Adding audio RepetitionDetector in AudioProcessingModule.

webrtc/modules/audio_processing/repetition_detector.h

+/*
+ *  Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_REPETITION_DETECTOR_H_
+#define WEBRTC_MODULES_AUDIO_PROCESSING_REPETITION_DETECTOR_H_
+
+#include <vector>
+#include <memory>

[hlundin-webrtc, 2015/08/31 13:46:46]

Order of includes.

[minyue-webrtc, 2015/09/03 13:23:58]

Yes, but memory is not needed any longer now

+
+#include "webrtc/base/constructormagic.h"
+#include "webrtc/typedefs.h"
+
+namespace webrtc {
+
+class RepetitionDetector {
+ public:
+  RepetitionDetector();
+  virtual ~RepetitionDetector();
+
+  struct Pattern {
+    int id_;
+    // All followings are in milliseconds, since repetition patterns are
+    // supposedly bounded to certain duration in time.
+    int look_back_ms_;
+    int min_length_ms_;
+  };
+
+  // Detect repetition given audio samples. When multichannel, samples should be
+  // interleaved and |bytes_per_sample| is the number of bytes of a stereo
+  // sample.
+  void Detect(const void* data, size_t bytes_per_sample,

[Andrew MacDonald, 2015/08/28 17:43:42]

Don't use void. Be confident about your type safety! Use templates or overloaded
methods.

[minyue-webrtc, 2015/09/01 10:21:48]

A benefit of this is that we can compare even larger sizes (a stereo float
sample is 2 * sizeof(float), e.g.), but of course, we can make channels explicit
too.

[ajm, 2015/09/02 05:28:28]

Yes, make them explicit. I see that lack of type safety as a clear negative, not
a benefit.

[minyue-webrtc, 2015/09/03 13:23:58]

Ok. I take it. To make it simple enough, I even restrict it to float.

+              size_t samples_per_channel, int sample_rate_hz);

[Andrew MacDonald, 2015/08/28 17:43:42]

We've been using num_frames for the samples_per_channel quantity in newer code.
This is consistent with Chromium.

[minyue-webrtc, 2015/09/01 10:21:48]

Ok. will change.

+
+ protected:
+  void RegisterRepetitionPatterns(const Pattern* patterns,

[Andrew MacDonald, 2015/08/28 17:43:42]

Why do you need these protected methods? Is it just for the test? Instead
consider a public RegisterRepetitionPatternsForTest or make this private and
friend the needed test.

+                                  size_t num_patterns);
+  void ClearRepetitionPatterns();
+  virtual void ReportRepetition(int id) { }
+
+ private:
+  class State {
+   public:
+    State(int id, int look_back_ms, int min_length_ms);
+    void Increment(bool zero);
+    bool HasValidReport(int sample_rate_khz) const;
+    bool AlreadyReported() const;
+    void SetReported();
+    void Reset();
+    int id() const { return id_; }
+    int look_back_ms() const { return look_back_ms_; }
+
+   private:
+    const int id_;
+    const int look_back_ms_;
+    const int min_length_ms_;
+    size_t count_samples_;
+    bool all_zero_;
+    bool reported_;
+  };
+
+  void Reset(size_t bytes_per_sample, int sample_rate_hz);
+
+  void AddSampleToBuffer(const void* sample);
+
+  std::vector<State*> states_;
+  int max_look_back_ms_;
+
+  std::unique_ptr<char[]> audio_buffer_;  // Ring buffers to store input audio.

[Andrew MacDonald, 2015/08/28 17:43:42]

unique_ptr is disallowed because it's C++11 standard library. What's wrong with
vectors?

Also did you consider AudioRingBuffer? It may already do exactly what you want.
(I'm not certain).

[minyue-webrtc, 2015/09/01 10:21:48]

Now I tried to switch to RingBuffer and I see a point that AudioRingBuffer does
not fit well. I need random access to the data, for different look back values.
But AudioRingBuffer (and RingBuffer) uses an internal read pointer.

[ajm, 2015/09/02 05:28:28]

Ah OK. Use a vector here then.

[minyue-webrtc, 2015/09/03 13:23:58]

will scoped_ptr be better, I feel that vector has a lot of overhead, and it is
weird or maybe wrong if we do something like
float* sample = &audio_buffer_[0]
if (*(sample + 1) == 0)
...

[Andrew MacDonald, 2015/09/07 06:52:18]

You can expect an empty vector to consume 12 bytes (pointer + size + capacity).
You store an explicit pointer and size here so the overhead is 4 bytes. Don't
worry about it.

audio_buffer_ can be a vector in your example and it will behave exactly as you
would expect. There is no problem.

+  size_t bytes_per_sample_;  // Number of bytes in each sample.
+  int sample_rate_hz_;  // Sample rate in kHz.

[hlundin-webrtc, 2015/08/31 13:46:46]

Name says Hz, comment says kHz.

[minyue-webrtc, 2015/09/01 10:21:48]

Thanks, it was kHz, but to handle 44.1, I made it Hz

+  size_t buffer_size_samples_;  // Number of samples in |audio_buffer|.
+  size_t buffer_end_index_;  // The index of the last sample in |audio_buffer|.
+
+  DISALLOW_COPY_AND_ASSIGN(RepetitionDetector);
+};
+
+}  // namespace webrtc
+
+#endif  // WEBRTC_MODULES_AUDIO_PROCESSING_REPETITION_DETECTOR_H_