MLECO-3173: Add AD, KWS_ASR and Noise reduction use case API's

Signed-off-by: Richard Burton <richard.burton@arm.com>

Change-Id: I36f61ce74bf17f7b327cdae9704a22ca54144f37

4 files changed, 175 insertions, 118 deletions

diff --git a/source/use_case/noise_reduction/src/MainLoop.cc b/source/use_case/noise_reduction/src/MainLoop.cc
index 5fd7823..fd72127 100644
--- a/source/use_case/noise_reduction/src/MainLoop.cc
+++ b/source/use_case/noise_reduction/src/MainLoop.cc
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * Copyright (c) 2021-2022 Arm Limited. All rights reserved.
  * SPDX-License-Identifier: Apache-2.0
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -14,12 +14,10 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#include "hal.h"                    /* Brings in platform definitions. */
 #include "UseCaseHandler.hpp"       /* Handlers for different user options. */
 #include "UseCaseCommonUtils.hpp"   /* Utils functions. */
 #include "RNNoiseModel.hpp"         /* Model class for running inference. */
 #include "InputFiles.hpp"           /* For input audio clips. */
-#include "RNNoiseProcess.hpp"       /* Pre-processing class */
 #include "log_macros.h"
 
 enum opcodes
diff --git a/source/use_case/noise_reduction/src/RNNoiseProcess.cc b/source/use_case/noise_reduction/src/RNNoiseFeatureProcessor.cc
index 4c568fa..036894c 100644
--- a/source/use_case/noise_reduction/src/RNNoiseProcess.cc
+++ b/source/use_case/noise_reduction/src/RNNoiseFeatureProcessor.cc
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * Copyright (c) 2021-2022 Arm Limited. All rights reserved.
  * SPDX-License-Identifier: Apache-2.0
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -14,7 +14,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#include "RNNoiseProcess.hpp"
+#include "RNNoiseFeatureProcessor.hpp"
 #include "log_macros.h"
 
 #include <algorithm>
@@ -33,7 +33,7 @@ do {                                                \
     }                                               \
 } while(0)
 
-RNNoiseProcess::RNNoiseProcess() :
+RNNoiseFeatureProcessor::RNNoiseFeatureProcessor() :
         m_halfWindow(FRAME_SIZE, 0),
         m_dctTable(NB_BANDS * NB_BANDS),
         m_analysisMem(FRAME_SIZE, 0),
@@ -54,9 +54,9 @@ RNNoiseProcess::RNNoiseProcess() :
     this->InitTables();
 }
 
-void RNNoiseProcess::PreprocessFrame(const float*   audioData,
-                                     const size_t   audioLen,
-                                     FrameFeatures& features)
+void RNNoiseFeatureProcessor::PreprocessFrame(const float*   audioData,
+                                              const size_t   audioLen,
+                                              FrameFeatures& features)
 {
     /* Note audioWindow is modified in place */
     const arrHp aHp {-1.99599, 0.99600 };
@@ -68,7 +68,7 @@ void RNNoiseProcess::PreprocessFrame(const float*   audioData,
     this->ComputeFrameFeatures(audioWindow, features);
 }
 
-void RNNoiseProcess::PostProcessFrame(vec1D32F& modelOutput, FrameFeatures& features, vec1D32F& outFrame)
+void RNNoiseFeatureProcessor::PostProcessFrame(vec1D32F& modelOutput, FrameFeatures& features, vec1D32F& outFrame)
 {
     std::vector<float> outputBands = modelOutput;
     std::vector<float> gain(FREQ_SIZE, 0);
@@ -92,7 +92,7 @@ void RNNoiseProcess::PostProcessFrame(vec1D32F& modelOutput, FrameFeatures& feat
     FrameSynthesis(outFrame, features.m_fftX);
 }
 
-void RNNoiseProcess::InitTables()
+void RNNoiseFeatureProcessor::InitTables()
 {
     constexpr float pi = M_PI;
     constexpr float halfPi = M_PI / 2;
@@ -111,7 +111,7 @@ void RNNoiseProcess::InitTables()
     }
 }
 
-void RNNoiseProcess::BiQuad(
+void RNNoiseFeatureProcessor::BiQuad(
         const arrHp& bHp,
         const arrHp& aHp,
         arrHp& memHpX,
@@ -126,8 +126,8 @@ void RNNoiseProcess::BiQuad(
     }
 }
 
-void RNNoiseProcess::ComputeFrameFeatures(vec1D32F& audioWindow,
-                                          FrameFeatures& features)
+void RNNoiseFeatureProcessor::ComputeFrameFeatures(vec1D32F& audioWindow,
+                                                   FrameFeatures& features)
 {
     this->FrameAnalysis(audioWindow,
                         features.m_fftX,
@@ -264,7 +264,7 @@ void RNNoiseProcess::ComputeFrameFeatures(vec1D32F& audioWindow,
     features.m_featuresVec[NB_BANDS + 3 * NB_DELTA_CEPS + 1] = specVariability / CEPS_MEM - 2.1;
 }
 
-void RNNoiseProcess::FrameAnalysis(
+void RNNoiseFeatureProcessor::FrameAnalysis(
     const vec1D32F& audioWindow,
     vec1D32F& fft,
     vec1D32F& energy,
@@ -289,7 +289,7 @@ void RNNoiseProcess::FrameAnalysis(
     ComputeBandEnergy(fft, energy);
 }
 
-void RNNoiseProcess::ApplyWindow(vec1D32F& x)
+void RNNoiseFeatureProcessor::ApplyWindow(vec1D32F& x)
 {
     if (WINDOW_SIZE != x.size()) {
         printf_err("Invalid size for vector to be windowed\n");
@@ -305,7 +305,7 @@ void RNNoiseProcess::ApplyWindow(vec1D32F& x)
     }
 }
 
-void RNNoiseProcess::ForwardTransform(
+void RNNoiseFeatureProcessor::ForwardTransform(
     vec1D32F& x,
     vec1D32F& fft)
 {
@@ -327,7 +327,7 @@ void RNNoiseProcess::ForwardTransform(
      * first half of the FFT's. The conjugates are not present. */
 }
 
-void RNNoiseProcess::ComputeBandEnergy(const vec1D32F& fftX, vec1D32F& bandE)
+void RNNoiseFeatureProcessor::ComputeBandEnergy(const vec1D32F& fftX, vec1D32F& bandE)
 {
     bandE = vec1D32F(NB_BANDS, 0);
 
@@ -351,7 +351,7 @@ void RNNoiseProcess::ComputeBandEnergy(const vec1D32F& fftX, vec1D32F& bandE)
     bandE[NB_BANDS - 1] *= 2;
 }
 
-void RNNoiseProcess::ComputeBandCorr(const vec1D32F& X, const vec1D32F& P, vec1D32F& bandC)
+void RNNoiseFeatureProcessor::ComputeBandCorr(const vec1D32F& X, const vec1D32F& P, vec1D32F& bandC)
 {
     bandC = vec1D32F(NB_BANDS, 0);
     VERIFY(this->m_eband5ms.size() >= NB_BANDS);
@@ -374,7 +374,7 @@ void RNNoiseProcess::ComputeBandCorr(const vec1D32F& X, const vec1D32F& P, vec1D
     bandC[NB_BANDS - 1] *= 2;
 }
 
-void RNNoiseProcess::DCT(vec1D32F& input, vec1D32F& output)
+void RNNoiseFeatureProcessor::DCT(vec1D32F& input, vec1D32F& output)
 {
     VERIFY(this->m_dctTable.size() >= NB_BANDS * NB_BANDS);
     for (uint32_t i = 0; i < NB_BANDS; ++i) {
@@ -387,7 +387,7 @@ void RNNoiseProcess::DCT(vec1D32F& input, vec1D32F& output)
     }
 }
 
-void RNNoiseProcess::PitchDownsample(vec1D32F& pitchBuf, size_t pitchBufSz) {
+void RNNoiseFeatureProcessor::PitchDownsample(vec1D32F& pitchBuf, size_t pitchBufSz) {
     for (size_t i = 1; i < (pitchBufSz >> 1); ++i) {
         pitchBuf[i] = 0.5 * (
                         0.5 * (this->m_pitchBuf[2 * i - 1] + this->m_pitchBuf[2 * i + 1])
@@ -431,7 +431,7 @@ void RNNoiseProcess::PitchDownsample(vec1D32F& pitchBuf, size_t pitchBufSz) {
     this->Fir5(lpc2, pitchBufSz >> 1, pitchBuf);
 }
 
-int RNNoiseProcess::PitchSearch(vec1D32F& xLp, vec1D32F& y, uint32_t len, uint32_t maxPitch) {
+int RNNoiseFeatureProcessor::PitchSearch(vec1D32F& xLp, vec1D32F& y, uint32_t len, uint32_t maxPitch) {
     uint32_t lag = len + maxPitch;
     vec1D32F xLp4(len >> 2, 0);
     vec1D32F yLp4(lag >> 2, 0);
@@ -488,7 +488,7 @@ int RNNoiseProcess::PitchSearch(vec1D32F& xLp, vec1D32F& y, uint32_t len, uint32
     return 2*bestPitch[0] - offset;
 }
 
-arrHp RNNoiseProcess::FindBestPitch(vec1D32F& xCorr, vec1D32F& y, uint32_t len, uint32_t maxPitch)
+arrHp RNNoiseFeatureProcessor::FindBestPitch(vec1D32F& xCorr, vec1D32F& y, uint32_t len, uint32_t maxPitch)
 {
     float Syy = 1;
     arrHp bestNum {-1, -1};
@@ -527,7 +527,7 @@ arrHp RNNoiseProcess::FindBestPitch(vec1D32F& xCorr, vec1D32F& y, uint32_t len,
     return bestPitch;
 }
 
-int RNNoiseProcess::RemoveDoubling(
+int RNNoiseFeatureProcessor::RemoveDoubling(
     vec1D32F& pitchBuf,
     uint32_t maxPeriod,
     uint32_t minPeriod,
@@ -679,12 +679,12 @@ int RNNoiseProcess::RemoveDoubling(
     return this->m_lastPeriod;
 }
 
-float RNNoiseProcess::ComputePitchGain(float xy, float xx, float yy)
+float RNNoiseFeatureProcessor::ComputePitchGain(float xy, float xx, float yy)
 {
     return xy / math::MathUtils::SqrtF32(1+xx*yy);
 }
 
-void RNNoiseProcess::AutoCorr(
+void RNNoiseFeatureProcessor::AutoCorr(
     const vec1D32F& x,
     vec1D32F& ac,
     size_t lag,
@@ -711,7 +711,7 @@ void RNNoiseProcess::AutoCorr(
 }
 
 
-void RNNoiseProcess::PitchXCorr(
+void RNNoiseFeatureProcessor::PitchXCorr(
     const vec1D32F& x,
     const vec1D32F& y,
     vec1D32F& xCorr,
@@ -728,7 +728,7 @@ void RNNoiseProcess::PitchXCorr(
 }
 
 /* Linear predictor coefficients */
-void RNNoiseProcess::LPC(
+void RNNoiseFeatureProcessor::LPC(
     const vec1D32F& correlation,
     int32_t p,
     vec1D32F& lpc)
@@ -766,7 +766,7 @@ void RNNoiseProcess::LPC(
     }
 }
 
-void RNNoiseProcess::Fir5(
+void RNNoiseFeatureProcessor::Fir5(
     const vec1D32F &num,
     uint32_t N,
     vec1D32F &x)
@@ -794,7 +794,7 @@ void RNNoiseProcess::Fir5(
     }
 }
 
-void RNNoiseProcess::PitchFilter(FrameFeatures &features, vec1D32F &gain) {
+void RNNoiseFeatureProcessor::PitchFilter(FrameFeatures &features, vec1D32F &gain) {
     std::vector<float> r(NB_BANDS, 0);
     std::vector<float> rf(FREQ_SIZE, 0);
     std::vector<float> newE(NB_BANDS);
@@ -835,7 +835,7 @@ void RNNoiseProcess::PitchFilter(FrameFeatures &features, vec1D32F &gain) {
     }
 }
 
-void RNNoiseProcess::FrameSynthesis(vec1D32F& outFrame, vec1D32F& fftY) {
+void RNNoiseFeatureProcessor::FrameSynthesis(vec1D32F& outFrame, vec1D32F& fftY) {
     std::vector<float> x(WINDOW_SIZE, 0);
     InverseTransform(x, fftY);
     ApplyWindow(x);
@@ -845,7 +845,7 @@ void RNNoiseProcess::FrameSynthesis(vec1D32F& outFrame, vec1D32F& fftY) {
     memcpy((m_synthesisMem.data()), &x[FRAME_SIZE], FRAME_SIZE*sizeof(float));
 }
 
-void RNNoiseProcess::InterpBandGain(vec1D32F& g, vec1D32F& bandE) {
+void RNNoiseFeatureProcessor::InterpBandGain(vec1D32F& g, vec1D32F& bandE) {
     for (size_t i = 0; i < NB_BANDS - 1; i++) {
         int bandSize = (m_eband5ms[i + 1] - m_eband5ms[i]) << FRAME_SIZE_SHIFT;
         for (int j = 0; j < bandSize; j++) {
@@ -855,7 +855,7 @@ void RNNoiseProcess::InterpBandGain(vec1D32F& g, vec1D32F& bandE) {
     }
 }
 
-void RNNoiseProcess::InverseTransform(vec1D32F& out, vec1D32F& fftXIn) {
+void RNNoiseFeatureProcessor::InverseTransform(vec1D32F& out, vec1D32F& fftXIn) {
 
     std::vector<float> x(WINDOW_SIZE * 2);  /* This is complex. */
     vec1D32F newFFT;  /* This is complex. */
diff --git a/source/use_case/noise_reduction/src/RNNoiseProcessing.cc b/source/use_case/noise_reduction/src/RNNoiseProcessing.cc
new file mode 100644
index 0000000..f6a3ec4
--- /dev/null
+++ b/source/use_case/noise_reduction/src/RNNoiseProcessing.cc
@@ -0,0 +1,100 @@
+/*
+ * Copyright (c) 2022 Arm Limited. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "RNNoiseProcessing.hpp"
+#include "log_macros.h"
+
+namespace arm {
+namespace app {
+
+    RNNoisePreProcess::RNNoisePreProcess(TfLiteTensor* inputTensor,
+            std::shared_ptr<rnn::RNNoiseFeatureProcessor> featureProcessor, std::shared_ptr<rnn::FrameFeatures> frameFeatures)
+    :   m_inputTensor{inputTensor},
+        m_featureProcessor{featureProcessor},
+        m_frameFeatures{frameFeatures}
+    {}
+
+    bool RNNoisePreProcess::DoPreProcess(const void* data, size_t inputSize)
+    {
+        if (data == nullptr) {
+            printf_err("Data pointer is null");
+            return false;
+        }
+
+        auto input = static_cast<const int16_t*>(data);
+        this->m_audioFrame = rnn::vec1D32F(input, input + inputSize);
+        m_featureProcessor->PreprocessFrame(this->m_audioFrame.data(), inputSize, *this->m_frameFeatures);
+
+        QuantizeAndPopulateInput(this->m_frameFeatures->m_featuresVec,
+                this->m_inputTensor->params.scale, this->m_inputTensor->params.zero_point,
+                this->m_inputTensor);
+
+        debug("Input tensor populated \n");
+
+        return true;
+    }
+
+    void RNNoisePreProcess::QuantizeAndPopulateInput(rnn::vec1D32F& inputFeatures,
+            const float quantScale, const int quantOffset,
+            TfLiteTensor* inputTensor)
+    {
+        const float minVal = std::numeric_limits<int8_t>::min();
+        const float maxVal = std::numeric_limits<int8_t>::max();
+
+        auto* inputTensorData = tflite::GetTensorData<int8_t>(inputTensor);
+
+        for (size_t i=0; i < inputFeatures.size(); ++i) {
+            float quantValue = ((inputFeatures[i] / quantScale) + quantOffset);
+            inputTensorData[i] = static_cast<int8_t>(std::min<float>(std::max<float>(quantValue, minVal), maxVal));
+        }
+    }
+
+    RNNoisePostProcess::RNNoisePostProcess(TfLiteTensor* outputTensor,
+            std::vector<int16_t>& denoisedAudioFrame,
+            std::shared_ptr<rnn::RNNoiseFeatureProcessor> featureProcessor,
+            std::shared_ptr<rnn::FrameFeatures> frameFeatures)
+    :   m_outputTensor{outputTensor},
+        m_denoisedAudioFrame{denoisedAudioFrame},
+        m_featureProcessor{featureProcessor},
+        m_frameFeatures{frameFeatures}
+        {
+            this->m_denoisedAudioFrameFloat.reserve(denoisedAudioFrame.size());
+            this->m_modelOutputFloat.resize(outputTensor->bytes);
+        }
+
+    bool RNNoisePostProcess::DoPostProcess()
+    {
+        const auto* outputData = tflite::GetTensorData<int8_t>(this->m_outputTensor);
+        auto outputQuantParams = GetTensorQuantParams(this->m_outputTensor);
+
+        for (size_t i = 0; i < this->m_outputTensor->bytes; ++i) {
+            this->m_modelOutputFloat[i] = (static_cast<float>(outputData[i]) - outputQuantParams.offset)
+                                  * outputQuantParams.scale;
+        }
+
+        this->m_featureProcessor->PostProcessFrame(this->m_modelOutputFloat,
+                *this->m_frameFeatures, this->m_denoisedAudioFrameFloat);
+
+        for (size_t i = 0; i < this->m_denoisedAudioFrame.size(); ++i) {
+            this->m_denoisedAudioFrame[i] = static_cast<int16_t>(
+                    std::roundf(this->m_denoisedAudioFrameFloat[i]));
+        }
+
+        return true;
+    }
+
+} /* namespace app */
+} /* namespace arm */
\ No newline at end of file
diff --git a/source/use_case/noise_reduction/src/UseCaseHandler.cc b/source/use_case/noise_reduction/src/UseCaseHandler.cc
index acb8ba7..53bb43e 100644
--- a/source/use_case/noise_reduction/src/UseCaseHandler.cc
+++ b/source/use_case/noise_reduction/src/UseCaseHandler.cc
@@ -21,12 +21,10 @@
 #include "ImageUtils.hpp"
 #include "InputFiles.hpp"
 #include "RNNoiseModel.hpp"
-#include "RNNoiseProcess.hpp"
+#include "RNNoiseFeatureProcessor.hpp"
+#include "RNNoiseProcessing.hpp"
 #include "log_macros.h"
 
-#include <cmath>
-#include <algorithm>
-
 namespace arm {
 namespace app {
 
@@ -36,17 +34,6 @@ namespace app {
     **/
     static void IncrementAppCtxClipIdx(ApplicationContext& ctx);
 
-    /**
-    * @brief            Quantize the given features and populate the input Tensor.
-    * @param[in]        inputFeatures   Vector of floating point features to quantize.
-    * @param[in]        quantScale      Quantization scale for the inputTensor.
-    * @param[in]        quantOffset     Quantization offset for the inputTensor.
-    * @param[in,out]    inputTensor     TFLite micro tensor to populate.
-    **/
-    static void QuantizeAndPopulateInput(rnn::vec1D32F& inputFeatures,
-                                         float quantScale, int quantOffset,
-                                         TfLiteTensor* inputTensor);
-
     /* Noise reduction inference handler. */
     bool NoiseReductionHandler(ApplicationContext& ctx, bool runAll)
     {
@@ -57,7 +44,7 @@ namespace app {
         size_t memDumpMaxLen = 0;
         uint8_t* memDumpBaseAddr = nullptr;
         size_t undefMemDumpBytesWritten = 0;
-        size_t *pMemDumpBytesWritten = &undefMemDumpBytesWritten;
+        size_t* pMemDumpBytesWritten = &undefMemDumpBytesWritten;
         if (ctx.Has("MEM_DUMP_LEN") && ctx.Has("MEM_DUMP_BASE_ADDR") && ctx.Has("MEM_DUMP_BYTE_WRITTEN")) {
             memDumpMaxLen = ctx.Get<size_t>("MEM_DUMP_LEN");
             memDumpBaseAddr = ctx.Get<uint8_t*>("MEM_DUMP_BASE_ADDR");
@@ -74,8 +61,8 @@ namespace app {
         }
 
         /* Populate Pre-Processing related parameters. */
-        auto audioParamsWinLen = ctx.Get<uint32_t>("frameLength");
-        auto audioParamsWinStride = ctx.Get<uint32_t>("frameStride");
+        auto audioFrameLen = ctx.Get<uint32_t>("frameLength");
+        auto audioFrameStride = ctx.Get<uint32_t>("frameStride");
         auto nrNumInputFeatures = ctx.Get<uint32_t>("numInputFeatures");
 
         TfLiteTensor* inputTensor = model.GetInputTensor(0);
@@ -103,7 +90,7 @@ namespace app {
         if (ctx.Has("featureFileNames")) {
             audioFileAccessorFunc = ctx.Get<std::function<const char*(const uint32_t)>>("featureFileNames");
         }
-        do{
+        do {
             hal_lcd_clear(COLOR_BLACK);
 
             auto startDumpAddress = memDumpBaseAddr + memDumpBytesWritten;
@@ -112,32 +99,38 @@ namespace app {
             /* Creating a sliding window through the audio. */
             auto audioDataSlider = audio::SlidingWindow<const int16_t>(
                     audioAccessorFunc(currentIndex),
-                    audioSizeAccessorFunc(currentIndex), audioParamsWinLen,
-                    audioParamsWinStride);
+                    audioSizeAccessorFunc(currentIndex), audioFrameLen,
+                    audioFrameStride);
 
             info("Running inference on input feature map %" PRIu32 " => %s\n", currentIndex,
                  audioFileAccessorFunc(currentIndex));
 
             memDumpBytesWritten += DumpDenoisedAudioHeader(audioFileAccessorFunc(currentIndex),
-                 (audioDataSlider.TotalStrides() + 1) * audioParamsWinLen,
+                                                           (audioDataSlider.TotalStrides() + 1) * audioFrameLen,
                  memDumpBaseAddr + memDumpBytesWritten,
                  memDumpMaxLen - memDumpBytesWritten);
 
-            rnn::RNNoiseProcess featureProcessor = rnn::RNNoiseProcess();
-            rnn::vec1D32F audioFrame(audioParamsWinLen);
-            rnn::vec1D32F inputFeatures(nrNumInputFeatures);
-            rnn::vec1D32F denoisedAudioFrameFloat(audioParamsWinLen);
-            std::vector<int16_t> denoisedAudioFrame(audioParamsWinLen);
+            /* Set up pre and post-processing. */
+            std::shared_ptr<rnn::RNNoiseFeatureProcessor> featureProcessor =
+                    std::make_shared<rnn::RNNoiseFeatureProcessor>();
+            std::shared_ptr<rnn::FrameFeatures> frameFeatures =
+                    std::make_shared<rnn::FrameFeatures>();
+
+            RNNoisePreProcess preProcess = RNNoisePreProcess(inputTensor, featureProcessor, frameFeatures);
+
+            std::vector<int16_t> denoisedAudioFrame(audioFrameLen);
+            RNNoisePostProcess postProcess = RNNoisePostProcess(outputTensor, denoisedAudioFrame,
+                    featureProcessor, frameFeatures);
 
-            std::vector<float> modelOutputFloat(outputTensor->bytes);
-            rnn::FrameFeatures frameFeatures;
             bool resetGRU = true;
 
             while (audioDataSlider.HasNext()) {
                 const int16_t* inferenceWindow = audioDataSlider.Next();
-                audioFrame = rnn::vec1D32F(inferenceWindow, inferenceWindow+audioParamsWinLen);
 
-                featureProcessor.PreprocessFrame(audioFrame.data(), audioParamsWinLen, frameFeatures);
+                if (!preProcess.DoPreProcess(inferenceWindow, audioFrameLen)) {
+                    printf_err("Pre-processing failed.");
+                    return false;
+                }
 
                 /* Reset or copy over GRU states first to avoid TFLu memory overlap issues. */
                 if (resetGRU){
@@ -148,53 +141,35 @@ namespace app {
                     model.CopyGruStates();
                 }
 
-                QuantizeAndPopulateInput(frameFeatures.m_featuresVec,
-                        inputTensor->params.scale, inputTensor->params.zero_point,
-                        inputTensor);
-
                 /* Strings for presentation/logging. */
                 std::string str_inf{"Running inference... "};
 
                 /* Display message on the LCD - inference running. */
-                hal_lcd_display_text(
-                            str_inf.c_str(), str_inf.size(),
-                            dataPsnTxtInfStartX, dataPsnTxtInfStartY, false);
+                hal_lcd_display_text(str_inf.c_str(), str_inf.size(),
+                        dataPsnTxtInfStartX, dataPsnTxtInfStartY, false);
 
                 info("Inference %zu/%zu\n", audioDataSlider.Index() + 1, audioDataSlider.TotalStrides() + 1);
 
                 /* Run inference over this feature sliding window. */
-                profiler.StartProfiling("Inference");
-                bool success = model.RunInference();
-                profiler.StopProfiling();
-                resetGRU = false;
-
-                if (!success) {
+                if (!RunInference(model, profiler)) {
+                    printf_err("Inference failed.");
                     return false;
                 }
+                resetGRU = false;
 
-                /* De-quantize main model output ready for post-processing. */
-                const auto* outputData = tflite::GetTensorData<int8_t>(outputTensor);
-                auto outputQuantParams = arm::app::GetTensorQuantParams(outputTensor);
-
-                for (size_t i = 0; i < outputTensor->bytes; ++i) {
-                    modelOutputFloat[i] = (static_cast<float>(outputData[i]) - outputQuantParams.offset)
-                            * outputQuantParams.scale;
-                }
-
-                /* Round and cast the post-processed results for dumping to wav. */
-                featureProcessor.PostProcessFrame(modelOutputFloat, frameFeatures, denoisedAudioFrameFloat);
-                for (size_t i = 0; i < audioParamsWinLen; ++i) {
-                    denoisedAudioFrame[i] = static_cast<int16_t>(std::roundf(denoisedAudioFrameFloat[i]));
+                /* Carry out post-processing. */
+                if (!postProcess.DoPostProcess()) {
+                    printf_err("Post-processing failed.");
+                    return false;
                 }
 
                 /* Erase. */
                 str_inf = std::string(str_inf.size(), ' ');
-                hal_lcd_display_text(
-                                str_inf.c_str(), str_inf.size(),
-                                dataPsnTxtInfStartX, dataPsnTxtInfStartY, false);
+                hal_lcd_display_text(str_inf.c_str(), str_inf.size(),
+                        dataPsnTxtInfStartX, dataPsnTxtInfStartY, false);
 
                 if (memDumpMaxLen > 0) {
-                    /* Dump output tensors to memory. */
+                    /* Dump final post processed output to memory. */
                     memDumpBytesWritten += DumpOutputDenoisedAudioFrame(
                             denoisedAudioFrame,
                             memDumpBaseAddr + memDumpBytesWritten,
@@ -209,6 +184,7 @@ namespace app {
                      valMemDumpBytesWritten, startDumpAddress);
             }
 
+            /* Finish by dumping the footer. */
             DumpDenoisedAudioFooter(memDumpBaseAddr + memDumpBytesWritten, memDumpMaxLen - memDumpBytesWritten);
 
             info("All inferences for audio clip complete.\n");
@@ -216,15 +192,13 @@ namespace app {
             IncrementAppCtxClipIdx(ctx);
 
             std::string clearString{' '};
-            hal_lcd_display_text(
-                    clearString.c_str(), clearString.size(),
+            hal_lcd_display_text(clearString.c_str(), clearString.size(),
                     dataPsnTxtInfStartX, dataPsnTxtInfStartY, false);
 
             std::string completeMsg{"Inference complete!"};
 
             /* Display message on the LCD - inference complete. */
-            hal_lcd_display_text(
-                    completeMsg.c_str(), completeMsg.size(),
+            hal_lcd_display_text(completeMsg.c_str(), completeMsg.size(),
                     dataPsnTxtInfStartX, dataPsnTxtInfStartY, false);
 
         } while (runAll && ctx.Get<uint32_t>("clipIndex") != startClipIdx);
@@ -233,7 +207,7 @@ namespace app {
     }
 
     size_t DumpDenoisedAudioHeader(const char* filename, size_t dumpSize,
-                                   uint8_t *memAddress, size_t memSize){
+                                   uint8_t* memAddress, size_t memSize){
 
         if (memAddress == nullptr){
             return 0;
@@ -284,7 +258,7 @@ namespace app {
         return numBytesWritten;
     }
 
-    size_t DumpDenoisedAudioFooter(uint8_t *memAddress, size_t memSize){
+    size_t DumpDenoisedAudioFooter(uint8_t* memAddress, size_t memSize){
         if ((memAddress == nullptr) || (memSize < 4)) {
             return 0;
         }
@@ -294,8 +268,8 @@ namespace app {
         return sizeof(int32_t);
      }
 
-    size_t DumpOutputDenoisedAudioFrame(const std::vector<int16_t> &audioFrame,
-                                        uint8_t *memAddress, size_t memSize)
+    size_t DumpOutputDenoisedAudioFrame(const std::vector<int16_t>& audioFrame,
+                                        uint8_t* memAddress, size_t memSize)
     {
         if (memAddress == nullptr) {
             return 0;
@@ -324,7 +298,7 @@ namespace app {
             const TfLiteTensor* tensor = model.GetOutputTensor(i);
             const auto* tData = tflite::GetTensorData<uint8_t>(tensor);
 #if VERIFY_TEST_OUTPUT
-            arm::app::DumpTensor(tensor);
+            DumpTensor(tensor);
 #endif /* VERIFY_TEST_OUTPUT */
             /* Ensure that we don't overflow the allowed limit. */
             if (numBytesWritten + tensor->bytes <= memSize) {
@@ -360,20 +334,5 @@ namespace app {
         ctx.Set<uint32_t>("clipIndex", curClipIdx);
     }
 
-    void QuantizeAndPopulateInput(rnn::vec1D32F& inputFeatures,
-            const float quantScale, const int quantOffset, TfLiteTensor* inputTensor)
-    {
-        const float minVal = std::numeric_limits<int8_t>::min();
-        const float maxVal = std::numeric_limits<int8_t>::max();
-
-        auto* inputTensorData = tflite::GetTensorData<int8_t>(inputTensor);
-
-        for (size_t i=0; i < inputFeatures.size(); ++i) {
-            float quantValue = ((inputFeatures[i] / quantScale) + quantOffset);
-            inputTensorData[i] = static_cast<int8_t>(std::min<float>(std::max<float>(quantValue, minVal), maxVal));
-        }
-    }
-
-
 } /* namespace app */
 } /* namespace arm */