The analysis of real-world conversational signal-to-noise ratios (SNRs) can provide insight into people's communicative strategies and difficulties and guide the development of hearing devices. However, measuring SNRs accurately is challenging in everyday recording conditions in which only a mixture of sound sources can be captured. This study introduces a method for accurate in situ SNR estimation where the speech signal of a target talker in natural conversation is captured by a cheek-mounted microphone, adjusted for free-field conditions and convolved with a measured impulse response to estimate its power at the receiving talker. A microphone near the receiver provides the noise-only component through voice activity detection. The method is applied to in situ recordings of conversations in two real-world sound scenarios. It is shown that the broadband speech level and SNR distributions are estimated more accurately by the proposed method compared to a typical single-channel method, especially in challenging, low-SNR environments. The application of the proposed two-channel method may render more realistic estimates of conversational SNRs and provide valuable input to hearing instrument processing strategies whose operating points are determined by accurate SNR estimates.