Dear Seiedmehdi Pourseiedrezaei,

for natural sounds, as a first approach you could try the MP3 time window ~ 13ms.

Best regards  

you can test for stationarity by using time windows of different lengths (starting with the smallest) and extracting statistics for each epoch and comparing with other epochs.

then, u can pick the time window which gives the best result for your stationarity test.

of course, smaller the time window, higher the probability of the signal being stationary there, but you do not want to choose a time window which is too small, hence the above sweeping technique will help.