audio: wov-upstream backport (PR 2 of 3)

.github/workflows/testbench.yml

@@ -47,6 +47,19 @@ jobs:
         with:
           path: sof
 
+      # MFCC with PCAN needs the TFLM microfrontend sources next to sof/,
+      # use the revision pinned in the west submanifest.
+      - name: Checkout TFLM microfrontend
+        run: |
+          rev=$(awk '/name: tflite-micro/ {f=1} f && /revision:/ {print $2; exit}' \
+                sof/submanifests/tflm.yml)
+          git init tflite-micro
+          cd tflite-micro
+          git remote add origin https://github.com/tensorflow/tflite-micro
+          git sparse-checkout set tensorflow/lite/experimental/microfrontend/lib
+          git fetch --depth 1 --filter=blob:none origin "$rev"
+          git checkout FETCH_HEAD
+
       # note: libasound-dev needed for testbench build which still
       #       requires system ALSA headers to be present
       - name: apt get

src/arch/host/configs/library_defconfig

@@ -11,6 +11,7 @@ CONFIG_COMP_IIR=y
 CONFIG_COMP_IGO_NR=y
 CONFIG_COMP_LEVEL_MULTIPLIER=y
 CONFIG_COMP_MFCC=y
+CONFIG_COMP_MFCC_PCAN=y
 CONFIG_COMP_MODULE_ADAPTER=y
 CONFIG_COMP_MULTIBAND_DRC=y
 CONFIG_COMP_MUX=y

src/audio/mfcc/CMakeLists.txt

@@ -1,10 +1,28 @@
 # SPDX-License-Identifier: BSD-3-Clause
 
+if(CONFIG_COMP_MFCC_PCAN)
+  if(NOT DEFINED TFLM_PATH)
+    if(DEFINED SOF_ROOT_SOURCE_DIRECTORY AND EXISTS "${SOF_ROOT_SOURCE_DIRECTORY}/../tflite-micro")
+      set(TFLM_PATH "${SOF_ROOT_SOURCE_DIRECTORY}/../tflite-micro")
+    elseif(DEFINED sof_top_dir AND EXISTS "${sof_top_dir}/../tflite-micro")
+      set(TFLM_PATH "${sof_top_dir}/../tflite-micro")
+    elseif(EXISTS "${CMAKE_CURRENT_LIST_DIR}/../../../../tflite-micro")
+      set(TFLM_PATH "${CMAKE_CURRENT_LIST_DIR}/../../../../tflite-micro")
+    endif()
+  endif()
+  is_zephyr(zephyr)
+  if(zephyr)
+    zephyr_include_directories(${TFLM_PATH})
+  else()
+    target_include_directories(sof PRIVATE ${TFLM_PATH})
+  endif()
+endif()
+
 if(CONFIG_COMP_MFCC STREQUAL "m" AND DEFINED CONFIG_LLEXT)
   add_subdirectory(llext ${PROJECT_BINARY_DIR}/mfcc_llext)
   add_dependencies(app mfcc)
 else()
-  add_local_sources(sof mfcc.c mfcc_setup.c mfcc_common.c mfcc_generic.c mfcc_hifi4.c mfcc_hifi3.c mfcc_vad.c)
+  add_local_sources(sof mfcc.c mfcc_setup.c mfcc_common.c mfcc_generic.c mfcc_hifi4.c mfcc_hifi3.c mfcc_vad.c mfcc_agc8.c)
   if(CONFIG_IPC_MAJOR_4)
     add_local_sources(sof mfcc_ipc4.c)
   endif()

src/audio/mfcc/Kconfig

@@ -24,3 +24,26 @@ config COMP_MFCC
 	  The characteristic of the audio features are defined in the binary
 	  control blob. Directory tools/tune/mfcc contains a tool to create
 	  the configurations.
+
+if COMP_MFCC
+
+config COMP_MFCC_PCAN
+	bool "Enable PCAN (Per-Channel AGC Normalization) in MFCC"
+	default n
+	select MATH_PCAN
+	help
+	  Select to build MFCC with PCAN (Per-Channel Automatic Gain Control
+	  Normalization) support on the Mel filterbank energies. PCAN performs
+	  dynamic per-channel gain control and piecewise polynomial root
+	  compression to suppress stationary noise and enhance transients,
+	  producing normalized features suitable for direct int8 quantization
+	  downstream (e.g. microWakeWord). The topology PCAN configuration
+	  fields are ignored when this option is disabled.
+
+config COMP_MFCC_DEBUG_TRACE
+	bool "Verbose per-process debug traces in MFCC"
+	default n
+	help
+	  Enable MFCC per-process debug prints in the hot path.
+
+endif # COMP_MFCC

src/audio/mfcc/README.md

@@ -1,25 +1,99 @@
-# MFCC Feature Extraction Architecture
+# MFCC & PCAN Feature Extraction Architecture
 
-This directory contains the Mel-Frequency Cepstral Coefficients (MFCC) feature extractor.
+This directory contains the **Mel-Frequency Cepstral Coefficients (MFCC)** feature extractor and integrated **PCAN (Per-Channel AGC Normalization)** audio pre-processing component for Sound Open Firmware (SOF).
 
-## Overview
+---
 
-MFCC extracts audio features commonly used as inputs for machine learning models, such as wake-word detection or speech recognition.
+## 1. Overview
 
-## Architecture Diagram
+The MFCC module converts raw streaming time-domain PCM audio into compact acoustic feature representations (e.g. 13-bin cepstral coefficients, 40-bin or 80-bin Mel spectrograms) suitable for on-device machine learning models such as wake-word classifiers (TFLM, microWakeWord) and speech recognition engines (Whisper).
+
+When enabled, **PCAN** applies per-channel adaptive dynamic gain control and piecewise root dynamic range compression across time to suppress stationary background noise and normalize channel energy before feature quantization.
+
+### Output modes
+
+One build supports all output modes, selected at runtime by the configuration blob:
+
+| Mode | Blob settings | Output per hop |
+|---|---|---|
+| Cepstral coefficients | `num_ceps > 0` | int16 Q9.7 widened to int32 |
+| Mel log spectra | `num_ceps == 0` | int32 Q9.23 |
+| PCAN Mel | `num_ceps == 0`, `enable_pcan` (needs `CONFIG_COMP_MFCC_PCAN`) | int8 |
+| Soft AGC Mel (AGC8) | `num_ceps == 0`, `enable_agc8` | int8 Q1.7 |
+
+AGC8 is a light alternative to PCAN without the per-channel noise estimate and the `tflite-micro` dependency. A single gain is added to all Q9.23 Mel log values of a hop. The gain attacks instantly to keep the hop peak at or below +1.0, and releases towards a target with a slower step when the VAD reports speech. The values are clamped to -1.0 .. +1.0 and converted to int8 Q1.7. The blob fields `agc_target`, `agc_floor` (Q8.7) and `agc_release`, `agc_release_speech` (Q9.23 per hop) set the parameters, zero selects the default.
+
+---
+
+## 2. Architecture & Data Flow
 
 ```mermaid
 graph LR
-  In[Audio Frame] --> Win[Windowing]
-  Win --> FFT[FFT]
-  FFT --> Mel[Mel Filterbank]
-  Mel --> DCT[DCT]
-  DCT --> Out[MFCC Output Features]
+  In[Audio PCM Frame] --> Win[Windowing: Hamming/Hann]
+  Win --> FFT[FFT: 512-pt]
+  FFT --> Mel[Mel Filterbank: 40/80 Bins]
+  Mel --> PCAN[PCAN Gain Normalization]
+  PCAN --> Log[Log Scaling / dB]
+  Log --> VAD[VAD / DTX Silence Gating]
+  Log --> AGC8[Soft AGC int8]
+  VAD --> AGC8
+  Log --> DCT[DCT: Cepstral Matrix]
+  DCT --> Lifter[Cepstral Lifter]
+  Lifter --> Out[MFCC Output Features]
+```
+
+---
+
+## 3. Dependencies
+
+| Dependency | Purpose |
+|---|---|
+| **West Manifest (`west.yml`)** | Manages workspace dependencies including `tflite-micro` |
+| **`tflite-micro`** | Provides Google's upstream Apache-2.0 `microfrontend` PCAN library (`pcan_gain_control.c`, `noise_reduction.c`) |
+| **GNU Octave ($\ge 5.0$)** | Required for topology configuration generation (`tune/setup_mfcc.m`) and reference vector export |
+| **CMake ($\ge 3.13$)** | SOF build system |
+| **Zephyr SDK / Xtensa XCC** | DSP cross-compilation toolchains for Cadence HiFi3 / HiFi4 SIMD kernels |
+| **CMocka** | Host unit testing framework |
+
+---
+
+## 4. Build & Tuning Instructions
+
+### 4.1 Synchronize Dependencies via West
+```bash
+west update
+```
+
+### 4.2 Build Host Unit Tests
+```bash
+# Configure unit tests
+cmake -S sof -B build_ut -DBUILD_UNIT_TESTS=ON -DBUILD_UNIT_TESTS_HOST=ON -DINIT_CONFIG=unit_test_defconfig
+
+# Build all math and MFCC tests
+cmake --build build_ut --target pcan
+ctest --test-dir build_ut -R "pcan|auditory|dct|matrix|window|fft" --output-on-failure
 ```
 
-## Configuration and Scripts
+### 4.3 Build Target DSP Firmware
+```bash
+west build -b intel_adsp_ace15_mtpm app
+```
+
+### 4.4 Generate Topology Configurations (Octave)
+The tuning script generates binary configuration blobs and ALSA topology configurations:
+```bash
+cd src/audio/mfcc/tune
+octave-cli --eval "setup_mfcc"
+```
+Available presets generated:
+- `ceps13_compress_dtx.conf`: 13-bin cepstral features with VAD and DTX silence suppression.
+- `mel80_compress.conf`: 80-bin linear Mel spectrogram features.
+- `mel80_compress_dtx.conf`: 80-bin Mel spectrogram with DTX gating.
+- `mel80_pcan_compress.conf`: 80-bin Mel spectrogram with Google PCAN dynamic AGC normalization.
+
+---
+
+## 5. Configuration Options
 
-- **Kconfig**: Enables the MFCC component (`COMP_MFCC`) which selects necessary math libraries (`MATH_FFT`, `MATH_DCT`, `MATH_16BIT_MEL_FILTERBANK`, etc.). Depends on `COMP_MODULE_ADAPTER`.
-- **CMakeLists.txt**: Compiles generic, common, and HIFI implementations (`mfcc_hifi3.c`, `mfcc_hifi4.c`). Provides support for Zephyr loadable extensions (`llext`).
-- **mfcc.toml**: Specifies the topology configuration for the MFCC module (UUID, affinity, memory parameters, and pin formats).
-- **Topology (.conf)**: Derived from `tools/topology/topology2/include/components/mfcc.conf`, configuring a `mfcc` widget object of type `effect` with UUID `73:a7:10:db:a4:1a:ea:4c:a2:1f:2d:57:a5:c9:82:eb`.
+- **`CONFIG_COMP_MFCC`**: Enables the MFCC component and automatically selects required math libraries (`CONFIG_MATH_FFT`, `CONFIG_MATH_AUDITORY`, `CONFIG_MATH_DCT`, `CONFIG_MATH_PCAN`, `CONFIG_MATH_WINDOW`).
+- **`CONFIG_COMP_MODULE_ADAPTER`**: Module adapter infrastructure for SOF processing components.

src/audio/mfcc/llext/CMakeLists.txt

@@ -8,5 +8,8 @@ sof_llext_build("mfcc"
 		../mfcc_generic.c
 		../mfcc_hifi3.c
 		../mfcc_hifi4.c
+		../mfcc_vad.c
+		../mfcc_agc8.c
+		../mfcc_ipc4.c
 	LIB openmodules
 )