The fine tuning mechanisms involved in the normal processing of sound in the cochlea are non-linear, hence combination tones are generated inside the cochlea when a pair of low-level pure tones with neighbouring frequencies f1 and f2 is used as a stimulus. Their detection as sounds in the ear canal proves that they undergo backward propagation in the cochlea and through the middle ear, and the non-invasive measurement of the combination tone at 2f1-f2, called the cubic difference tone (CDT), has become a routine method of monitoring cochlear function. In order to gain information on the hypothetical places where CDTs are generated, on their intracochlear levels and propagation velocities, direct measurements of CDT pressure waves were carried out in scala vestibuli and tympani of the first and second turn of the guinea-pig cochlea. Cubic difference tones at 2f1-f2 varied from 0.75 to 9 kHz and were measured with a miniature piezoresistive transducer. Its high sensitivity allowed the detection of CDTs whenever their levels exceeded 5 dB SPL in the ear canal, i.e. 40 dB SPL (re: 20 microPa) inside the cochlea. The levels of CDTs were similar in scala vestibuli of the first and second turn. Phase comparisons between measurements at 2f1-f2 in the first and second turn allowed determination of the place where the CDT phase was minimum. It provided an estimation of the generation site of the CDT, which appeared to be close to the place tuned to f2 for stimulus levels lower than 70 dB SPL. Forward and backward travel times from one turn to the other were assessed at several frequencies, and both values were shorter than 0.2 ms. In contrast, the overall 'round-trip' delay of CDTs, measured in the ear canal, was about five times larger, suggesting that local filtering processes rather than propagation delays account for the overall CDT delay.