Add PCAN support to MFCC by wrapping Tensorflow Lite Microfrontend.

CMakeLists.txt

@@ -2,6 +2,9 @@
 
 cmake_minimum_required(VERSION 3.13)
 
+if(CMAKE_C_COMPILER_ID MATCHES "Clang")
+	add_compile_options(-mllvm --text-section-literals=false)
+endif()
 if("${CMAKE_CURRENT_SOURCE_DIR}" STREQUAL "${CMAKE_CURRENT_BINARY_DIR}")
 	message(FATAL_ERROR
 		" In-source builds are not supported.\n"
@@ -58,7 +61,10 @@ set(CMAKE_ASM_FLAGS -DASSEMBLY)
 # that may be exported / installed
 add_library(sof_public_headers INTERFACE)
 
-target_include_directories(sof_public_headers INTERFACE ${PROJECT_SOURCE_DIR}/src/include)
+target_include_directories(sof_public_headers INTERFACE
+	${PROJECT_SOURCE_DIR}/src/include
+	${PROJECT_SOURCE_DIR}/../tflite-micro
+)
 
 # interface library that is used only as container for sof binary options
 # other targets can use it to build with the same options

src/audio/mfcc/Kconfig

@@ -11,6 +11,7 @@ config COMP_MFCC
 	select MATH_DECIBELS
 	select MATH_FFT
 	select MATH_MATRIX
+	select MATH_PCAN
 	select MATH_WINDOW
 	select NATURAL_LOGARITHM_FIXED
 	select NUMBERS_NORM

src/audio/mfcc/README.md

@@ -1,25 +1,84 @@
-# 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.
+
+---
+
+## 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 --> 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
 ```
 
-## Configuration and Scripts
+### 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
+```
+
+### 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/mfcc_common.c

@@ -13,6 +13,7 @@
 #include <sof/math/auditory.h>
 #include <sof/math/fft.h>
 #include <sof/math/matrix.h>
+#include <sof/math/pcan.h>
 #include <sof/math/sqrt.h>
 #include <sof/math/trig.h>
 #include <sof/math/window.h>
@@ -355,6 +356,11 @@ int mfcc_stft_process(struct processing_module *mod, struct mfcc_comp_data *cd)
 		psy_apply_mel_filterbank_32(&state->melfb, fft->fft_out, state->power_spectra,
 					    state->mel_log_32, mel_scale_shift);
 
+		if (state->pcan.enable_pcan) {
+			pcan_update_noise_estimate(&state->pcan, (const uint32_t *)state->mel_log_32);
+			pcan_apply(&state->pcan, (uint32_t *)state->mel_log_32);
+		}
+
 		if (state->mel_only) {
 			/* In Mel-only mode output Mel log spectra directly */
 			cc_count += state->dct.num_in;

src/audio/mfcc/mfcc_setup.c

@@ -10,6 +10,7 @@
 #include <sof/math/auditory.h>
 #include <sof/math/icomplex16.h>
 #include <sof/math/icomplex32.h>
+#include <sof/math/pcan.h>
 #include <sof/math/trig.h>
 #include <sof/math/window.h>
 #include <sof/trace/trace.h>
@@ -459,9 +460,63 @@ int mfcc_setup(struct processing_module *mod, int max_frames, int sample_rate, i
 		}
 	}
 
+	if (config->enable_pcan) {
+		struct pcan_config pcfg;
+		uint32_t *pcan_noise;
+		int16_t *pcan_lut;
+
+		pcfg.strength = (config->pcan_strength > 0) ?
+			((float)config->pcan_strength / 32768.0f) : 0.95f;
+		pcfg.offset = (config->pcan_offset > 0) ?
+			((float)config->pcan_offset / 128.0f) : 80.0f;
+		pcfg.gain_bits = (config->pcan_gain_bits > 0) ?
+			config->pcan_gain_bits : 21;
+		pcfg.smoothing_coef = (config->pcan_smoothing_coef > 0) ?
+			(uint16_t)config->pcan_smoothing_coef : 819;
+		pcfg.smoothing_bits = 10;
+		pcfg.input_correction_bits = 0;
+		pcfg.enable_pcan = true;
+
+		pcan_noise = mod_zalloc(mod, config->num_mel_bins * sizeof(uint32_t));
+		if (!pcan_noise) {
+			comp_err(dev, "Failed PCAN noise estimate alloc");
+			ret = -ENOMEM;
+			goto free_vad;
+		}
+
+		pcan_lut = mod_zalloc(mod, PCAN_LUT_SIZE * sizeof(int16_t));
+		if (!pcan_lut) {
+			comp_err(dev, "Failed PCAN gain LUT alloc");
+			mod_free(mod, pcan_noise);
+			ret = -ENOMEM;
+			goto free_vad;
+		}
+
+		ret = pcan_populate_state(&pcfg, &state->pcan, pcan_noise, pcan_lut,
+					  config->num_mel_bins, 10, 0);
+		if (ret < 0) {
+			comp_err(dev, "Failed PCAN state init");
+			mod_free(mod, pcan_noise);
+			mod_free(mod, pcan_lut);
+			goto free_vad;
+		}
+	} else {
+		state->pcan.enable_pcan = false;
+		state->pcan.noise_estimate = NULL;
+		state->pcan.gain_lut = NULL;
+	}
+
 	comp_dbg(dev, "done");
 	return 0;
 
+free_vad:
+	if (config->enable_vad) {
+		mod_free(mod, cd->vad.noise_floor);
+		mod_free(mod, cd->vad.weights);
+		cd->vad.noise_floor = NULL;
+		cd->vad.weights = NULL;
+	}
+
 free_out_stage:
 	mod_free(mod, state->out_stage);
 
@@ -528,6 +583,8 @@ void mfcc_free_buffers(struct processing_module *mod)
 	mfcc_free_and_null(mod, (void **)&cd->state.dct.matrix);
 	mfcc_free_and_null(mod, (void **)&cd->state.lifter.matrix);
 	mfcc_free_and_null(mod, (void **)&cd->state.out_stage);
+	mfcc_free_and_null(mod, (void **)&cd->state.pcan.noise_estimate);
+	mfcc_free_and_null(mod, (void **)&cd->state.pcan.gain_lut);
 	mfcc_free_and_null(mod, (void **)&cd->vad.noise_floor);
 	mfcc_free_and_null(mod, (void **)&cd->vad.weights);
 }

src/audio/mfcc/tune/setup_mfcc.m

@@ -51,6 +51,13 @@ function setup_mfcc()
 	setup.tplg_fn = 'ceps13_compress_dtx.conf';
 	export_mfcc_setup(gen_cfg, setup);
 
+	% Mel spectrogram with PCAN normalization and compress PCM output
+	setup = get_mel_spectrogram_config();
+	setup.enable_pcan = true;
+	setup.compress_output = true;
+	setup.tplg_fn = 'mel80_pcan_compress.conf';
+	export_mfcc_setup(gen_cfg, setup);
+
 end
 
 function cfg = get_mfcc_default_config()
@@ -94,6 +101,11 @@ function setup_mfcc()
 	cfg.dtx_silence_hops_interval = 0;
 	cfg.update_controls = false;
 	cfg.compress_output = false;
+	cfg.enable_pcan = false;
+	cfg.pcan_strength = 0.95;
+	cfg.pcan_offset = 80.0;
+	cfg.pcan_gain_bits = 21;
+	cfg.pcan_smoothing_coef = 819;
 end
 
 function cfg = get_mel_spectrogram_config()
@@ -113,7 +125,7 @@ function setup_mfcc()
 	cfg.num_mel_bins = 80;
 	cfg.preemphasis_coefficient = 0;
 	cfg.raw_energy = false;
-	cfg.remove_dc_offset = false;
+	cfg.remove_dc_offset = true;
 	cfg.round_to_power_of_two = true;
 	cfg.sample_frequency = 16000;
 	cfg.snip_edges = true;
@@ -123,20 +135,25 @@ function setup_mfcc()
 	cfg.vtln_low = 0;
 	cfg.vtln_warp = 1.0;
 	cfg.window_type = 'hann';
-	cfg.mel_log = 'log10';
+	cfg.mel_log = 'log'; % Set to 'db' for librosa, set to 'log10' for matlab
 	cfg.pmin = 1e-10;
-	cfg.top_db = 8; % applied for log10, would be 80 dB clamp for decibels as 10*log10()
-	cfg.mel_offset = 4.0; % For whisper like Mel scale and normalize
-	cfg.mel_scale = 0.25; % For whisper like Mel scale and normalize
-	cfg.mmax_init = 0; % Initial value max Mel value, data clamp is mmax - top_db
-	cfg.mmax_coef = 0; % Dynamic max Mel value decay coefficient (zero lock to found max)
-	cfg.dynamic_mmax = true;
-	cfg.enable_vad = true;
+	cfg.top_db = 80.0;
+	cfg.mel_offset = 4.0; % Whisper: (mel + 4.0) * 0.25
+	cfg.mel_scale = 0.25;
+	cfg.mmax_init = 0;
+	cfg.mmax_coef = 0.005; % Whisper mmax tracking: slow decay
+	cfg.dynamic_mmax = false;
+	cfg.enable_vad = false;
 	cfg.enable_dtx = false;
 	cfg.dtx_trailing_silence_hops = 0;
 	cfg.dtx_silence_hops_interval = 0;
-	cfg.update_controls = true;
+	cfg.update_controls = false;
 	cfg.compress_output = false;
+	cfg.enable_pcan = false;
+	cfg.pcan_strength = 0.95;
+	cfg.pcan_offset = 80.0;
+	cfg.pcan_gain_bits = 21;
+	cfg.pcan_smoothing_coef = 819;
 end
 
 function export_mfcc_setup(gen_cfg, cfg)
@@ -173,8 +190,9 @@ function export_mfcc_setup(gen_cfg, cfg)
 
 v = cfg.dtx_trailing_silence_hops;                [b8, j] = add_w16b(v, b8, j); % DTX trailing silence hops
 v = cfg.dtx_silence_hops_interval;                [b8, j] = add_w16b(v, b8, j); % DTX silence frame interval
+v = cfg.pcan_smoothing_coef;                 [b8, j] = add_w32b(v, b8, j); % PCAN smoothing coef in Q14
 % Reserved
-for i = 1:5
+for i = 1:4
 	[b8, j] = add_w32b(0, b8, j);
 end
 
@@ -200,10 +218,9 @@ function export_mfcc_setup(gen_cfg, cfg)
 v = 0;                                           [b8, j] = add_w16b(v, b8, j); % vtln_high Qx.y TBD
 v = 0;                                           [b8, j] = add_w16b(v, b8, j); % vtln_low Qx.y TBD
 v = 0;                                           [b8, j] = add_w16b(v, b8, j); % vtln_warp Qx.y TBD
-% reserved16[3]
-for i = 1:3
-	[b8, j] = add_w16b(0, b8, j);
-end
+v = q_convert(cfg.pcan_strength, 15);            [b8, j] = add_w16b(v, b8, j); % PCAN strength in Q1.15
+v = q_convert(cfg.pcan_offset, 7);               [b8, j] = add_w16b(v, b8, j); % PCAN offset in Q8.7
+v = cfg.pcan_gain_bits;                          [b8, j] = add_w16b(v, b8, j); % PCAN gain bits
 v = cfg.htk_compat;                              [b8, j] = add_w8b(v, b8, j); % bool
 v = cfg.raw_energy;                              [b8, j] = add_w8b(v, b8, j); % bool
 v = cfg.remove_dc_offset;                        [b8, j] = add_w8b(v, b8, j); % bool
@@ -216,8 +233,9 @@ function export_mfcc_setup(gen_cfg, cfg)
 v = cfg.enable_dtx;                              [b8, j] = add_w8b(v, b8, j); % bool
 v = cfg.update_controls;                         [b8, j] = add_w8b(v, b8, j); % bool
 v = cfg.compress_output;                         [b8, j] = add_w8b(v, b8, j); % bool
-% reserved_bool[4]
-for i = 1:4
+v = cfg.enable_pcan;                             [b8, j] = add_w8b(v, b8, j); % bool
+% reserved_bool[3]
+for i = 1:3
 	[b8, j] = add_w8b(0, b8, j);
 end
 

src/include/sof/audio/mfcc/mfcc_comp.h

@@ -13,6 +13,7 @@
 #include <sof/math/auditory.h>
 #include <sof/math/dct.h>
 #include <sof/math/fft.h>
+#include <sof/math/pcan.h>
 #include <sof/audio/mfcc/mfcc_vad.h>
 #include <sof/ipc/msg.h>
 #include <stddef.h>
@@ -99,6 +100,7 @@ struct mfcc_state {
 	struct mfcc_fft fft; /**< FFT related */
 	struct dct_plan_16 dct; /**< DCT related */
 	struct psy_mel_filterbank melfb; /**< Mel filter bank */
+	struct pcan_state pcan; /**< PCAN state */
 	struct mfcc_cepstral_lifter lifter; /**< Cepstral lifter coefficients */
 	struct mat_matrix_16b *mel_spectra; /**< Pointer to scratch */
 	struct mat_matrix_16b *cepstral_coef; /**< Pointer to scratch */