理解RTF和RTX指标_asr rtf

RTF(real_time_factor实时率,在线识别指标)

RTF(real time factor)表示如下：
$$RTF=\frac{ASR识别所有音频处理时间(单位秒)}{所有音频总时间(单位秒)}$$
【意义】：平均1秒时长音频给ASR算法需要多少秒处理，小于1才能达到实时效果。越小越好，正常0.2-0.3
          该指标针对的是在线ASR识别指标
具体见官方代码(参考文献)：

 BaseFloat frame_shift = 0.01;
 BaseFloat real_time_factor =
     global_timer.Elapsed() / (frame_shift * num_frames);
 if (num_frames > 0)
   KALDI_LOG << "Real-time factor was " << real_time_factor
             << " assuming frame shift of " << frame_shift;
1
2
3
4
5
6

另外一段代码(参考文献)：

double real_time_factor = total_time_taken_ / total_audio_,
    average_wait = (total_time_taken_ - total_audio_) / num_utts_,
    idle_proportion = total_time_waited_ / total_audio_,
    idle_percent = 100.0 * idle_proportion;
1
2
3
4

    double elapsed = timer.Elapsed();
    KALDI_LOG << "Time taken "<< elapsed
              << "s: real-time factor assuming 100 frames/sec is "
              << (elapsed*100.0/frame_count);
1
2
3
4

【注解】 会发现上面三段代码都是一个意思
r e a l _ t i m e _ f a c t o r = t o t a l _ t i m e _ t a k e n t o t a l _ a u d i o = t o t a l _ t i m e _ t a k e n f r a m e _ s h i f t ⋅ n u m _ f r a m e s = t o t a l _ t i m e _ t a k e n 0.01 ⋅ n u m _ f r a m e s = 100 ⋅ t o t a l _ t i m e _ t a k e n n u m _ f r a m e s

$$\begin{aligned} real\_time\_factor & = \frac{total\_time\_taken}{total\_audio} \\ & = \frac{total\_time\_taken}{frame\_shift \cdot num\_frames} \\ & = \frac{total\_time\_taken}{0.01 \cdot num\_frames} \\ & = \frac{100 \cdot total\_time\_taken}{num\_frames} \\ \end{aligned}$$ real_time_factor​=total_audiototal_time_taken​=frame_shift⋅num_framestotal_time_taken​=0.01⋅num_framestotal_time_taken​=num_frames100⋅total_time_taken​​
即：$音频总时长=total\_audio=frame\_shift\cdot num\_frames=0.01\cdot num\_frames$

---

RTX(real_time_x实时效率,离线转写指标)

RTX(real time x)表示如下：
$$RTX=\frac{所有音频总时间(单位秒)}{ASR识别所有音频处理时间(单位秒)}$$
【意义】：平均1秒时间内我们的ASR算法可以处理多长时间的音频，越大越好，正常在GPU上是500多
参考代码：

KALDI_LOG << "Overall: "
              << " Aggregate Total Time: " << total_time
              << " Total Audio: " << total_audio * iterations
              << " RealTimeX: " << total_audio * iterations / total_time;
1
2
3
4

Reference

online-timing.cc
online2-wav-nnet2-latgen-threaded.cc
gmm-latgen-simple.cc
batched-wav-nnet3-cuda.cc