Recommended input pre-processing steps to achieve maximum VAD accuracy?

[tojo423]

As the title states.

[snakers4]

It depends on whether your have a streaming or a non-streaming scenario.

For both - the best approach is to make sure that the highest possible quality audio with high dynamic range is used.

Typically most problems arise from:

• Very poor audio quality and / or noise louder than voice
• Very quiet audio or very quiet voice (for example audio is coded as int8 or int16 with very low values, but after renormalization sounds ok. it is hard to renormalize during streaming)
• Some loud atypical noise