7.3.5): still to be
+procured — download from the ITU test-signal database into
+`tests/data/itu/` (git-ignored; local test use only, never committed).
+Until then the affected rows run on the CSS/AM-FM signals and are
+reported `method-equivalent`.
+
+## Simulation assets required (original Stage 1 list, for reference)
 
 **Signals (all generatable from P.501's algorithmic descriptions):**
 - CSS single-talk, fullband: voiced 48.62 ms (literal sample table 7-1,

tests/CMakeLists.txt

@@ -37,7 +37,8 @@ add_executable(mutap_tests
     test_adaptation_control.cpp
     test_closed_loop.cpp
     test_rir_fixtures.cpp
-    test_aec.cpp)
+    test_aec.cpp
+    test_itu_signals.cpp)
 target_link_libraries(mutap_tests PRIVATE
     MuTap::MuTap
     mutap_warnings)