5.3.2 Fine-Tuning VAD Performance on DM3 Boards

The VAD algorithm on DM3 boards performs sophisticated calculations on each input signal to determine whether the signal is speech or not. The probability of speech is computed based on the current energy level estimate and zero-crossing frequency. The calculations include a combination of long-term and short-term energy and zero-crossing based probabilities.

Thus, the adaptive nature of the VAD on DM3 boards reduces the need to fine-tune VAD parameters.

When using the SVAD mode on DM3 boards (ECCH_SVAD = 0), which is a combination of energy and zero-crossing based probability calculations, you can adjust the DXCH_SPEECHPLAYWINDOW and DXCH_SPEECHPLAYTRIGG parameters to fine-tune the performance of VAD. If you find that your application triggers on echo or background noise, you may want to:

• increase speech trigger (DXCH_SPEECHPLAYTRIGG) so that the incoming speech energy is present for a greater duration of the speech window.
• increase speech window (DXCH_SPEECHPLAYWINDOW) thereby requiring the incoming speech energy be present for a longer time period. This should protect against false triggers due to noise spikes or other short duration (non-speech) noises.

Note:

To reduce sensitivity to background noise, you must perform these actions in certain combinations only: increase speech trigger alone, or increase speech trigger and speech window together to reduce sensitivity to background noise. Increasing speech window alone will not help reduce sensitivity to background noise.

When using the energy-only mode on DM3 boards (ECCH_SVAD = 1), you can adjust the DXCH_SPEECHPLAYTRIGG, DXCH_SPEECHPLAYWINDOW and DXCH_SPEECHPLAYTHRESH parameters to fine-tune the performance of VAD.

On DM3 boards, non-play parameters, DXCH_SPEECHNONPLAYTRIGG, DXCH_SPEECHNONPLAYWINDOW and DXCH_SPEECHNONPLAYTHRESH, are not supported.

For more information on all VAD parameters, see the ec_setparm( ) function description in the Continuous Speech Processing API Library Reference.

Click here to contact Dialogic Customer Engineering