blob: 3b46d16afeff4e6296ab61d8bc80f57435ab16ea [file] [log] [blame]
ekm030249d2015-06-15 13:02:24 -07001/*
2 * Copyright (c) 2014 The WebRTC project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_INTELLIGIBILITY_INTELLIGIBILITY_ENHANCER_H_
12#define WEBRTC_MODULES_AUDIO_PROCESSING_INTELLIGIBILITY_INTELLIGIBILITY_ENHANCER_H_
13
14#include <complex>
kwiberg88788ad2016-02-19 07:04:49 -080015#include <memory>
ekm35b72fb2015-07-10 14:11:52 -070016#include <vector>
ekm030249d2015-06-15 13:02:24 -070017
18#include "webrtc/common_audio/lapped_transform.h"
ekmeyerson60d9b332015-08-14 10:35:55 -070019#include "webrtc/common_audio/channel_buffer.h"
ekm030249d2015-06-15 13:02:24 -070020#include "webrtc/modules/audio_processing/intelligibility/intelligibility_utils.h"
Alejandro Luebs18fcbcf2016-02-22 15:57:38 -080021#include "webrtc/modules/audio_processing/vad/voice_activity_detector.h"
ekm030249d2015-06-15 13:02:24 -070022
ekm030249d2015-06-15 13:02:24 -070023namespace webrtc {
24
25// Speech intelligibility enhancement module. Reads render and capture
26// audio streams and modifies the render stream with a set of gains per
27// frequency bin to enhance speech against the noise background.
Alejandro Luebs32348192016-02-17 20:04:19 -080028// Details of the model and algorithm can be found in the original paper:
29// http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6882788
aluebs0a007592016-02-26 17:17:38 -080030class IntelligibilityEnhancer : public LappedTransform::Callback {
ekm030249d2015-06-15 13:02:24 -070031 public:
Alejandro Luebs18fcbcf2016-02-22 15:57:38 -080032 IntelligibilityEnhancer(int sample_rate_hz, size_t num_render_channels);
ekm030249d2015-06-15 13:02:24 -070033
aluebsc466bad2016-02-10 12:03:00 -080034 // Sets the capture noise magnitude spectrum estimate.
35 void SetCaptureNoiseEstimate(std::vector<float> noise);
ekmb7553df2015-06-16 18:57:32 -070036
ekmdb4fecf2015-06-22 17:49:08 -070037 // Reads chunk of speech in time domain and updates with modified signal.
ekmeyerson60d9b332015-08-14 10:35:55 -070038 void ProcessRenderAudio(float* const* audio,
39 int sample_rate_hz,
Peter Kasting69558702016-01-12 16:26:35 -080040 size_t num_channels);
ekmeyerson60d9b332015-08-14 10:35:55 -070041 bool active() const;
ekmdb4fecf2015-06-22 17:49:08 -070042
aluebs0a007592016-02-26 17:17:38 -080043 protected:
44 // All in frequency domain, receives input |in_block|, applies
45 // intelligibility enhancement, and writes result to |out_block|.
46 void ProcessAudioBlock(const std::complex<float>* const* in_block,
47 size_t in_channels,
48 size_t frames,
49 size_t out_channels,
50 std::complex<float>* const* out_block) override;
51
ekm030249d2015-06-15 13:02:24 -070052 private:
ekm35b72fb2015-07-10 14:11:52 -070053 FRIEND_TEST_ALL_PREFIXES(IntelligibilityEnhancerTest, TestErbCreation);
54 FRIEND_TEST_ALL_PREFIXES(IntelligibilityEnhancerTest, TestSolveForGains);
ekm030249d2015-06-15 13:02:24 -070055
ekm35b72fb2015-07-10 14:11:52 -070056 // Bisection search for optimal |lambda|.
aluebsf99af6b2016-02-24 17:25:42 -080057 void SolveForLambda(float power_target);
ekm35b72fb2015-07-10 14:11:52 -070058
59 // Transforms freq gains to ERB gains.
60 void UpdateErbGains();
61
ekmdb4fecf2015-06-22 17:49:08 -070062 // Returns number of ERB filters.
Peter Kastingdce40cf2015-08-24 14:52:23 -070063 static size_t GetBankSize(int sample_rate, size_t erb_resolution);
ekmdb4fecf2015-06-22 17:49:08 -070064
65 // Initializes ERB filterbank.
aluebsc466bad2016-02-10 12:03:00 -080066 std::vector<std::vector<float>> CreateErbBank(size_t num_freqs);
ekmdb4fecf2015-06-22 17:49:08 -070067
68 // Analytically solves quadratic for optimal gains given |lambda|.
69 // Negative gains are set to 0. Stores the results in |sols|.
Peter Kastingdce40cf2015-08-24 14:52:23 -070070 void SolveForGainsGivenLambda(float lambda, size_t start_freq, float* sols);
ekmdb4fecf2015-06-22 17:49:08 -070071
Alejandro Luebs18fcbcf2016-02-22 15:57:38 -080072 // Returns true if the audio is speech.
73 bool IsSpeech(const float* audio);
74
Peter Kastingdce40cf2015-08-24 14:52:23 -070075 const size_t freqs_; // Num frequencies in frequency domain.
Peter Kastingdce40cf2015-08-24 14:52:23 -070076 const size_t chunk_length_; // Chunk size in samples.
77 const size_t bank_size_; // Num ERB filters.
ekm030249d2015-06-15 13:02:24 -070078 const int sample_rate_hz_;
Peter Kasting69558702016-01-12 16:26:35 -080079 const size_t num_render_channels_;
ekmeyerson60d9b332015-08-14 10:35:55 -070080
Alejandro Luebs18fcbcf2016-02-22 15:57:38 -080081 intelligibility::PowerEstimator<std::complex<float>> clear_power_estimator_;
82 std::unique_ptr<intelligibility::PowerEstimator<float>>
83 noise_power_estimator_;
aluebs0a007592016-02-26 17:17:38 -080084 std::vector<float> filtered_clear_pow_;
85 std::vector<float> filtered_noise_pow_;
86 std::vector<float> center_freqs_;
aluebsc466bad2016-02-10 12:03:00 -080087 std::vector<std::vector<float>> capture_filter_bank_;
88 std::vector<std::vector<float>> render_filter_bank_;
Peter Kastingdce40cf2015-08-24 14:52:23 -070089 size_t start_freq_;
Alejandro Luebs18fcbcf2016-02-22 15:57:38 -080090
aluebs0a007592016-02-26 17:17:38 -080091 std::vector<float> gains_eq_; // Pre-filter modified gains.
ekm030249d2015-06-15 13:02:24 -070092 intelligibility::GainApplier gain_applier_;
93
kwiberg88788ad2016-02-19 07:04:49 -080094 std::unique_ptr<LappedTransform> render_mangler_;
Alejandro Luebs18fcbcf2016-02-22 15:57:38 -080095
96 VoiceActivityDetector vad_;
97 std::vector<int16_t> audio_s16_;
98 size_t chunks_since_voice_;
99 bool is_speech_;
ekm030249d2015-06-15 13:02:24 -0700100};
101
102} // namespace webrtc
103
104#endif // WEBRTC_MODULES_AUDIO_PROCESSING_INTELLIGIBILITY_INTELLIGIBILITY_ENHANCER_H_