The a.m. noise spectrum in reflex klystrons is treated experimentally and theoretically. The frequency range of the measurements is 1–100 kc/s from the signal frequencies, which are 4.7 and 9.3 Gc/s. It is shown that the measurement equipment cannot consist of a crystal detector immediately followed by a low-frequency amplifier because the crystal noise up to about 20 kc/s is greater than the klystron noise. To reduce the effect of crystal noise the klystron is frequency modulated and the measurement equipment consists of a crystal diode as first detector, an i.f. amplifier, a vacuum diode as second detector and a selective l.f. amplifier. The measurements show that the noise density is, with great accuracy, constant in the actual range. This agrees with the theory, which is based on a treatment by Knipp. The deviation from Knipp's noise formula is due to the introduction of an inherent synchronizing effect in the noise current. The theory indicates that the noise is approximately constant up to a displacement of 10−4to 10−3from the signal frequency.