tomwalters@284: // Copyright 2010, Thomas Walters
tomwalters@284: //
tomwalters@284: // AIM-C: A C++ implementation of the Auditory Image Model
tomwalters@284: // http://www.acousticscale.org/AIMC
tomwalters@284: //
tomwalters@318: // Licensed under the Apache License, Version 2.0 (the "License");
tomwalters@318: // you may not use this file except in compliance with the License.
tomwalters@318: // You may obtain a copy of the License at
tomwalters@284: //
tomwalters@318: //     http://www.apache.org/licenses/LICENSE-2.0
tomwalters@284: //
tomwalters@318: // Unless required by applicable law or agreed to in writing, software
tomwalters@318: // distributed under the License is distributed on an "AS IS" BASIS,
tomwalters@318: // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
tomwalters@318: // See the License for the specific language governing permissions and
tomwalters@318: // limitations under the License.
tomwalters@284: 
tomwalters@284: /*!
tomwalters@284:  * \author Thomas Walters <tom@acousticscale.org>
tomwalters@284:  * \date created 2010/02/19
tomwalters@284:  * \version \$Id$
tomwalters@284:  */
tomwalters@284: 
tomwalters@287: #include <cmath>
tomwalters@287: 
tomwalters@284: #include "Modules/SSI/ModuleSSI.h"
tomwalters@284: 
tomwalters@284: namespace aimc {
tomwalters@330: #ifdef _MSC_VER
tomwalters@330: // MSVC doesn't define log2()
tomwalters@330: float log2(float n) {
tomwalters@330:   return log(n) / log(2.0);  
tomwalters@330: }
tomwalters@330: #endif
tomwalters@330: 
tomwalters@284: ModuleSSI::ModuleSSI(Parameters *params) : Module(params) {
tomwalters@284:   module_description_ = "Size-shape image (aka the 'sscAI')";
tomwalters@284:   module_identifier_ = "ssi";
tomwalters@284:   module_type_ = "ssi";
tomwalters@284:   module_version_ = "$Id$";
tomwalters@284: 
tomwalters@305:   // Cut off the SSI at the end of the first cycle
tomwalters@305:   do_pitch_cutoff_ = parameters_->DefaultBool("ssi.pitch_cutoff", false);
tomwalters@305: 
tomwalters@305:   // Weight the values in each channel more strongly if the channel was
tomwalters@305:   // truncated due to the pitch cutoff. This ensures that the same amount of
tomwalters@305:   // energy remains in the SSI spectral profile
tomwalters@305:   weight_by_cutoff_ = parameters_->DefaultBool("ssi.weight_by_cutoff", false);
tomwalters@305: 
tomwalters@305:   // Weight the values in each channel more strongly if the channel was
tomwalters@305:   // scaled such that the end goes off the edge of the computed SSI.
tomwalters@305:   // Again, this ensures that the overall energy of the spectral profile
tomwalters@305:   // remains the same.
tomwalters@305:   weight_by_scaling_ = parameters_->DefaultBool("ssi.weight_by_scaling",
tomwalters@305:                                                 false);
tomwalters@305: 
tomwalters@305:   // Time from the zero-lag line of the SAI from which to start searching
tomwalters@305:   // for a maximum in the input SAI's temporal profile.
tomwalters@305:   pitch_search_start_ms_ = parameters_->DefaultFloat(
tomwalters@411:       "ssi.pitch_search_start_ms", 2.0f);
tomwalters@305: 
tomwalters@305:   // Total width in cycles of the whole SSI
tomwalters@305:   ssi_width_cycles_ = parameters_->DefaultFloat("ssi.width_cycles", 10.0f);
tomwalters@305: 
tomwalters@305:   // Set to true to make the cycles axis logarithmic (ie indexing by gamma
tomwalters@305:   // rather than by cycles) 
tomwalters@305:   log_cycles_axis_ = parameters_->DefaultBool("ssi.log_cycles_axis", true);
tomwalters@305: 
tomwalters@305:   // The centre frequency of the channel which will just fill the complete
tomwalters@305:   // width of the SSI buffer
tomwalters@305:   pivot_cf_ = parameters_->DefaultFloat("ssi.pivot_cf", 1000.0f);
tomwalters@397:   
tomwalters@397:   // Whether or not to do smooth offset when the pitch cutoff is active.
tomwalters@397:   do_smooth_offset_ = parameters_->DefaultBool("ssi.do_smooth_offset", false);
tomwalters@397:   
tomwalters@397:   // The number of cycles, centered on the pitch line, over which the SSI is taken
tomwalters@397:   // to zero when doing the pitch cutoff.
tomwalters@397:   smooth_offset_cycles_ = parameters_->DefaultFloat("ssi.smooth_offset_cycles", 3.0f);
tomwalters@284: }
tomwalters@284: 
tomwalters@284: ModuleSSI::~ModuleSSI() {
tomwalters@284: }
tomwalters@284: 
tomwalters@284: bool ModuleSSI::InitializeInternal(const SignalBank &input) {
tomwalters@284:   // Copy the parameters of the input signal bank into internal variables, so
tomwalters@284:   // that they can be checked later.
tomwalters@284:   sample_rate_ = input.sample_rate();
tomwalters@284:   buffer_length_ = input.buffer_length();
tomwalters@284:   channel_count_ = input.channel_count();
tomwalters@284: 
tomwalters@305:   ssi_width_samples_ = sample_rate_ * ssi_width_cycles_ / pivot_cf_;
tomwalters@287:   if (ssi_width_samples_ > buffer_length_) {
tomwalters@287:     ssi_width_samples_ = buffer_length_;
tomwalters@305:     float cycles = ssi_width_samples_ * pivot_cf_ / sample_rate_;
tomwalters@287:     LOG_INFO(_T("Requested SSI width of %f cycles is too long for the "
tomwalters@287:                 "input buffer length of %d samples. The SSI will be "
tomwalters@287:                 "truncated at %d samples wide. This corresponds to a width "
tomwalters@287:                 "of %f cycles."), ssi_width_cycles_, buffer_length_,
tomwalters@287:                 ssi_width_samples_, cycles);
tomwalters@287:     ssi_width_cycles_ = cycles;
tomwalters@287:   }
tomwalters@411:   for (int i = 0; i < input.channel_count(); ++i) {
tomwalters@411:     output_.set_centre_frequency(i, input.centre_frequency(i));
tomwalters@411:   }
tomwalters@411:   
tom@420:   h_.resize(ssi_width_samples_, 0.0);
tom@420:   float gamma_min = -1.0f;
tom@420:   float gamma_max = log2(ssi_width_cycles_);
tom@420:   for (int i = 0; i < ssi_width_samples_; ++i) {
tom@420:     if (log_cycles_axis_) {
tom@420:       float gamma = gamma_min + (gamma_max - gamma_min)
tom@420:                                  * static_cast<float>(i)
tom@420:                                  / static_cast<float>(ssi_width_samples_);
tom@420:       h_[i]= pow(2.0f, gamma);
tom@420:     } else {
tom@420:       h_[i] = static_cast<float>(i) * ssi_width_cycles_
tom@420:               / static_cast<float>(ssi_width_samples_);
tom@420:     }
tom@420:   }
tom@420:   
tomwalters@287:   output_.Initialize(channel_count_, ssi_width_samples_, sample_rate_);
tomwalters@284:   return true;
tomwalters@284: }
tomwalters@284: 
tomwalters@284: void ModuleSSI::ResetInternal() {
tomwalters@284: }
tomwalters@284: 
tomwalters@305: int ModuleSSI::ExtractPitchIndex(const SignalBank &input) const {
tomwalters@305:   // Generate temporal profile of the SAI
tomwalters@305:   vector<float> sai_temporal_profile(buffer_length_, 0.0f);
tomwalters@305:   for (int i = 0; i < buffer_length_; ++i) {
tomwalters@305:     float val = 0.0f;
tomwalters@305:     for (int ch = 0; ch < channel_count_; ++ch) {
tomwalters@305:       val += input.sample(ch, i);
tomwalters@305:     }
tomwalters@305:     sai_temporal_profile[i] = val;
tomwalters@305:   }
tomwalters@305: 
tomwalters@305:   // Find pitch value
tomwalters@305:   int start_sample = floor(pitch_search_start_ms_ * sample_rate_ / 1000.0f);
tomwalters@305:   int max_idx = 0;
tomwalters@305:   float max_val = 0.0f;
tomwalters@305:   for (int i = start_sample; i < buffer_length_; ++i) {
tomwalters@305:     if (sai_temporal_profile[i] > max_val) {
tomwalters@305:       max_idx = i;
tomwalters@305:       max_val = sai_temporal_profile[i];
tomwalters@305:     }
tomwalters@305:   }
tomwalters@305:   return max_idx;
tomwalters@305: }
tomwalters@305: 
tomwalters@284: void ModuleSSI::Process(const SignalBank &input) {
tomwalters@284:   // Check to see if the module has been initialized. If not, processing
tomwalters@284:   // should not continue.
tomwalters@284:   if (!initialized_) {
tomwalters@285:     LOG_ERROR(_T("Module %s not initialized."), module_identifier_.c_str());
tomwalters@284:     return;
tomwalters@284:   }
tomwalters@284: 
tomwalters@284:   // Check that ths input this time is the same as the input passed to
tomwalters@284:   // Initialize()
tomwalters@284:   if (buffer_length_ != input.buffer_length()
tomwalters@284:       || channel_count_ != input.channel_count()) {
tomwalters@284:     LOG_ERROR(_T("Mismatch between input to Initialize() and input to "
tomwalters@285:                  "Process() in module %s."), module_identifier_.c_str());
tomwalters@284:     return;
tomwalters@284:   }
tomwalters@284: 
tomwalters@287:   output_.set_start_time(input.start_time());
tomwalters@284: 
tomwalters@305:   int pitch_index = buffer_length_ - 1;
tomwalters@305:   if (do_pitch_cutoff_) {
tomwalters@305:     pitch_index = ExtractPitchIndex(input);
tomwalters@305:   }
tomwalters@305: 
tomwalters@287:   for (int ch = 0; ch < channel_count_; ++ch) {
tomwalters@305:     float centre_frequency = input.centre_frequency(ch);
tomwalters@411:     float cycle_samples = sample_rate_ / centre_frequency;
tomwalters@397:     
tomwalters@397:     float channel_weight = 1.0f;
tomwalters@397:     int cutoff_index = buffer_length_ - 1;
tomwalters@397:     if (do_pitch_cutoff_) {
tomwalters@397:       if (pitch_index < cutoff_index) {
tomwalters@397:         if (weight_by_cutoff_) {
tomwalters@397:           channel_weight = static_cast<float>(buffer_length_)
tomwalters@411:                            / static_cast<float>(pitch_index);
tomwalters@397:         }
tomwalters@397:         cutoff_index = pitch_index;
tomwalters@397:       }
tomwalters@397:     }
tomwalters@397:     
tomwalters@397:     // tanh(3) is about 0.995. Seems reasonable. 
tomwalters@411:     float smooth_pitch_constant = 3.0f / smooth_offset_cycles_;
tomwalters@397:     float pitch_h = 0.0f;
tomwalters@397:     if (do_smooth_offset_) {
tomwalters@411:       pitch_h = static_cast<float>(pitch_index) / cycle_samples;
tomwalters@397:     }
tomwalters@397:     
tomwalters@411:     // Copy the buffer from input to output, addressing by h-value.
tomwalters@287:     for (int i = 0; i < ssi_width_samples_; ++i) {
tom@420:       
tomwalters@287:       // The index into the input array is a floating-point number, which is
tomwalters@287:       // split into a whole part and a fractional part. The whole part and
tomwalters@287:       // fractional part are found, and are used to linearly interpolate
tomwalters@287:       // between input samples to yield an output sample.
tomwalters@287:       double whole_part;
tom@420:       float frac_part = modf(h_[i] * cycle_samples, &whole_part);
tomwalters@305:       int sample = floor(whole_part);
tomwalters@305: 
tomwalters@397:       float weight = channel_weight;
tomwalters@397:       
tomwalters@397:       if (do_smooth_offset_ && do_pitch_cutoff_) {
tomwalters@397:         // Smoothing around the pitch cutoff line.
tom@420:         float pitch_weight = (1.0f + tanh((pitch_h - h_[i])
tomwalters@411:                                           * smooth_pitch_constant)) / 2.0f;
tomwalters@411:         weight *= pitch_weight;
tomwalters@411:         //LOG_INFO("Channel %d, Sample %d. Pitch weight: %f", ch, i, pitch_weight);
tomwalters@305:       }
tomwalters@305: 
tomwalters@305:       if (weight_by_scaling_) {
tomwalters@305:         if (centre_frequency > pivot_cf_) {
tomwalters@305:           weight *= (centre_frequency / pivot_cf_);
tomwalters@305:         }
tomwalters@305:       }
tomwalters@287: 
tomwalters@287:       float val;
tomwalters@397:       if (sample < cutoff_index || do_smooth_offset_) {
tomwalters@287:         float curr_sample = input.sample(ch, sample);
tomwalters@287:         float next_sample = input.sample(ch, sample + 1);
tomwalters@305:         val = weight * (curr_sample
tomwalters@305:                         + frac_part * (next_sample - curr_sample));
tomwalters@287:       } else {
tomwalters@309:         val = 0.0f;
tomwalters@287:       }
tomwalters@287:       output_.set_sample(ch, i, val);
tomwalters@287:     }
tomwalters@287:   }
tomwalters@284:   PushOutput();
tomwalters@284: }
tomwalters@284: }  // namespace aimc
tomwalters@284: