Echo cancellation delay time estimation for webrtc

It is not reasonable to estimate the delay time called echo cancellation. The core here is to estimate the condition boundary of calling webrtc. We all know that the prerequisite for webrtc echo cancellation to take effect is to obtain the information of the far-end sound, and then get the near-end sound and the far-end sound. Alignment, from the near-end sound, combined with the far-end sound to simulate the echo information of the far-end sound in the near-end sound and then kill the echo information. Therefore, if there is no corresponding sound for the far-end sound saved in the alignment, webrtc The echo cancellation effect will be poor or ineffective, so this time difference is very important, and it is necessary to comprehensively weigh delay_time in combination with the hardware situation.

The echo cancellation (aec, aecm) algorithm of webrtc mainly includes the following important modules:

• Echo delay estimation
• NLMS (Normalized Least Mean Square Adaptive Algorithm)
• NLP (Nonlinear Filtering)
• CNG (Comfort Noise Generation)

Echo delay estimation

 

There are many things in this picture that can be ignored. Let's focus on T0, T1, and T2.

• T0 represents the time for the sound to pass from the speaker to the microphone. This time can be ignored, because generally the distance between the microphone and the speaker is not too far. Considering the speed of sound at 340 meters per second, this time will not exceed 1 millisecond. .
• T1 represents the sound coming to you from a distance, the time from when this sound is passed to the echo cancellation remote interface (WebRtcAec_BufferFarend) to when it is played. Generally speaking, when the received audio data is transmitted to this interface, it is the moment when the upper layer transmits to the speaker, so it can be understood as the time when the sound starts timing in the playback queue and is played out.
• T2 represents the moment when a piece of sound is collected by the speaker, and then sent to the near-end processing function (WebRtcAec_Process). Since the sound is collected, the echo cancellation process will be performed immediately, so this time can be understood as the start of timing when the sound is collected by the microphone, and then The time it takes for your code to get the audio PCM data.
• delay=T0+T1+T2, which is actually T1+T2.

Generally speaking, if a device can find a suitable delay, then the echo cancellation processing of this device is almost as difficult as the noise reduction gain. For example, the fixed delay of iPhone is 60ms. However, this depends on the location of the code. If it is inside the chip, the time is relatively small and it is easy to fix. If it is in the system application layer software, the entire time is uncertain. Relatively large. apm_->set_stream_delay_ms(delay_ms_);

For example, the package call of webrtc under the android system:

#include "webrtc_apm.h"
//#include "webrtc/common_types.h"
#include "webrtc/modules/audio_processing/include/audio_processing.h"
#include "webrtc/modules/include/module_common_types.h"
#include "webrtc/api/audio/audio_frame.h"
#include "YuvConvert.h"

using namespace webrtc;
//using namespace cbase;

WebrtcAPM::WebrtcAPM(int process_smp, int reverse_smp)
: apm_(nullptr)
, far_frame_(new AudioFrame)
, near_frame_(new AudioFrame)
, ref_cnt_(0)
, sample_rate_(8000)
, samples_per_channel_(8000/100)
, channels_(1)
, frame_size_10ms_(8000/100*sizeof(int16_t))
, delay_ms_(60)
, process_sample_rate_(process_smp)//44100
, reverse_sample_rate_(reverse_smp)//48000
{
    audio_send_ = new unsigned char[kMaxDataSizeSamples_];
    audio_reverse_ = new unsigned char[kMaxDataSizeSamples_];

#if defined(__APPLE__)
    delay_ms_ = 60;
#endif

#if defined(VAD_TEST)
        //create webrtc vad
        ty_vad_create(8000, 1);
        ty_set_vad_level(vad_level::LOW);
        ty_vad_set_recordfile("/sdcard/vadfile.pcm");
#endif
}

WebrtcAPM::~WebrtcAPM()
{
    if(far_frame_)
    {
        delete far_frame_;
        far_frame_ = NULL ;
    }

    if (near_frame_) {
        delete near_frame_ ;
        near_frame_ = NULL ;
    }

    if (audio_send_) {
        delete[] audio_send_;
    }

    if (audio_reverse_) {
        delete[] audio_reverse_;
    }

#if defined(VAD_TEST)
        //destory webrtc vad
        ty_vad_destory();
#endif
}

void WebrtcAPM::set_sample_rate(int sample_rate)
{
    sample_rate_ = sample_rate;
    samples_per_channel_ = sample_rate_ / 100;
}

int WebrtcAPM::frame_size()
{
    return frame_size_10ms_;
}

void WebrtcAPM::set_ace_delay(int delay)
{
    LOGI("set aec delay to %d ms \n", delay);
    delay_ms_ = delay;
}

void WebrtcAPM::set_reverse_stream(int reverse_sample_rate)
{
    std::lock_guard<std::mutex> guard(mutex_);

    reverse_sample_rate_ = reverse_sample_rate;
    if (resampleReverse) {
        delete resampleReverse;
        resampleReverse = NULL ;
    }

    resampleReverse = new webrtc::Resampler(reverse_sample_rate_,sample_rate_,channels_);
    int result = resampleReverse->Reset(reverse_sample_rate_,sample_rate_,channels_);
    if (result != 0) {
        LOGE("reset resampleReverse fail,%d!\n", result);
    }
}

int WebrtcAPM::init()
{
    std::lock_guard<std::mutex> guard(mutex_);

    resampleReverse = new webrtc::Resampler(reverse_sample_rate_,sample_rate_,channels_);
    int result = resampleReverse->Reset(reverse_sample_rate_,sample_rate_,channels_);
    if (result != 0) {
        LOGE("reset resampleReverse fail,%d!\n", result);
    }

    resampleIn = new webrtc::Resampler(sample_rate_,process_sample_rate_,channels_);
    result = resampleIn->Reset(sample_rate_,process_sample_rate_,channels_);
    if (result != 0) {
        LOGE("reset resampleIn fail,%d!\n", result);
    }

    initAPM();

    LOGI("initAPM success!");

    return 0;//crash without this
}

void WebrtcAPM::uninit()
{
    std::lock_guard<std::mutex> guard(mutex_);

    deInitAPM();
    safe_delete(resampleIn);
    safe_delete(resampleReverse);
}

void WebrtcAPM::reset_apm(){
    std::lock_guard<std::mutex> guard(mutex_);
    deInitAPM();
    initAPM();
}

int WebrtcAPM::initAPM(){
    apm_ = AudioProcessingBuilder().Create();
    if (apm_ == nullptr) {
        LOGE("AudioProcessing create failed");
        return -1;
    }

    AudioProcessing::Config config;
    config.echo_canceller.enabled = true;
    config.echo_canceller.mobile_mode = true;

    config.gain_controller1.enabled = true;
    config.gain_controller1.mode = AudioProcessing::Config::GainController1::kAdaptiveDigital;
    config.gain_controller1.analog_level_minimum = 0;
    config.gain_controller1.analog_level_maximum = 255;

    config.noise_suppression.enabled = true;
    config.noise_suppression.level = AudioProcessing::Config::NoiseSuppression::Level::kModerate;

    config.gain_controller2.enabled = true;
    config.high_pass_filter.enabled = true;
    config.voice_detection.enabled = true;

    apm_->ApplyConfig(config);
    LOGI("AudioProcessing initialize success \n");

    far_frame_->sample_rate_hz_ = sample_rate_;
    far_frame_->samples_per_channel_ = samples_per_channel_;
    far_frame_->num_channels_ = channels_;

    near_frame_->sample_rate_hz_ = sample_rate_;
    near_frame_->samples_per_channel_ = samples_per_channel_;
    near_frame_->num_channels_ = channels_;

    frame_size_10ms_ = samples_per_channel_ * channels_ * sizeof(int16_t);

    //LOGI("AudioProcessing initialize success end\n");
    return 0;

}

void WebrtcAPM::deInitAPM(){
    //std::lock_guard<std::mutex> guard(mutex_);
    //if (--ref_cnt_ == 0) {
        LOGI("destroy WebrtcAPM \n");
        safe_delete(apm_);
    //}
}

//8000->44100
void WebrtcAPM::process_stream(uint8_t *buffer,int bufferLength, uint8_t *bufferOut, int* pOutLen, bool bUseAEC)
{
    if (bUseAEC) {
        std::lock_guard<std::mutex> guard(mutex_);
        if (apm_) {
            int frame_count = bufferLength / frame_size_10ms_;
        
            for (int i = 0; i < frame_count; i++) {
                 apm_->set_stream_delay_ms(delay_ms_);
                 // webrtc apm process 10ms datas every time
                 memcpy((void*)near_frame_->data(), buffer + i*frame_size_10ms_, frame_size_10ms_);
                 int res = apm_->ProcessStream(near_frame_->data(),
                                        StreamConfig(near_frame_->sample_rate_hz_, near_frame_->num_channels_),
                                        StreamConfig(near_frame_->sample_rate_hz_, near_frame_->num_channels_),
                                        (int16_t * const)near_frame_->data());
                 if (res != 0) {
                     LOGE("ProcessStream failed, ret %d \n",res);
                 }

#if  defined(VAD_TEST)                        
            bool ret = ty_vad_process(buffer + i*frame_size_10ms_, frame_size_10ms_);
#endif

                 memcpy(buffer + i*frame_size_10ms_, near_frame_->data(), frame_size_10ms_);
            }
        }
    }

    if (resampleIn && apm_) {
        //resample
        size_t outlen = 0 ;
        int result = resampleIn->Push((int16_t*)buffer, bufferLength/sizeof(int16_t), (int16_t*)bufferOut, kMaxDataSizeSamples_/sizeof(int16_t), outlen);
        if (result != 0) {
            LOGE("resampleIn error, result = %d, outlen = %d\n", result, outlen);
        }
        *pOutLen = outlen;
    }
}

//48000->8000
void WebrtcAPM::process_reverse_10ms_stream(uint8_t *bufferIn, int bufferLength, uint8_t *bufferOut, int* pOutLen, bool bUseAEC)
{
    size_t outlen = 0 ;
    if (resampleReverse && apm_) {
        //resample
        int result = resampleReverse->Push((int16_t*)bufferIn, bufferLength/sizeof(int16_t), (int16_t*)audio_reverse_, kMaxDataSizeSamples_/sizeof(int16_t), outlen);
        if (result != 0) {
            LOGE("resampleReverse error, result = %d, outlen = %d\n", result, outlen);
        }
    }
    else {
		memcpy(audio_reverse_, bufferIn, bufferLength);
		outlen = bufferLength;
	}

    *pOutLen = outlen;

    if (!bUseAEC){
        //copy data and return
        memcpy(bufferOut, audio_reverse_, frame_size_10ms_);
        return;
    }

    std::lock_guard<std::mutex> guard(mutex_);
    if (apm_) {
        memcpy((void*)far_frame_->data(), audio_reverse_, frame_size_10ms_);
        int res = apm_->ProcessReverseStream(far_frame_->data(),
                                        StreamConfig(far_frame_->sample_rate_hz_, far_frame_->num_channels_),
                                        StreamConfig(far_frame_->sample_rate_hz_, far_frame_->num_channels_),
                                        (int16_t * const)far_frame_->data());
        if (res != 0) {
            LOGE("ProcessReverseStream failed, ret %d \n",res);
        }
        memcpy(bufferOut, audio_reverse_, frame_size_10ms_);//far_frame_->data()
    }
}