vad_startspeech equivalent for triggering end of utterance

[Carlo Benussi]

Hi,

I am using pocketsphinx, and while I am trying to tune the recognizer to achieve the best performances, I noticed the parameters:

-vad_postspeech     50      Num of silence frames to keep after from speech to silence.
-vad_prespeech      20      Num of speech frames to keep before silence to speech.
-vad_startspeech    10      Num of speech frames to trigger vad from silence to speech.

I found vad_startspeech a really useful parameter to tune, but I would like also (especially) to set the number of speech frames to trigger vad from speech to silence, and it seems this parameter is not present. Or am I missing something? Is it possible to set this parameter in some way?

Thanks in advance

[Nickolay V. Shmyrev]

There is no such parameter unfortuantely, you can check the corresponding code in fe_noise.c:fe_vad_hangover.

[Carlo Benussi]

From the function fe_vad_hangover in fe_noise.c it seems that the parameter to tune is vad_postspeech. Indeed in the function there are the two checks for silence->speech and speech->silence transitions:

    if (is_speech) {
        fe->vad_data->post_speech_frames = 0;
        if (!fe->vad_data->in_speech) {
            fe->vad_data->pre_speech_frames++;
            /* check for transition sil->speech */
            if (fe->vad_data->pre_speech_frames >= fe->start_speech) {
                fe->vad_data->pre_speech_frames = 0;
                fe->vad_data->in_speech = 1;
            }
        }
    } else {
        fe->vad_data->pre_speech_frames = 0;
        if (fe->vad_data->in_speech) {
            fe->vad_data->post_speech_frames++;
            /* check for transition speech->sil */
            if (fe->vad_data->post_speech_frames >= fe->post_speech) {
                fe->vad_data->post_speech_frames = 0;
                fe->vad_data->in_speech = 0;
                fe_prespch_reset_cep(fe->vad_data->prespch_buf);
            fe_prespch_reset_pcm(fe->vad_data->prespch_buf);
            }
        }
    }

So fe->post_speech is checked as a threshold for triggering the speech->silence transition.
Since from a quick check I believe this parameter is used only in this function, I kindly suggest to change the name and/or the description, for differentiating it from vad_prespeech and linking it to the similar vad_startspeech.

I suggest this change in the name + description:

Previous:

-vad_postspeech     50      Num of silence frames to keep after from speech to silence.

Correction:

-vad_stopspeech     50      Num of silence frames to trigger vad from speech to silence.

Thanks for the help. If I am suggesting something wrong, I apologize

[Nickolay V. Shmyrev]

Well, yes, the name could be better. I do not think think we need to change the option name though to avoid backward compatibility issues.