ronw@642: // Copyright 2013, Google, Inc.
ronw@642: // Author: Ron Weiss <ronw@google.com>
ronw@642: //
ronw@642: // This C++ file is part of an implementation of Lyon's cochlear model:
ronw@642: // "Cascade of Asymmetric Resonators with Fast-Acting Compression"
ronw@642: // to supplement Lyon's upcoming book "Human and Machine Hearing"
ronw@642: //
ronw@642: // Licensed under the Apache License, Version 2.0 (the "License");
ronw@642: // you may not use this file except in compliance with the License.
ronw@642: // You may obtain a copy of the License at
ronw@642: //
ronw@642: //     http://www.apache.org/licenses/LICENSE-2.0
ronw@642: //
ronw@642: // Unless required by applicable law or agreed to in writing, software
ronw@642: // distributed under the License is distributed on an "AS IS" BASIS,
ronw@642: // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
ronw@642: // See the License for the specific language governing permissions and
ronw@642: // limitations under the License.
ronw@642: 
ronw@642: #include <assert.h>
ronw@642: 
ronw@642: #include "sai.h"
ronw@642: 
ronw@642: SAI::SAI(const SAIParams& params) : params_(params) {
ronw@642:   assert(params_.window_width > params_.width &&
ronw@642:          "SAI window_width must be larger than width.");
ronw@642: 
ronw@642:   int buffer_width = params_.width +
ronw@642:       static_cast<int>((1 + static_cast<float>(params_.n_window_pos - 1)/2) *
ronw@642:                        params_.window_width);
ronw@642:   input_buffer_.setZero(params_.n_ch, buffer_width);
ronw@642:   output_buffer_.setZero(params_.n_ch, params_.width);
ronw@642: 
ronw@642:   window_.setLinSpaced(params_.window_width, kPi / params_.window_width, kPi)
ronw@642:       .sin();
ronw@642: }
ronw@642: 
alexbrandmeyer@643: void SAI::RunSegment(const std::vector<ArrayX>& input,
alexbrandmeyer@643:                      ArrayXX* output_frame) {
ronw@642:   assert(!input.empty() || input.size() <= params_.window_width &&
ronw@642:          "Unexpected input size.");
ronw@642:   assert(input[0].size() == params_.n_ch &&
ronw@642:          "Unexpected input frame size.");
ronw@642: 
ronw@642:   // Append new data to the input buffer.
ronw@642:   int n_shift = input.size();
ronw@642:   int shift_width = input_buffer_.cols() - n_shift;
ronw@642:   input_buffer_.topLeftCorner(params_.n_ch, shift_width).swap(
ronw@642:       input_buffer_.block(0, n_shift, params_.n_ch, shift_width));
ronw@642:   for (int i = 0; i < input.size(); ++i) {
ronw@642:     input_buffer_.block(0, shift_width + i, input[i].size(), 1) = input[i];
ronw@642:   }
ronw@642:   // Zero-pad the buffer if necessary.
ronw@642:   if (input.size() < params_.window_width) {
ronw@642:     int pad_width = params_.window_width - input.size();
ronw@642:     input_buffer_.topRightCorner(params_.n_ch, pad_width).setZero();
ronw@642:   }
ronw@642: 
ronw@642:   StabilizeSegment(input_buffer_, &output_buffer_);
ronw@642:   *output_frame = output_buffer_;
ronw@642: }
ronw@642: 
alexbrandmeyer@643: void SAI::StabilizeSegment(const ArrayXX& input_buffer,
alexbrandmeyer@643:                            ArrayXX* output_buffer) const {
ronw@642:   // Windows are always approximately 50% overlapped.
ronw@642:   float window_hop = params_.window_width / 2;
ronw@642:   int window_start = (input_buffer.cols() - params_.window_width) -
ronw@642:       (params_.n_window_pos - 1) * window_hop;
ronw@642:   int window_range_start = window_start - params_.future_lags - 1;
ronw@642:   int offset_range_start = window_start - params_.width;
ronw@642:   assert(offset_range_start >= 0);
ronw@642:   for (int i = 0; i < params_.n_ch; ++i) {
ronw@642:     // TODO(ronw): Rename this here and in the Matlab code since the
ronw@642:     // input doesn't have to contain naps.
alexbrandmeyer@643:     const ArrayX& nap_wave = input_buffer.row(i);
ronw@642:     // TODO(ronw): Smooth row.
ronw@642: 
ronw@642:     for (int w = 0; w < params_.n_window_pos; ++w) {
ronw@642:       int current_window_offset = w * window_hop;
ronw@642:       // Choose a trigger point.
ronw@642:       int trigger_time;
alexbrandmeyer@643:       const ArrayX& trigger_window =
ronw@642:           nap_wave.segment(window_range_start + current_window_offset,
ronw@642:                            params_.window_width);
ronw@642:       FPType peak_val = (trigger_window * window_).maxCoeff(&trigger_time);
ronw@642:       if (peak_val <= 0) {
ronw@642:         peak_val = window_.maxCoeff(&trigger_time);
ronw@642:       }
ronw@642:       trigger_time += current_window_offset;
ronw@642: 
ronw@642:       // Blend the window following the trigger into the output
ronw@642:       // buffer, weighted according to the the trigger strength (0.05
ronw@642:       // to near 1.0).
ronw@642:       FPType alpha = (0.025 + peak_val) / (0.5 + peak_val);
ronw@642:       output_buffer->row(i) *= 1 - alpha;
ronw@642:       output_buffer->row(i) += alpha *
ronw@642:           nap_wave.segment(trigger_time + offset_range_start, params_.width);
ronw@642:     }
ronw@642:   }
ronw@642: }