michael@0: // Copyright (c) 2013 The Chromium Authors. All rights reserved.
michael@0: //
michael@0: // Redistribution and use in source and binary forms, with or without
michael@0: // modification, are permitted provided that the following conditions are
michael@0: // met:
michael@0: //
michael@0: //    * Redistributions of source code must retain the above copyright
michael@0: // notice, this list of conditions and the following disclaimer.
michael@0: //    * Redistributions in binary form must reproduce the above
michael@0: // copyright notice, this list of conditions and the following disclaimer
michael@0: // in the documentation and/or other materials provided with the
michael@0: // distribution.
michael@0: //    * Neither the name of Google Inc. nor the names of its
michael@0: // contributors may be used to endorse or promote products derived from
michael@0: // this software without specific prior written permission.
michael@0: //
michael@0: // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
michael@0: // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
michael@0: // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
michael@0: // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
michael@0: // OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
michael@0: // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
michael@0: // LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
michael@0: // DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
michael@0: // THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
michael@0: // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
michael@0: // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
michael@0: 
michael@0: #include "energy_endpointer.h"
michael@0: 
michael@0: #include <math.h>
michael@0: 
michael@0: namespace {
michael@0: 
michael@0: // Returns the RMS (quadratic mean) of the input signal.
michael@0: float RMS(const int16_t* samples, int num_samples) {
michael@0:   int64_t ssq_int64_t = 0;
michael@0:   int64_t sum_int64_t = 0;
michael@0:   for (int i = 0; i < num_samples; ++i) {
michael@0:     sum_int64_t += samples[i];
michael@0:     ssq_int64_t += samples[i] * samples[i];
michael@0:   }
michael@0:   // now convert to floats.
michael@0:   double sum = static_cast<double>(sum_int64_t);
michael@0:   sum /= num_samples;
michael@0:   double ssq = static_cast<double>(ssq_int64_t);
michael@0:   return static_cast<float>(sqrt((ssq / num_samples) - (sum * sum)));
michael@0: }
michael@0: 
michael@0: int64_t Secs2Usecs(float seconds) {
michael@0:   return static_cast<int64_t>(0.5 + (1.0e6 * seconds));
michael@0: }
michael@0: 
michael@0: float GetDecibel(float value) {
michael@0:   if (value > 1.0e-100)
michael@0:     return 20 * log10(value);
michael@0:   return -2000.0;
michael@0: }
michael@0: 
michael@0: }  // namespace
michael@0: 
michael@0: namespace mozilla {
michael@0: 
michael@0: // Stores threshold-crossing histories for making decisions about the speech
michael@0: // state.
michael@0: class EnergyEndpointer::HistoryRing {
michael@0:  public:
michael@0:   HistoryRing() : insertion_index_(0) {}
michael@0: 
michael@0:   // Resets the ring to |size| elements each with state |initial_state|
michael@0:   void SetRing(int size, bool initial_state);
michael@0: 
michael@0:   // Inserts a new entry into the ring and drops the oldest entry.
michael@0:   void Insert(int64_t time_us, bool decision);
michael@0: 
michael@0:   // Returns the time in microseconds of the most recently added entry.
michael@0:   int64_t EndTime() const;
michael@0: 
michael@0:   // Returns the sum of all intervals during which 'decision' is true within
michael@0:   // the time in seconds specified by 'duration'. The returned interval is
michael@0:   // in seconds.
michael@0:   float RingSum(float duration_sec);
michael@0: 
michael@0:  private:
michael@0:   struct DecisionPoint {
michael@0:     int64_t time_us;
michael@0:     bool decision;
michael@0:   };
michael@0: 
michael@0:   std::vector<DecisionPoint> decision_points_;
michael@0:   int insertion_index_;  // Index at which the next item gets added/inserted.
michael@0: 
michael@0:   HistoryRing(const HistoryRing&);
michael@0:   void operator=(const HistoryRing&);
michael@0: };
michael@0: 
michael@0: void EnergyEndpointer::HistoryRing::SetRing(int size, bool initial_state) {
michael@0:   insertion_index_ = 0;
michael@0:   decision_points_.clear();
michael@0:   DecisionPoint init = { -1, initial_state };
michael@0:   decision_points_.resize(size, init);
michael@0: }
michael@0: 
michael@0: void EnergyEndpointer::HistoryRing::Insert(int64_t time_us, bool decision) {
michael@0:   decision_points_[insertion_index_].time_us = time_us;
michael@0:   decision_points_[insertion_index_].decision = decision;
michael@0:   insertion_index_ = (insertion_index_ + 1) % decision_points_.size();
michael@0: }
michael@0: 
michael@0: int64_t EnergyEndpointer::HistoryRing::EndTime() const {
michael@0:   int ind = insertion_index_ - 1;
michael@0:   if (ind < 0)
michael@0:     ind = decision_points_.size() - 1;
michael@0:   return decision_points_[ind].time_us;
michael@0: }
michael@0: 
michael@0: float EnergyEndpointer::HistoryRing::RingSum(float duration_sec) {
michael@0:   if (!decision_points_.size())
michael@0:     return 0.0;
michael@0: 
michael@0:   int64_t sum_us = 0;
michael@0:   int ind = insertion_index_ - 1;
michael@0:   if (ind < 0)
michael@0:     ind = decision_points_.size() - 1;
michael@0:   int64_t end_us = decision_points_[ind].time_us;
michael@0:   bool is_on = decision_points_[ind].decision;
michael@0:   int64_t start_us = end_us - static_cast<int64_t>(0.5 + (1.0e6 * duration_sec));
michael@0:   if (start_us < 0)
michael@0:     start_us = 0;
michael@0:   size_t n_summed = 1;  // n points ==> (n-1) intervals
michael@0:   while ((decision_points_[ind].time_us > start_us) &&
michael@0:          (n_summed < decision_points_.size())) {
michael@0:     --ind;
michael@0:     if (ind < 0)
michael@0:       ind = decision_points_.size() - 1;
michael@0:     if (is_on)
michael@0:       sum_us += end_us - decision_points_[ind].time_us;
michael@0:     is_on = decision_points_[ind].decision;
michael@0:     end_us = decision_points_[ind].time_us;
michael@0:     n_summed++;
michael@0:   }
michael@0: 
michael@0:   return 1.0e-6f * sum_us;  //  Returns total time that was super threshold.
michael@0: }
michael@0: 
michael@0: EnergyEndpointer::EnergyEndpointer()
michael@0:     : status_(EP_PRE_SPEECH),
michael@0:       offset_confirm_dur_sec_(0),
michael@0:       endpointer_time_us_(0),
michael@0:       fast_update_frames_(0),
michael@0:       frame_counter_(0),
michael@0:       max_window_dur_(4.0),
michael@0:       sample_rate_(0),
michael@0:       history_(new HistoryRing()),
michael@0:       decision_threshold_(0),
michael@0:       estimating_environment_(false),
michael@0:       noise_level_(0),
michael@0:       rms_adapt_(0),
michael@0:       start_lag_(0),
michael@0:       end_lag_(0),
michael@0:       user_input_start_time_us_(0) {
michael@0: }
michael@0: 
michael@0: EnergyEndpointer::~EnergyEndpointer() {
michael@0: }
michael@0: 
michael@0: int EnergyEndpointer::TimeToFrame(float time) const {
michael@0:   return static_cast<int32_t>(0.5 + (time / params_.frame_period()));
michael@0: }
michael@0: 
michael@0: void EnergyEndpointer::Restart(bool reset_threshold) {
michael@0:   status_ = EP_PRE_SPEECH;
michael@0:   user_input_start_time_us_ = 0;
michael@0: 
michael@0:   if (reset_threshold) {
michael@0:     decision_threshold_ = params_.decision_threshold();
michael@0:     rms_adapt_ = decision_threshold_;
michael@0:     noise_level_ = params_.decision_threshold() / 2.0f;
michael@0:     frame_counter_ = 0;  // Used for rapid initial update of levels.
michael@0:   }
michael@0: 
michael@0:   // Set up the memories to hold the history windows.
michael@0:   history_->SetRing(TimeToFrame(max_window_dur_), false);
michael@0: 
michael@0:   // Flag that indicates that current input should be used for
michael@0:   // estimating the environment. The user has not yet started input
michael@0:   // by e.g. pressed the push-to-talk button. By default, this is
michael@0:   // false for backward compatibility.
michael@0:   estimating_environment_ = false;
michael@0: }
michael@0: 
michael@0: void EnergyEndpointer::Init(const EnergyEndpointerParams& params) {
michael@0:   params_ = params;
michael@0: 
michael@0:   // Find the longest history interval to be used, and make the ring
michael@0:   // large enough to accommodate that number of frames.  NOTE: This
michael@0:   // depends upon ep_frame_period being set correctly in the factory
michael@0:   // that did this instantiation.
michael@0:   max_window_dur_ = params_.onset_window();
michael@0:   if (params_.speech_on_window() > max_window_dur_)
michael@0:     max_window_dur_ = params_.speech_on_window();
michael@0:   if (params_.offset_window() > max_window_dur_)
michael@0:     max_window_dur_ = params_.offset_window();
michael@0:   Restart(true);
michael@0: 
michael@0:   offset_confirm_dur_sec_ = params_.offset_window() -
michael@0:                             params_.offset_confirm_dur();
michael@0:   if (offset_confirm_dur_sec_ < 0.0)
michael@0:     offset_confirm_dur_sec_ = 0.0;
michael@0: 
michael@0:   user_input_start_time_us_ = 0;
michael@0: 
michael@0:   // Flag that indicates that  current input should be used for
michael@0:   // estimating the environment. The user has not yet started input
michael@0:   // by e.g. pressed the push-to-talk button. By default, this is
michael@0:   // false for backward compatibility.
michael@0:   estimating_environment_ = false;
michael@0:   // The initial value of the noise and speech levels is inconsequential.
michael@0:   // The level of the first frame will overwrite these values.
michael@0:   noise_level_ = params_.decision_threshold() / 2.0f;
michael@0:   fast_update_frames_ =
michael@0:       static_cast<int64_t>(params_.fast_update_dur() / params_.frame_period());
michael@0: 
michael@0:   frame_counter_ = 0;  // Used for rapid initial update of levels.
michael@0: 
michael@0:   sample_rate_ = params_.sample_rate();
michael@0:   start_lag_ = static_cast<int>(sample_rate_ /
michael@0:                                 params_.max_fundamental_frequency());
michael@0:   end_lag_ = static_cast<int>(sample_rate_ /
michael@0:                               params_.min_fundamental_frequency());
michael@0: }
michael@0: 
michael@0: void EnergyEndpointer::StartSession() {
michael@0:   Restart(true);
michael@0: }
michael@0: 
michael@0: void EnergyEndpointer::EndSession() {
michael@0:   status_ = EP_POST_SPEECH;
michael@0: }
michael@0: 
michael@0: void EnergyEndpointer::SetEnvironmentEstimationMode() {
michael@0:   Restart(true);
michael@0:   estimating_environment_ = true;
michael@0: }
michael@0: 
michael@0: void EnergyEndpointer::SetUserInputMode() {
michael@0:   estimating_environment_ = false;
michael@0:   user_input_start_time_us_ = endpointer_time_us_;
michael@0: }
michael@0: 
michael@0: void EnergyEndpointer::ProcessAudioFrame(int64_t time_us,
michael@0:                                          const int16_t* samples,
michael@0:                                          int num_samples,
michael@0:                                          float* rms_out) {
michael@0:   endpointer_time_us_ = time_us;
michael@0:   float rms = RMS(samples, num_samples);
michael@0: 
michael@0:   // Check that this is user input audio vs. pre-input adaptation audio.
michael@0:   // Input audio starts when the user indicates start of input, by e.g.
michael@0:   // pressing push-to-talk. Audio recieved prior to that is used to update
michael@0:   // noise and speech level estimates.
michael@0:   if (!estimating_environment_) {
michael@0:     bool decision = false;
michael@0:     if ((endpointer_time_us_ - user_input_start_time_us_) <
michael@0:         Secs2Usecs(params_.contamination_rejection_period())) {
michael@0:       decision = false;
michael@0:       //PR_LOG(GetSpeechRecognitionLog(), PR_LOG_DEBUG, ("decision: forced to false, time: %d", endpointer_time_us_));
michael@0:     } else {
michael@0:       decision = (rms > decision_threshold_);
michael@0:     }
michael@0: 
michael@0:     history_->Insert(endpointer_time_us_, decision);
michael@0: 
michael@0:     switch (status_) {
michael@0:       case EP_PRE_SPEECH:
michael@0:         if (history_->RingSum(params_.onset_window()) >
michael@0:             params_.onset_detect_dur()) {
michael@0:           status_ = EP_POSSIBLE_ONSET;
michael@0:         }
michael@0:         break;
michael@0: 
michael@0:       case EP_POSSIBLE_ONSET: {
michael@0:         float tsum = history_->RingSum(params_.onset_window());
michael@0:         if (tsum > params_.onset_confirm_dur()) {
michael@0:           status_ = EP_SPEECH_PRESENT;
michael@0:         } else {  // If signal is not maintained, drop back to pre-speech.
michael@0:           if (tsum <= params_.onset_detect_dur())
michael@0:             status_ = EP_PRE_SPEECH;
michael@0:         }
michael@0:         break;
michael@0:       }
michael@0: 
michael@0:       case EP_SPEECH_PRESENT: {
michael@0:         // To induce hysteresis in the state residency, we allow a
michael@0:         // smaller residency time in the on_ring, than was required to
michael@0:         // enter the SPEECH_PERSENT state.
michael@0:         float on_time = history_->RingSum(params_.speech_on_window());
michael@0:         if (on_time < params_.on_maintain_dur())
michael@0:           status_ = EP_POSSIBLE_OFFSET;
michael@0:         break;
michael@0:       }
michael@0: 
michael@0:       case EP_POSSIBLE_OFFSET:
michael@0:         if (history_->RingSum(params_.offset_window()) <=
michael@0:             offset_confirm_dur_sec_) {
michael@0:           // Note that this offset time may be beyond the end
michael@0:           // of the input buffer in a real-time system.  It will be up
michael@0:           // to the RecognizerSession to decide what to do.
michael@0:           status_ = EP_PRE_SPEECH;  // Automatically reset for next utterance.
michael@0:         } else {  // If speech picks up again we allow return to SPEECH_PRESENT.
michael@0:           if (history_->RingSum(params_.speech_on_window()) >=
michael@0:               params_.on_maintain_dur())
michael@0:             status_ = EP_SPEECH_PRESENT;
michael@0:         }
michael@0:         break;
michael@0: 
michael@0:       default:
michael@0:         break;
michael@0:     }
michael@0: 
michael@0:     // If this is a quiet, non-speech region, slowly adapt the detection
michael@0:     // threshold to be about 6dB above the average RMS.
michael@0:     if ((!decision) && (status_ == EP_PRE_SPEECH)) {
michael@0:       decision_threshold_ = (0.98f * decision_threshold_) + (0.02f * 2 * rms);
michael@0:       rms_adapt_ = decision_threshold_;
michael@0:     } else {
michael@0:       // If this is in a speech region, adapt the decision threshold to
michael@0:       // be about 10dB below the average RMS. If the noise level is high,
michael@0:       // the threshold is pushed up.
michael@0:       // Adaptation up to a higher level is 5 times faster than decay to
michael@0:       // a lower level.
michael@0:       if ((status_ == EP_SPEECH_PRESENT) && decision) {
michael@0:         if (rms_adapt_ > rms) {
michael@0:           rms_adapt_ = (0.99f * rms_adapt_) + (0.01f * rms);
michael@0:         } else {
michael@0:           rms_adapt_ = (0.95f * rms_adapt_) + (0.05f * rms);
michael@0:         }
michael@0:         float target_threshold = 0.3f * rms_adapt_ +  noise_level_;
michael@0:         decision_threshold_ = (.90f * decision_threshold_) +
michael@0:                               (0.10f * target_threshold);
michael@0:       }
michael@0:     }
michael@0: 
michael@0:     // Set a floor
michael@0:     if (decision_threshold_ < params_.min_decision_threshold())
michael@0:       decision_threshold_ = params_.min_decision_threshold();
michael@0:   }
michael@0: 
michael@0:   // Update speech and noise levels.
michael@0:   UpdateLevels(rms);
michael@0:   ++frame_counter_;
michael@0: 
michael@0:   if (rms_out)
michael@0:     *rms_out = GetDecibel(rms);
michael@0: }
michael@0: 
michael@0: float EnergyEndpointer::GetNoiseLevelDb() const {
michael@0:   return GetDecibel(noise_level_);
michael@0: }
michael@0: 
michael@0: void EnergyEndpointer::UpdateLevels(float rms) {
michael@0:   // Update quickly initially. We assume this is noise and that
michael@0:   // speech is 6dB above the noise.
michael@0:   if (frame_counter_ < fast_update_frames_) {
michael@0:     // Alpha increases from 0 to (k-1)/k where k is the number of time
michael@0:     // steps in the initial adaptation period.
michael@0:     float alpha = static_cast<float>(frame_counter_) /
michael@0:         static_cast<float>(fast_update_frames_);
michael@0:     noise_level_ = (alpha * noise_level_) + ((1 - alpha) * rms);
michael@0:     //PR_LOG(GetSpeechRecognitionLog(), PR_LOG_DEBUG, ("FAST UPDATE, frame_counter_ %d, fast_update_frames_ %d", frame_counter_, fast_update_frames_));
michael@0:   } else {
michael@0:     // Update Noise level. The noise level adapts quickly downward, but
michael@0:     // slowly upward. The noise_level_ parameter is not currently used
michael@0:     // for threshold adaptation. It is used for UI feedback.
michael@0:     if (noise_level_ < rms)
michael@0:       noise_level_ = (0.999f * noise_level_) + (0.001f * rms);
michael@0:     else
michael@0:       noise_level_ = (0.95f * noise_level_) + (0.05f * rms);
michael@0:   }
michael@0:   if (estimating_environment_ || (frame_counter_ < fast_update_frames_)) {
michael@0:     decision_threshold_ = noise_level_ * 2; // 6dB above noise level.
michael@0:     // Set a floor
michael@0:     if (decision_threshold_ < params_.min_decision_threshold())
michael@0:       decision_threshold_ = params_.min_decision_threshold();
michael@0:   }
michael@0: }
michael@0: 
michael@0: EpStatus EnergyEndpointer::Status(int64_t* status_time)  const {
michael@0:   *status_time = history_->EndTime();
michael@0:   return status_;
michael@0: }
michael@0: 
michael@0: }  // namespace mozilla