Upgrade AudioRendererAlgorithm to use WSOLA,

media/filters/audio_renderer_algorithm.cc

Index: media/filters/audio_renderer_algorithm.cc
diff --git a/media/filters/audio_renderer_algorithm.cc b/media/filters/audio_renderer_algorithm.cc
index 97f0811384159ff35868bb45f10eabdc7a94d86b..1068a2342ed67023cf8229e94cc127a9299c0b01 100644
--- a/media/filters/audio_renderer_algorithm.cc
+++ b/media/filters/audio_renderer_algorithm.cc
@@ -12,41 +12,78 @@
 #include "media/audio/audio_util.h"
 #include "media/base/audio_buffer.h"
 #include "media/base/audio_bus.h"
+#include "media/filters/wsola_internals.h"
 
 namespace media {
 
-// The starting size in frames for |audio_buffer_|. Previous usage maintained a
-// queue of 16 AudioBuffers, each of 512 frames. This worked well, so we
-// maintain this number of frames.
-static const int kStartingBufferSizeInFrames = 16 * 512;
+
+// Waveform Similarity Overlap-and-add (WSOLA).
+//
+// One WSOLA iteration
+//
+// 1) Extract |target_block_| as input frames at indices
+//    [|target_block_index_|, |target_block_index_| + |ola_window_size_|).
+//    Note that |target_block_| is the "natural" continuation of the output.
+//
+// 2) Extract |search_block_| as input frames at indices
+//    [|search_block_index_|,
+//     |search_block_index_| + |num_candidate_blocks_| + |ola_window_size_|).
+//
+// 3) Find a block within the |search_block_| that is most similar
+//    to |target_block_|. Let |optimal_index| be the index of such block and
+//    write it to |optimal_block_|.
+//
+// 4) Update:
+//    |optimal_block_| = |transition_window_| * |target_block_| +
+//    (1 - |transition_window_|) * |optimal_block_|.
+//
+// 5) Overlap-and-add |optimal_block_| to the |wsola_output_|.
+//
+// 6) Update:
+//    |target_block_| = |optimal_index| + |ola_window_size_| / 2.
+//    |output_index_| = |output_index_| + |ola_window_size_| / 2,
+//    |search_block_center_offset_| = |output_index_| * |playback_rate_|, and
+//    |search_block_index_| = |search_block_center_offset_| -
+//        |search_block_center_offset_|.
 
 // The maximum size in frames for the |audio_buffer_|. Arbitrarily determined.
 // This number represents 3 seconds of 96kHz/16 bit 7.1 surround sound.
 static const int kMaxBufferSizeInFrames = 3 * 96000;
 
-// Duration of audio segments used for crossfading (in seconds).
-static const double kWindowDuration = 0.08;
-
-// Duration of crossfade between audio segments (in seconds).
-static const double kCrossfadeDuration = 0.008;
-
 // Max/min supported playback rates for fast/slow audio. Audio outside of these
 // ranges are muted.
 // Audio at these speeds would sound better under a frequency domain algorithm.
 static const float kMinPlaybackRate = 0.5f;
 static const float kMaxPlaybackRate = 4.0f;
 
+// Overlap-and-add window size in milliseconds.
+static const int kOlaWindowSizeMs = 20;
+
+// Size of search interval in milliseconds. The search interval is
+// [-delta delta] around |output_index_| * |playback_rate_|. So the search
+// interval is 2 * delta.
+static const int kWsolaSearchIntervalMs = 30;
+
+// The starting size in frames for |audio_buffer_|. Previous usage maintained a
+// queue of 16 AudioBuffers, each of 512 frames. This worked well, so we
+// maintain this number of frames.
+static const int kStartingBufferSizeInFrames = 16 * 512;
+
 AudioRendererAlgorithm::AudioRendererAlgorithm()
     : channels_(0),
       samples_per_second_(0),
       playback_rate_(0),
-      frames_in_crossfade_(0),
-      index_into_window_(0),
-      crossfade_frame_number_(0),
       muted_(false),
       muted_partial_frame_(0),
-      window_size_(0),
-      capacity_(kStartingBufferSizeInFrames) {
+      capacity_(kStartingBufferSizeInFrames),
+      output_time_(0),
+      search_block_center_offset_(0),
+      search_block_index_(0),
+      num_candidate_blocks_(0),
+      target_block_index_(0),
+      ola_window_size_(0),
+      ola_hop_size_(0),
+      num_complete_frames_(0) {
 }
 
 AudioRendererAlgorithm::~AudioRendererAlgorithm() {}
@@ -58,16 +95,58 @@ void AudioRendererAlgorithm::Initialize(float initial_playback_rate,
   channels_ = params.channels();
   samples_per_second_ = params.sample_rate();
   SetPlaybackRate(initial_playback_rate);
-
-  window_size_ = samples_per_second_ * kWindowDuration;
-  frames_in_crossfade_ = samples_per_second_ * kCrossfadeDuration;
-  crossfade_buffer_ = AudioBus::Create(channels_, frames_in_crossfade_);
+  num_candidate_blocks_ = (kWsolaSearchIntervalMs * samples_per_second_) / 1000;
+  ola_window_size_ = kOlaWindowSizeMs * samples_per_second_ / 1000;
+
+  // Make sure window size in an even number.
+  ola_window_size_ += ola_window_size_ & 1;
+  ola_hop_size_ = ola_window_size_ / 2;
+
+  // |num_candidate_blocks_| / 2 is the offset of the center of the search
+  // block to the center of the first (left most) candidate block. The offset
+  // of the center of a candidate block to its left most point is
+  // |ola_window_size_| / 2 - 1. Note that |ola_window_size_| is even and in
+  // our convention the center belongs to the left half, so we need to subtract
+  // one frame to get the correct offset.
+  //
+  //                             Search Block
+  //              <------------------------------------------->
+  //
+  //   |ola_window_size_| / 2 - 1
+  //              <----
+  //
+  //             |num_candidate_blocks_| / 2
+  //                   <----------------
+  //                                 center
+  //              X----X----------------X---------------X-----X
+  //              <---------->                     <---------->
+  //                Candidate      ...               Candidate
+  //                   1,          ...         |num_candidate_blocks_|
+  search_block_center_offset_ = num_candidate_blocks_ / 2 +
+      (ola_window_size_ / 2 - 1);
+
+  ola_window_.reset(new float[ola_window_size_]);
+  internal::GetSymmetricHanningWindow(ola_window_size_, ola_window_.get());
+
+  transition_window_.reset(new float[ola_window_size_ * 2]);
+  internal::GetSymmetricHanningWindow(2 * ola_window_size_,
+                                      transition_window_.get());
+
+  wsola_output_ = AudioBus::Create(channels_, ola_window_size_ + ola_hop_size_);
+
+  // Auxiliary containers.
+  optimal_block_ = AudioBus::Create(channels_, ola_window_size_);
+  search_block_ = AudioBus::Create(
+      channels_, num_candidate_blocks_ + (ola_window_size_ - 1));
+  target_block_ = AudioBus::Create(channels_, ola_window_size_);
 }
 
 int AudioRendererAlgorithm::FillBuffer(AudioBus* dest, int requested_frames) {
   if (playback_rate_ == 0)
     return 0;
 
+  DCHECK_EQ(channels_, dest->channels());
+
   // Optimize the |muted_| case to issue a single clear instead of performing
   // the full crossfade and clearing each crossfaded frame.
   if (muted_) {
@@ -93,12 +172,12 @@ int AudioRendererAlgorithm::FillBuffer(AudioBus* dest, int requested_frames) {
     return frames_to_render;
   }
 
-  int slower_step = ceil(window_size_ * playback_rate_);
-  int faster_step = ceil(window_size_ / playback_rate_);
+  int slower_step = ceil(ola_window_size_ * playback_rate_);
+  int faster_step = ceil(ola_window_size_ / playback_rate_);
 
   // Optimize the most common |playback_rate_| ~= 1 case to use a single copy
   // instead of copying frame by frame.
-  if (window_size_ <= faster_step && slower_step >= window_size_) {
+  if (ola_window_size_ <= faster_step && slower_step >= ola_window_size_) {
     const int frames_to_copy =
         std::min(audio_buffer_.frames(), requested_frames);
     const int frames_read = audio_buffer_.ReadFrames(frames_to_copy, 0, dest);
@@ -106,277 +185,200 @@ int AudioRendererAlgorithm::FillBuffer(AudioBus* dest, int requested_frames) {
     return frames_read;
   }
 
-  int total_frames_rendered = 0;
-  while (total_frames_rendered < requested_frames) {
-    if (index_into_window_ >= window_size_)
-      ResetWindow();
-
-    int rendered_frames = 0;
-    if (window_size_ > faster_step) {
-      rendered_frames =
-          OutputFasterPlayback(dest,
-                               total_frames_rendered,
-                               requested_frames - total_frames_rendered,
-                               window_size_,
-                               faster_step);
-    } else if (slower_step < window_size_) {
-      rendered_frames =
-          OutputSlowerPlayback(dest,
-                               total_frames_rendered,
-                               requested_frames - total_frames_rendered,
-                               slower_step,
-                               window_size_);
-    } else {
-      NOTREACHED();
-    }
+  int rendered_frames = 0;
+  do {
+    rendered_frames += WriteCompletedFramesTo(
+        requested_frames - rendered_frames, rendered_frames, dest);
+  } while (rendered_frames < requested_frames && RunOneWsolaIteration());
+  return rendered_frames;
+}
 
-    if (rendered_frames == 0)
-      break;
+void AudioRendererAlgorithm::SetPlaybackRate(float new_rate) {
+  DCHECK_GE(new_rate, 0);
+  playback_rate_ = new_rate;
+  muted_ =
+      playback_rate_ < kMinPlaybackRate || playback_rate_ > kMaxPlaybackRate;
+}
 
-    total_frames_rendered += rendered_frames;
-  }
-  return total_frames_rendered;
+void AudioRendererAlgorithm::FlushBuffers() {
+  // Clear the queue of decoded packets (releasing the buffers).
+  audio_buffer_.Clear();
+  output_time_ = 0;
+  search_block_index_ = 0;
+  target_block_index_ = 0;
+  wsola_output_->Zero();
+  num_complete_frames_ = 0;
 }
 
-void AudioRendererAlgorithm::ResetWindow() {
-  DCHECK_LE(index_into_window_, window_size_);
-  index_into_window_ = 0;
-  crossfade_frame_number_ = 0;
+base::TimeDelta AudioRendererAlgorithm::GetTime() {
+  return audio_buffer_.current_time();
 }
 
-int AudioRendererAlgorithm::OutputFasterPlayback(AudioBus* dest,
-                                                 int dest_offset,
-                                                 int requested_frames,
-                                                 int input_step,
-                                                 int output_step) {
-  // Ensure we don't run into OOB read/write situation.
-  CHECK_GT(input_step, output_step);
-  DCHECK_LT(index_into_window_, window_size_);
-  DCHECK_GT(playback_rate_, 1.0);
-  DCHECK(!muted_);
-
-  if (audio_buffer_.frames() < 1)
-    return 0;
+void AudioRendererAlgorithm::EnqueueBuffer(
+    const scoped_refptr<AudioBuffer>& buffer_in) {
+  DCHECK(!buffer_in->end_of_stream());
+  audio_buffer_.Append(buffer_in);
+}
 
-  // The audio data is output in a series of windows. For sped-up playback,
-  // the window is comprised of the following phases:
-  //
-  //  a) Output raw data.
-  //  b) Save bytes for crossfade in |crossfade_buffer_|.
-  //  c) Drop data.
-  //  d) Output crossfaded audio leading up to the next window.
-  //
-  // The duration of each phase is computed below based on the |window_size_|
-  // and |playback_rate_|.
-  DCHECK_LE(frames_in_crossfade_, output_step);
+bool AudioRendererAlgorithm::IsQueueFull() {
+  return audio_buffer_.frames() >= capacity_;
+}
 
-  // This is the index of the end of phase a, beginning of phase b.
-  int outtro_crossfade_begin = output_step - frames_in_crossfade_;
+void AudioRendererAlgorithm::IncreaseQueueCapacity() {
+  capacity_ = std::min(2 * capacity_, kMaxBufferSizeInFrames);
+}
 
-  // This is the index of the end of phase b, beginning of phase c.
-  int outtro_crossfade_end = output_step;
+bool AudioRendererAlgorithm::CanPerformWsola() const {
+  const int search_block_size = num_candidate_blocks_ + (ola_window_size_ - 1);
+  const int frames = audio_buffer_.frames();
+  return target_block_index_ + ola_window_size_ <= frames &&
+      search_block_index_ + search_block_size <= frames;
+}
 
-  // This is the index of the end of phase c, beginning of phase d.
-  // This phase continues until |index_into_window_| reaches |window_size_|, at
-  // which point the window restarts.
-  int intro_crossfade_begin = input_step - frames_in_crossfade_;
+bool AudioRendererAlgorithm::RunOneWsolaIteration() {
+  if (!CanPerformWsola())
+    return false;
 
-  // a) Output raw frames if we haven't reached the crossfade section.
-  if (index_into_window_ < outtro_crossfade_begin) {
-    // Read as many frames as we can and return the count. If it's not enough,
-    // we will get called again.
-    const int frames_to_copy =
-        std::min(requested_frames, outtro_crossfade_begin - index_into_window_);
-    int copied = audio_buffer_.ReadFrames(frames_to_copy, dest_offset, dest);
-    index_into_window_ += copied;
-    return copied;
-  }
-
-  // b) Save outtro crossfade frames into intermediate buffer, but do not output
-  //    anything to |dest|.
-  if (index_into_window_ < outtro_crossfade_end) {
-    // This phase only applies if there are bytes to crossfade.
-    DCHECK_GT(frames_in_crossfade_, 0);
-    int crossfade_start = index_into_window_ - outtro_crossfade_begin;
-    int crossfade_count = outtro_crossfade_end - index_into_window_;
-    int copied = audio_buffer_.ReadFrames(
-        crossfade_count, crossfade_start, crossfade_buffer_.get());
-    index_into_window_ += copied;
-
-    // Did we get all the frames we need? If not, return and let subsequent
-    // calls try to get the rest.
-    if (copied != crossfade_count)
-      return 0;
-  }
+  GetOptimalBlock();
 
-  // c) Drop frames until we reach the intro crossfade section.
-  if (index_into_window_ < intro_crossfade_begin) {
-    // Check if there is enough data to skip all the frames needed. If not,
-    // return 0 and let subsequent calls try to skip it all.
-    int seek_frames = intro_crossfade_begin - index_into_window_;
-    if (audio_buffer_.frames() < seek_frames)
-      return 0;
-    audio_buffer_.SeekFrames(seek_frames);
+  // Overlap-and-add.
+  for (int k = 0; k < channels_; ++k) {
+    const float* const ch_opt_frame = optimal_block_->channel(k);
+    float* ch_output = wsola_output_->channel(k) + num_complete_frames_;
+    for (int n = 0; n < ola_hop_size_; ++n) {
+      ch_output[n] = ch_output[n] * ola_window_[ola_hop_size_ + n] +
+          ch_opt_frame[n] * ola_window_[n];
+    }
 
-    // We've dropped all the frames that need to be dropped.
-    index_into_window_ += seek_frames;
+    // Copy the second half to the output.
+    memcpy(&ch_output[ola_hop_size_], &ch_opt_frame[ola_hop_size_],
+           sizeof(*ch_opt_frame) * ola_hop_size_);
   }
 
-  // d) Crossfade and output a frame, as long as we have data.
-  if (audio_buffer_.frames() < 1)
-    return 0;
-  DCHECK_GT(frames_in_crossfade_, 0);
-  DCHECK_LT(index_into_window_, window_size_);
-
-  int offset_into_buffer = index_into_window_ - intro_crossfade_begin;
-  int copied = audio_buffer_.ReadFrames(1, dest_offset, dest);
-  DCHECK_EQ(copied, 1);
-  CrossfadeFrame(crossfade_buffer_.get(),
-                 offset_into_buffer,
-                 dest,
-                 dest_offset,
-                 offset_into_buffer);
-  index_into_window_ += copied;
-  return copied;
+  num_complete_frames_ += ola_hop_size_;
+  UpdateOutputTime(ola_hop_size_);
+  RemoveOldInputFrames();
+  return true;
 }
 
-int AudioRendererAlgorithm::OutputSlowerPlayback(AudioBus* dest,
-                                                 int dest_offset,
-                                                 int requested_frames,
-                                                 int input_step,
-                                                 int output_step) {
-  // Ensure we don't run into OOB read/write situation.
-  CHECK_LT(input_step, output_step);
-  DCHECK_LT(index_into_window_, window_size_);
-  DCHECK_LT(playback_rate_, 1.0);
-  DCHECK_NE(playback_rate_, 0);
-  DCHECK(!muted_);
-
-  if (audio_buffer_.frames() < 1)
-    return 0;
-
-  // The audio data is output in a series of windows. For slowed down playback,
-  // the window is comprised of the following phases:
-  //
-  //  a) Output raw data.
-  //  b) Output and save bytes for crossfade in |crossfade_buffer_|.
-  //  c) Output* raw data.
-  //  d) Output* crossfaded audio leading up to the next window.
-  //
-  // * Phases c) and d) do not progress |audio_buffer_|'s cursor so that the
-  // |audio_buffer_|'s cursor is in the correct place for the next window.
-  //
-  // The duration of each phase is computed below based on the |window_size_|
-  // and |playback_rate_|.
-  DCHECK_LE(frames_in_crossfade_, input_step);
-
-  // This is the index of the end of phase a, beginning of phase b.
-  int intro_crossfade_begin = input_step - frames_in_crossfade_;
+void AudioRendererAlgorithm::UpdateOutputTime(float time_change) {
+  output_time_ += time_change;
+  // Center of the search region, in frames.
+  const int search_block_center_index = static_cast<int>(
+      output_time_ * playback_rate_ + 0.5f);
+  search_block_index_ = search_block_center_index - search_block_center_offset_;
+}
 
-  // This is the index of the end of phase b, beginning of phase c.
-  int intro_crossfade_end = input_step;
+void AudioRendererAlgorithm::RemoveOldInputFrames() {
+  const int earliest_used_index = std::min(target_block_index_,
+                                           search_block_index_);
+  if (earliest_used_index <= 0)
+    return;  // Nothing to remove.
 
-  // This is the index of the end of phase c,  beginning of phase d.
-  // This phase continues until |index_into_window_| reaches |window_size_|, at
-  // which point the window restarts.
-  int outtro_crossfade_begin = output_step - frames_in_crossfade_;
+  // Remove frames from input and adjust indices accordingly.
+  audio_buffer_.SeekFrames(earliest_used_index);
+  target_block_index_ -= earliest_used_index;
 
-  // a) Output raw frames.
-  if (index_into_window_ < intro_crossfade_begin) {
-    // Read as many frames as we can and return the count. If it's not enough,
-    // we will get called again.
-    const int frames_to_copy =
-        std::min(requested_frames, intro_crossfade_begin - index_into_window_);
-    int copied = audio_buffer_.ReadFrames(frames_to_copy, dest_offset, dest);
-    index_into_window_ += copied;
-    return copied;
-  }
+  // Adjust output index.
+  float output_time_change = earliest_used_index / playback_rate_;
+  CHECK_GE(output_time_, output_time_change);
+  UpdateOutputTime(-output_time_change);
+}
 
-  // b) Save the raw frames for the intro crossfade section, then copy the
-  //    same frames to |dest|.
-  if (index_into_window_ < intro_crossfade_end) {
-    const int frames_to_copy =
-        std::min(requested_frames, intro_crossfade_end - index_into_window_);
-    int offset = index_into_window_ - intro_crossfade_begin;
-    int copied = audio_buffer_.ReadFrames(
-        frames_to_copy, offset, crossfade_buffer_.get());
-    crossfade_buffer_->CopyPartialFramesTo(offset, copied, dest_offset, dest);
-    index_into_window_ += copied;
-    return copied;
-  }
+int AudioRendererAlgorithm::WriteCompletedFramesTo(
+    int requested_frames, int dest_offset, AudioBus* dest) {
+  int rendered_frames = std::min(num_complete_frames_, requested_frames);
 
-  // c) Output a raw frame into |dest| without advancing the |audio_buffer_|
-  //    cursor.
-  int audio_buffer_offset = index_into_window_ - intro_crossfade_end;
-  DCHECK_GE(audio_buffer_offset, 0);
-  if (audio_buffer_.frames() <= audio_buffer_offset)
-    return 0;
-  int copied =
-      audio_buffer_.PeekFrames(1, audio_buffer_offset, dest_offset, dest);
-  DCHECK_EQ(1, copied);
-
-  // d) Crossfade the next frame of |crossfade_buffer_| into |dest| if we've
-  //    reached the outtro crossfade section of the window.
-  if (index_into_window_ >= outtro_crossfade_begin) {
-    int offset_into_crossfade_buffer =
-        index_into_window_ - outtro_crossfade_begin;
-    CrossfadeFrame(dest,
-                   dest_offset,
-                   crossfade_buffer_.get(),
-                   offset_into_crossfade_buffer,
-                   offset_into_crossfade_buffer);
-  }
+  if (rendered_frames == 0)
+    return 0;  // There is nothing to read from |wsola_output_|, return.
 
-  index_into_window_ += copied;
-  return copied;
-}
+  wsola_output_->CopyPartialFramesTo(0, rendered_frames, dest_offset, dest);
 
-void AudioRendererAlgorithm::CrossfadeFrame(AudioBus* intro,
-                                            int intro_offset,
-                                            AudioBus* outtro,
-                                            int outtro_offset,
-                                            int fade_offset) {
-  float crossfade_ratio =
-      static_cast<float>(fade_offset) / frames_in_crossfade_;
-  for (int channel = 0; channel < channels_; ++channel) {
-    outtro->channel(channel)[outtro_offset] =
-        (1.0f - crossfade_ratio) * intro->channel(channel)[intro_offset] +
-        (crossfade_ratio) * outtro->channel(channel)[outtro_offset];
+  // Remove the frames which are read.
+  int frames_to_move = wsola_output_->frames() - rendered_frames;
+  for (int k = 0; k < channels_; ++k) {
+    float* ch = wsola_output_->channel(k);
+    memmove(ch, &ch[rendered_frames], sizeof(*ch) * frames_to_move);
   }
+  num_complete_frames_ -= rendered_frames;
+  return rendered_frames;
 }
 
-void AudioRendererAlgorithm::SetPlaybackRate(float new_rate) {
-  DCHECK_GE(new_rate, 0);
-  playback_rate_ = new_rate;
-  muted_ =
-      playback_rate_ < kMinPlaybackRate || playback_rate_ > kMaxPlaybackRate;
-
-  ResetWindow();
-}
-
-void AudioRendererAlgorithm::FlushBuffers() {
-  ResetWindow();
-
-  // Clear the queue of decoded packets (releasing the buffers).
-  audio_buffer_.Clear();
-}
+bool AudioRendererAlgorithm::TargetIsWithinSearchRegion() const {
+  const int search_block_size = num_candidate_blocks_ + (ola_window_size_ - 1);
 
-base::TimeDelta AudioRendererAlgorithm::GetTime() {
-  return audio_buffer_.current_time();
+  return target_block_index_ >= search_block_index_ &&
+      target_block_index_ + ola_window_size_ <=
+      search_block_index_ + search_block_size;
 }
 
-void AudioRendererAlgorithm::EnqueueBuffer(
-    const scoped_refptr<AudioBuffer>& buffer_in) {
-  DCHECK(!buffer_in->end_of_stream());
-  audio_buffer_.Append(buffer_in);
-}
+void AudioRendererAlgorithm::GetOptimalBlock() {
+  int optimal_index = 0;
+
+  // An interval around last optimal block which is excluded from the search.
+  // This is to reduce the buzzy sound. The number 160 is rather arbitrary and
+  // derived heuristically.
+  const int kExcludeIntervalLengthFrames = 160;
+  if (TargetIsWithinSearchRegion()) {
+    optimal_index = target_block_index_;
+    PeekAudioWithZeroPrepend(optimal_index, optimal_block_.get());
+  } else {
+    PeekAudioWithZeroPrepend(target_block_index_, target_block_.get());
+    PeekAudioWithZeroPrepend(search_block_index_, search_block_.get());
+    int last_optimal = target_block_index_ - ola_hop_size_ -
+        search_block_index_;
+    internal::Interval exclude_iterval = std::make_pair(
+        last_optimal - kExcludeIntervalLengthFrames / 2,
+        last_optimal + kExcludeIntervalLengthFrames / 2);
+
+    // |optimal_index| is in frames and it is relative to the beginning of the
+    // |search_block_|.
+    optimal_index = internal::OptimalIndex(
+        search_block_.get(), target_block_.get(), exclude_iterval);
+
+    // Translate |index| w.r.t. the beginning of |audio_buffer_| and extract the
+    // optimal block.
+    optimal_index += search_block_index_;
+    PeekAudioWithZeroPrepend(optimal_index, optimal_block_.get());
+
+    // Make a transition from target block to the optimal block if different.
+    // Target block has the best continuation to the current output.
+    // Optimal block is the most similar block to the target, however, it might
+    // introduce some discontinuity when over-lap-added. Therefore, we combine
+    // them for a smoother transition. The length of transition window is twice
+    // as that of the optimal-block which makes it like a weighting function
+    // where target-block has higher weight close to zero (weight of 1 at index
+    // 0) and lower weight close the end.
+    for (int k = 0; k < channels_; ++k) {
+      float* ch_opt = optimal_block_->channel(k);
+      const float* const ch_target = target_block_->channel(k);
+      for (int n = 0; n < ola_window_size_; ++n) {
+        ch_opt[n] = ch_opt[n] * transition_window_[n] + ch_target[n] *
+            transition_window_[ola_window_size_ + n];
+      }
+    }
+  }
 
-bool AudioRendererAlgorithm::IsQueueFull() {
-  return audio_buffer_.frames() >= capacity_;
+  // Next target is one hop ahead of the current optimal.
+  target_block_index_ = optimal_index + ola_hop_size_;
 }
 
-void AudioRendererAlgorithm::IncreaseQueueCapacity() {
-  capacity_ = std::min(2 * capacity_, kMaxBufferSizeInFrames);
+void AudioRendererAlgorithm::PeekAudioWithZeroPrepend(
+    int read_offset_frames, AudioBus* dest) {
+  CHECK_LE(read_offset_frames + dest->frames(), audio_buffer_.frames());
+
+  int write_offset = 0;
+  int num_frames_to_read = dest->frames();
+  if (read_offset_frames < 0) {
+    int num_zero_frames_appended = std::min(-read_offset_frames,
+                                            num_frames_to_read);
+    read_offset_frames = 0;
+    num_frames_to_read -= num_zero_frames_appended;
+    write_offset = num_zero_frames_appended;
+    dest->ZeroFrames(num_zero_frames_appended);
+  }
+  audio_buffer_.PeekFrames(num_frames_to_read, read_offset_frames,
+                           write_offset, dest);
 }
 
 }  // namespace media