unimrcp-voice-activity语音检测

研究 unimrcp有一段时间了，其中unimrcp voice acitve的算法，是遭到大家频繁吐槽。今天我们简单的介绍一下unimrcp voice activity 的这个简单粗暴的算法：

 

unimrcp 语音活动检测是通过能量来控制的，设定几个常量：

struct mpf_activity_detector_t {
/* 静音检测阀值 */
apr_size_t level_threshold;

/* 转换成active状态的超时时长*/
apr_size_t speech_timeout;
/* 转换成inactive状态超时时长 */
apr_size_t silence_timeout;
/* 没有输入的超时时长 */
apr_size_t noinput_timeout;

/* 当前的状态 */
mpf_detector_state_e state;
/* 当前状态的保持的时长 */
apr_size_t duration;
};

 

来看一下这几个参数的初始化的值，根据实际的测试，我们后期做过改动：

/** Create activity detector */
MPF_DECLARE(mpf_activity_detector_t*) mpf_activity_detector_create(apr_pool_t *pool)
{
    mpf_activity_detector_t *detector = apr_palloc(pool,sizeof(mpf_activity_detector_t));
    detector->level_threshold = 50; /* 0 .. 255 */
    detector->speech_timeout = 300; /* 0.3 s  = 300*/
    detector->silence_timeout = 1000; /* 0.3 s  =300 */
    detector->noinput_timeout = 5000; /* 5 s =5000*/
    detector->duration = 0;
    detector->state = DETECTOR_STATE_INACTIVITY;
    return detector;
}

 

看一下重要的函数，能量的计算：

根据frame的信息，对能量进行粗暴的累加，所以对于噪音，这个算法完全不可用。后面将会介绍如何采用webrtc的voice active来取代这个算法。

static apr_size_t mpf_activity_detector_level_calculate(const mpf_frame_t *frame)
{
    apr_size_t sum = 0;
    //计算多少个
    apr_size_t count = frame->codec_frame.size/2;
    //初始值
    const apr_int16_t *cur = frame->codec_frame.buffer;
    //最后一个值
    const apr_int16_t *end = cur + count;

    for(; cur < end; cur++) {
        if(*cur < 0) {
            sum -= *cur;
        }
        else {
            sum += *cur;
        }
    }
    //取平均值，简单粗暴，被吐槽的原因
    return sum / count;
}

 

最后看一下，状态切换的过程，下面mpf_activity_detector_process函数，主要是通过计算frame的平均值，来完成状态切换的逻辑：

处理过程共有四个状态：

      ACTIVITY状态

      INACTIVITY状态

     TRANS_ACTIVITY状态

     TRANS_INACTIVITY状态

     其中TRANS状态是中间状态，再切换为ACTIVITY状态和INACTIVITY状态的时，需要经过这个状态来累计设定时长，如果满足了，才会切换，否则不予切换。

/** Process current frame */
MPF_DECLARE(mpf_detector_event_e) mpf_activity_detector_process(mpf_activity_detector_t *detector, const mpf_frame_t *frame)
{
    mpf_detector_event_e det_event = MPF_DETECTOR_EVENT_NONE;
    apr_size_t level = 0;
    if((frame->type & MEDIA_FRAME_TYPE_AUDIO) == MEDIA_FRAME_TYPE_AUDIO) {
        /* first, calculate current activity level of processed frame */
        //此处计算得到level的值

        level = mpf_activity_detector_level_calculate(frame);
#if 0
        apt_log(APT_LOG_MARK,APT_PRIO_INFO,"Activity Detector --------------------- [%"APR_SIZE_T_FMT"]",level);
#endif
    }
    /*如果当前状态处于INACTIVITY状态，并且level 大于我们设定的阀值，开始向活动状态切换，但是并没有变成活动状态*/
    if(detector->state == DETECTOR_STATE_INACTIVITY) {
        if(level >= detector->level_threshold) {
            /* start to detect activity */
            mpf_activity_detector_state_change(detector,DETECTOR_STATE_ACTIVITY_TRANSITION);
        }
        else {
            detector->duration += CODEC_FRAME_TIME_BASE;
            if(detector->duration >= detector->noinput_timeout) {
                /* detected noinput */
                det_event = MPF_DETECTOR_EVENT_NOINPUT;
            }
        }
    }
    else if(detector->state == DETECTOR_STATE_ACTIVITY_TRANSITION) {
        //处于向活动状态转换的过程。
        if(level >= detector->level_threshold) {
            //如果level 大于阀值
            detector->duration += CODEC_FRAME_TIME_BASE;
            //并且超过了设定了向活动状态转换的超时时长
            if(detector->duration >= detector->speech_timeout) {
                /* finally detected activity */
                //切换为活动状态

                det_event = MPF_DETECTOR_EVENT_ACTIVITY;
                mpf_activity_detector_state_change(detector,DETECTOR_STATE_ACTIVITY);
            }
        }
        else {
            /* fallback to inactivity */
            //降级为非活动状态

            mpf_activity_detector_state_change(detector,DETECTOR_STATE_INACTIVITY);
        }
    }
    else if(detector->state == DETECTOR_STATE_ACTIVITY) {
        //处于活动状态
        if(level >= detector->level_threshold) {
             //如果level大于阀值，增加duration
            detector->duration += CODEC_FRAME_TIME_BASE;
        }
        else {
            /* start to detect inactivity */
            //准备转换成inactivity状态

            mpf_activity_detector_state_change(detector,DETECTOR_STATE_INACTIVITY_TRANSITION);
        }
    }
    else if(detector->state == DETECTOR_STATE_INACTIVITY_TRANSITION) {
        //处于inactivity transtion状态
        if(level >= detector->level_threshold) {
            /* fallback to activity */
            //如果大于阀值了，则回归到activity状态

            mpf_activity_detector_state_change(detector,DETECTOR_STATE_ACTIVITY);
        }
        else {
            //如果检测仍然小于阀值，增加判断时长，如果大于设定的时长了，则进入inactivity状态。
            detector->duration += CODEC_FRAME_TIME_BASE;
            if(detector->duration >= detector->silence_timeout) {
                /* detected inactivity */
                det_event = MPF_DETECTOR_EVENT_INACTIVITY;
                mpf_activity_detector_state_change(detector,DETECTOR_STATE_INACTIVITY);
            }
        }
    }

    return det_event;
}