Speech Bandwidth normally denotes the frequency range a "nromal" speaker is using (80Hz to 12kHz) including all harmonics. For telephone based applications, normally this range is limited to 300 to 3400Hz. The sampling frequency is the lower frequency given by the Nyquist–Shannon sampling theorem. According to this theorem, the sampling frequency should be at least(!) two times higher than the highest informative signal frequency (in the upper example this frequency 3400Hz). Thus, the sampling frequency is 2*3400Hz (roughly 3*4000Hz=8kHz).

Be aware that Nyquist sampling theorem what really states is that you should use a sampling frequency at least twice higher that the highest frequency component in the signal. Otherwise you will get the so-called aliasing problem. To avoid this you should be sure that your signal has not frequency component higher that half of your samplign frequency or that you use an anti-aliasing filter before sampling.

the speech bandwidth is typical between ( 500 to 3800) Hz. the  frequency of sampling must be equal or larger than double speech bandwidth (i.e. Fs> Fm)