1. AGC 初識(shí)

自動(dòng)增益控制電路的作用是：當(dāng)輸入信號(hào)電壓變化很大時(shí)，保持接收機(jī)輸出電壓恒定或基本不變。具體地說(shuō)，當(dāng)輸入信號(hào)很弱時(shí)，接收機(jī)的增益大，自動(dòng)增益控制電路不起作用；當(dāng)輸入信號(hào)很強(qiáng)時(shí)，自動(dòng)增益控制電路進(jìn)行控制，使接收機(jī)的增益減小。這樣，當(dāng)接收信號(hào)強(qiáng)度變化時(shí)，接收機(jī)的輸出端的電壓或功率基本不變或保持恒定。因此對(duì)AGC電路的要求是：在輸入信號(hào)較小時(shí)，AGC電路不起作用，只有當(dāng)輸入信號(hào)增大到一定程度后，AGC電路才起控制作用，使增益隨輸入信號(hào)的增大而減少。
為實(shí)現(xiàn)上述要求，必須有一個(gè)能隨外來(lái)信號(hào)強(qiáng)弱而變化的控制電壓或電流信號(hào)，利用這個(gè)信號(hào)對(duì)放大器的增益自動(dòng)進(jìn)行控制。由上述分析可知，調(diào)幅中頻信號(hào)經(jīng)幅度檢波后，在它的輸出中除音頻信號(hào)外，還含有直流分量。直流分量大小與中頻載波的振幅成正比，也即與外來(lái)高頻信號(hào)成正比。因此，可將檢波器輸出的直流分量作為AGC控制信號(hào)。

AGC電路工作原理：可以分為增益受控放大電路和控制電壓形成電路。增益受控放大電路位于正向放大通路，其增益隨控制電壓U0而改變?？刂齐妷盒纬呻娐返幕静考茿GC整流器和低通平滑濾波器，有時(shí)也包含門電路和直流放大器等部件。

放大器及AGC電路.png

2. webrtc 的 AGC算法

AGC是自動(dòng)增益補(bǔ)償功能（Automatic Gain Control），AGC可以自動(dòng)調(diào)麥克風(fēng)的收音量，使與會(huì)者收到一定的音量水平，不會(huì)因發(fā)言者與麥克風(fēng)的距離改變時(shí)，聲音有忽大忽小聲的缺點(diǎn)。
webbrtc中的結(jié)構(gòu)如下：

AGC算法目錄結(jié)構(gòu).png

包裝的頭文件目錄.png

說(shuō)明：gain_control.h是包裝的頭文件，在apm里頭gain_control_impl調(diào)用。主要包括了接口定義函數(shù)和參數(shù)配置。

3. 主要配置

include/gain_control.h 里面定義了agc的三種模式，kAdaptiveAnalog、kAdaptiveDigital和kFixedDigital。其中，kAdaptiveAnalog帶有模擬音量調(diào)節(jié)的功能。kAdaptiveDigital是可變?cè)鲆鎍gc，但是不調(diào)節(jié)系統(tǒng)音量。kFixedDigital是固定增益的agc。

enum Mode {
    // Adaptive mode intended for use if an analog volume control is available
    // on the capture device. It will require the user to provide coupling
    // between the OS mixer controls and AGC through the |stream_analog_level()|
    // functions.
    //
    // It consists of an analog gain prescription for the audio device and a
    // digital compression stage.
    kAdaptiveAnalog,

    // Adaptive mode intended for situations in which an analog volume control
    // is unavailable. It operates in a similar fashion to the adaptive analog
    // mode, but with scaling instead applied in the digital domain. As with
    // the analog mode, it additionally uses a digital compression stage.
    kAdaptiveDigital,

    // Fixed mode which enables only the digital compression stage also used by
    // the two adaptive modes.
    //
    // It is distinguished from the adaptive modes by considering only a
    // short time-window of the input signal. It applies a fixed gain through
    // most of the input level range, and compresses (gradually reduces gain
    // with increasing level) the input signal at higher levels. This mode is
    // preferred on embedded devices where the capture signal level is
    // predictable, so that a known gain can be applied.
    kFixedDigital
  };

analog_agc.h 定義了配置targetLevelDbfs和compressionGaindB用于調(diào)節(jié)agc的動(dòng)態(tài)范圍

typedef struct {
  // Configurable parameters/variables
  uint32_t fs;                // Sampling frequency
  int16_t compressionGaindB;  // Fixed gain level in dB
  int16_t targetLevelDbfs;    // Target level in -dBfs of envelope (default -3)
  int16_t agcMode;            // Hard coded mode (adaptAna/adaptDig/fixedDig)
  uint8_t limiterEnable;      // Enabling limiter (on/off (default off))
  WebRtcAgcConfig defaultConfig;
  WebRtcAgcConfig usedConfig;

  // General variables
  int16_t initFlag;
  int16_t lastError;

  // Target level parameters
  // Based on the above: analogTargetLevel = round((32767*10^(-22/20))^2*16/2^7)
  int32_t analogTargetLevel;    // = RXX_BUFFER_LEN * 846805;       -22 dBfs
  int32_t startUpperLimit;      // = RXX_BUFFER_LEN * 1066064;      -21 dBfs
  int32_t startLowerLimit;      // = RXX_BUFFER_LEN * 672641;       -23 dBfs
  int32_t upperPrimaryLimit;    // = RXX_BUFFER_LEN * 1342095;      -20 dBfs
  int32_t lowerPrimaryLimit;    // = RXX_BUFFER_LEN * 534298;       -24 dBfs
  int32_t upperSecondaryLimit;  // = RXX_BUFFER_LEN * 2677832;      -17 dBfs
  int32_t lowerSecondaryLimit;  // = RXX_BUFFER_LEN * 267783;       -27 dBfs
  uint16_t targetIdx;           // Table index for corresponding target level
#ifdef MIC_LEVEL_FEEDBACK
  uint16_t targetIdxOffset;  // Table index offset for level compensation
#endif
  int16_t analogTarget;  // Digital reference level in ENV scale

  // Analog AGC specific variables
  int32_t filterState[8];  // For downsampling wb to nb
  int32_t upperLimit;      // Upper limit for mic energy
  int32_t lowerLimit;      // Lower limit for mic energy
  int32_t Rxx160w32;       // Average energy for one frame
  int32_t Rxx16_LPw32;     // Low pass filtered subframe energies
  int32_t Rxx160_LPw32;    // Low pass filtered frame energies
  int32_t Rxx16_LPw32Max;  // Keeps track of largest energy subframe
  int32_t Rxx16_vectorw32[RXX_BUFFER_LEN];  // Array with subframe energies
  int32_t Rxx16w32_array[2][5];  // Energy values of microphone signal
  int32_t env[2][10];            // Envelope values of subframes

  int16_t Rxx16pos;          // Current position in the Rxx16_vectorw32
  int16_t envSum;            // Filtered scaled envelope in subframes
  int16_t vadThreshold;      // Threshold for VAD decision
  int16_t inActive;          // Inactive time in milliseconds
  int16_t msTooLow;          // Milliseconds of speech at a too low level
  int16_t msTooHigh;         // Milliseconds of speech at a too high level
  int16_t changeToSlowMode;  // Change to slow mode after some time at target
  int16_t firstCall;         // First call to the process-function
  int16_t msZero;            // Milliseconds of zero input
  int16_t msecSpeechOuterChange;  // Min ms of speech between volume changes
  int16_t msecSpeechInnerChange;  // Min ms of speech between volume changes
  int16_t activeSpeech;           // Milliseconds of active speech
  int16_t muteGuardMs;            // Counter to prevent mute action
  int16_t inQueue;                // 10 ms batch indicator

  // Microphone level variables
  int32_t micRef;         // Remember ref. mic level for virtual mic
  uint16_t gainTableIdx;  // Current position in virtual gain table
  int32_t micGainIdx;     // Gain index of mic level to increase slowly
  int32_t micVol;         // Remember volume between frames
  int32_t maxLevel;       // Max possible vol level, incl dig gain
  int32_t maxAnalog;      // Maximum possible analog volume level
  int32_t maxInit;        // Initial value of "max"
  int32_t minLevel;       // Minimum possible volume level
  int32_t minOutput;      // Minimum output volume level
  int32_t zeroCtrlMax;    // Remember max gain => don't amp low input
  int32_t lastInMicLevel;

  int16_t scale;  // Scale factor for internal volume levels
#ifdef MIC_LEVEL_FEEDBACK
  int16_t numBlocksMicLvlSat;
  uint8_t micLvlSat;
#endif
  // Structs for VAD and digital_agc
  AgcVad vadMic;
  DigitalAgc digitalAgc;

#ifdef WEBRTC_AGC_DEBUG_DUMP
  FILE* fpt;
  FILE* agcLog;
  int32_t fcount;
#endif

  int16_t lowLevelSignal;
} LegacyAgc;

相關(guān)源代碼如下：

int GainControlImpl::AnalyzeCaptureAudio(AudioBuffer* audio) {
  if (!enabled_) {
    return AudioProcessing::kNoError;
  }

  RTC_DCHECK(num_proc_channels_);
  RTC_DCHECK_GE(AudioBuffer::kMaxSplitFrameLength,
                audio->num_frames_per_band());
  RTC_DCHECK_EQ(audio->num_channels(), *num_proc_channels_);
  RTC_DCHECK_LE(*num_proc_channels_, gain_controllers_.size());

  int16_t split_band_data[AudioBuffer::kMaxNumBands]
                         [AudioBuffer::kMaxSplitFrameLength];
  int16_t* split_bands[AudioBuffer::kMaxNumBands] = {
      split_band_data[0], split_band_data[1], split_band_data[2]};

  if (mode_ == kAdaptiveAnalog) {
    int capture_channel = 0;
    for (auto& gain_controller : gain_controllers_) {
      gain_controller->set_capture_level(analog_capture_level_);

      audio->ExportSplitChannelData(capture_channel, split_bands);

      int err =
          WebRtcAgc_AddMic(gain_controller->state(), split_bands,
                           audio->num_bands(), audio->num_frames_per_band());

      audio->ImportSplitChannelData(capture_channel, split_bands);

      if (err != AudioProcessing::kNoError) {
        return AudioProcessing::kUnspecifiedError;
      }
      ++capture_channel;
    }
  } else if (mode_ == kAdaptiveDigital) {
    int capture_channel = 0;
    for (auto& gain_controller : gain_controllers_) {
      int32_t capture_level_out = 0;

      audio->ExportSplitChannelData(capture_channel, split_bands);

      int err =
          WebRtcAgc_VirtualMic(gain_controller->state(), split_bands,
                               audio->num_bands(), audio->num_frames_per_band(),
                               analog_capture_level_, &capture_level_out);

      audio->ImportSplitChannelData(capture_channel, split_bands);

      gain_controller->set_capture_level(capture_level_out);

      if (err != AudioProcessing::kNoError) {
        return AudioProcessing::kUnspecifiedError;
      }
      ++capture_channel;
    }
  }

  return AudioProcessing::kNoError;
}

4. 主要接口

analog_agc.h包括模擬的agc結(jié)構(gòu)體聲明，而gain_control.h中的接口函數(shù)在analog_agc.c中實(shí)現(xiàn)。

WebRtcAgc_AddFarend  計(jì)算遠(yuǎn)端信號(hào)的語(yǔ)音活度VAD
WebRtcAgc_AddMic 計(jì)算麥克風(fēng)輸入的語(yǔ)音活度，對(duì)于非常小的信號(hào)會(huì)乘增益系數(shù)
WebRtcAgc_VirtualMic    用虛擬的麥克風(fēng)音量來(lái)調(diào)節(jié)幅度
WebRtcAgc_Process   vad核心處理
WebRtcAgc_set_config    設(shè)置VAD參數(shù)

在analog_agc.c還包括以下函數(shù)：

WebRtcAgc_UpdateAgcThresholds   
WebRtcAgc_SaturationCtrl    
WebRtcAgc_ZeroCtrl  
WebRtcAgc_SpeakerInactiveCtrl   
WebRtcAgc_ExpCurve  
WebRtcAgc_ProcessAnalog 

digital_agc.h包括數(shù)字的agc結(jié)構(gòu)體聲明，Vad結(jié)構(gòu)聲明，而gain_control.h中的接口函數(shù)在analog_agc.c中實(shí)現(xiàn)。

WebRtcAgc_ProcessDigital    
WebRtcAgc_AddFarendToDigital    
WebRtcAgc_InitVad   
WebRtcAgc_ProcessVad    
WebRtcAgc_CalculateGainTable    

參考文章：
Webrtc AGC 算法（二）