Refactor(audio): Move TX Voice DSP to 48 kHz Float — Principle VIII.

CMakeLists.txt

@@ -655,6 +655,7 @@ set(CORE_SOURCES
     src/core/AetherDspModePolicy.cpp
     src/core/AudioEngine.cpp
     src/core/TxMicChannelNormalizer.cpp
+    src/core/TxVoiceProcessor.cpp
     src/core/ChannelStripPresets.cpp
     src/core/Biquad.cpp
     src/core/StereoBiquad.cpp
@@ -3655,6 +3656,13 @@ target_include_directories(tx_mic_channel_normalizer_test PRIVATE
 target_link_libraries(tx_mic_channel_normalizer_test PRIVATE Qt6::Core)
 add_test(NAME tx_mic_channel_normalizer_test COMMAND tx_mic_channel_normalizer_test)
 
+add_executable(tx_voice_processor_test
+    tests/tx_voice_processor_test.cpp
+)
+target_include_directories(tx_voice_processor_test PRIVATE src)
+target_link_libraries(tx_voice_processor_test PRIVATE aethercore Qt6::Core)
+add_test(NAME tx_voice_processor_test COMMAND tx_voice_processor_test)
+
 # Pins the SkyRoof-parity WFM DSP chain: NCO offset correction removes the
 # discriminator DC term (fixed pan + Doppler-stepped slice), twin linear-phase
 # resamplers deliver exactly 48 kHz from any native DAX IQ rate, and streaming
@@ -5665,6 +5673,7 @@ add_executable(device_diagnostics_test
 )
 target_include_directories(device_diagnostics_test PRIVATE src)
 target_link_libraries(device_diagnostics_test PRIVATE Qt6::Core)
+add_test(NAME device_diagnostics_test COMMAND device_diagnostics_test)
 
 add_executable(midi_settings_test
     tests/midi_settings_test.cpp

docs/architecture/pipelines.md

@@ -89,17 +89,18 @@ TX AUDIO ROUTING SUMMARY:                  ◄── AUDIO THREAD
               │                       └─→ radio-native VITA PCC 0x0123
               │
               ▼
-      test tone → client TX DSP chain → post-DSP monitor
-        → PC mic gain → Quindar → final limiter → meters/scopes
+      device-rate Int16 → canonical mono → 48 kHz float normalization
+        → RN2 → test tone → client TX DSP chain → PC mic gain
+        → Quindar → final limiter → 48-to-24 kHz SRC
+        → linked TPDF + Int16 quantization → monitors/meters/scopes
               │
-              ├─→ Opus remote_audio_tx VITA PCC 0x8005
-              └─→ uncompressed VITA fallback PCC 0x03E3
-                              │
-                              ▼ [queued to NETWORK]
-                      PanadapterStream.sendToRadio()
-                              │
-                              ▼
-                         Radio UDP 4991
+              ├─→ backend TX seam (HL2/Icom, 24 kHz Int16)
+              │       └─→ backend-native conversion/modulation
+              │
+              └─→ Flex Opus PCC 0x8005 / uncompressed PCC 0x03E3
+                              └─→ [queued to NETWORK]
+                                  PanadapterStream.sendToRadio()
+                                      └─→ Radio UDP 4991
 
 The PSK Reporter WSPR beacon is a one-shot, operator-armed digital source. A
 precise timer on the AudioEngine worker thread generates sample-accurate 4-FSK

docs/architecture/tx-audio-signal-path.md

@@ -18,6 +18,14 @@ PC mic voice TX stream before VITA/Opus packetization. The radio receives
 the already-shaped voice signal and treats it identically to any other
 PC-mic input (enters at SC_MIC, meter 26).
 
+Mic hardware remains Int16 at its negotiated device rate. The normal voice
+path canonicalizes the selected channel, converts to float once, normalizes to
+a fixed 48 kHz processing domain, and remains float through RN2 and the complete
+channel strip. A 12%-transition-band SRC then converts 48 kHz to the existing
+24 kHz Flex transport rate before linked TPDF dither and the sole Int16
+quantization on the Flex path. See `audio-pipeline.md` for measured SRC delay
+and the separate HL2/Icom backend seam behavior.
+
 DAX/TCI TX and RADE are intentionally not part of this voice strip.
 DAX/TCI bypasses client voice DSP in `AudioEngine::feedDaxTxAudio()`.
 RADE branches early from `AudioEngine::onTxAudioReady()` and bypasses
@@ -29,10 +37,16 @@ to bypass, double-click to open the floating editor.
 
 ```
 PC mic capture (QAudioSource)
+  │
+  ▼  Int16 at negotiated device rate
+Canonical channel selection → float conversion → 48 kHz normalization
+  │
+  ▼
+Optional native-48 kHz RN2 → 48 kHz test tone
   │
   ▼
 ┌───────────────────────────────────────────────────────────────────┐
-│  CHAIN widget — drag-drop ordered TX DSP pipeline                  │
+│  CHAIN widget — drag-drop ordered 48 kHz float TX DSP pipeline     │
 │                                                                     │
 │  [GATE] → [EQ] → [DESS] → [COMP] → [TUBE] → [PUDU] → [VERB]        │
 │                                                                     │
@@ -45,7 +59,7 @@ PC mic capture (QAudioSource)
 │  ClientReverb  — Freeverb (disabled by default)                     │
 │                                                                     │
 │  Audio thread loads the packed chain order once per block and       │
-│  dispatches each stage to its per-stage apply helper.               │
+│  TxVoiceProcessor dispatches each enabled processor directly.       │
 └─────────┬──────────────────────────────────────────────────────────┘
           │
           ▼  (meters: per-stage inputPeak/outputPeak/GR, ClientEq FFT
@@ -55,6 +69,9 @@ PC mic capture (QAudioSource)
      PC mic gain → Quindar → final limiter → meters/scopes
           │
           ▼
+     48-to-24 kHz SRC → linked TPDF dither → Int16 quantization
+          │
+          ▼
      Opus remote_audio_tx / VITA encode → UDP → radio
 ```
 

docs/audio-sink-factory.md

@@ -9,8 +9,9 @@
 Every audio sink and source in AetherSDR currently answers two questions on its
 own:
 
-1. **"What rate / sample format does this device want, and how do I get my
-   24 kHz canonical audio onto it?"** — reimplemented ~9 times with ~6 divergent
+1. **"What rate / sample format does this device want, and how do I bridge it
+   to the selected route's canonical device-boundary rate?"** — reimplemented
+   ~9 times with ~6 divergent
    fallback ladders and per-OS `#ifdef` branches that have drifted apart. This
    is the root of a recurring class of platform bugs (44.1k-only devices
    silently failing on some sinks, WASAPI Float32-only devices rejecting Int16,
@@ -84,7 +85,7 @@ The ladder, in one place, owns:
 | Int16 ↔ Float32 | both tried per rate; Float-first output, Int16-first input | [#2669] / [#1090] |
 | `preferredFormat()` catch-all | final rung for Float32-only virtual drivers / WASAPI | [#3231] |
 | macOS mic preferred-rate-first | avoids silent 48 k open on 16 k-native mics | PR [#2930] |
-| macOS Bluetooth-HFP mic | native 8/16/24 k first, never forced to 48 k | [#2615] |
+| macOS Bluetooth-HFP mic | open native 8/16/24 k first; normal voice then normalizes to 48 k, while RADE converts separately to 24 k | [#2615] |
 | Windows probe-at-open | skip unreliable `isFormatSupported`, try at open | [#2120] / [#2929] |
 
 ### Layer 2 — `AudioDeviceNegotiator` (live wrapper, Qt Multimedia) — next