This dataset resumes an urban sound corpus whose the realism has been proved through a perceptual test [1]. This corpus has been used in order to estimate the traffic sound level with the Non-negative Matrix Factorization formula [2]. This dataset presents 4 folders : dictionary where the traffic audio samples dedicated to the dictionary design of NMF are, recordings which contains the 74 original recordings, annotation which contains the annotations text files of the 74 audio files, transcribed scenes which contains the 74 transcribed audio files generated with SimScene software. 4 folders composed it, according to the sound environment of the audio files (park, quiet street, noisy street, very noisy street). In each folder, one can find the global sound mixtures, the audio of each sound class as well as the files that include all the elements associated with the traffic and the interfering class (which contains all the other sound sources). [1] Gloaguen, J. R., Can, A., Lagrange, M., & Petiot, J. F. (2017, June). Creation of a corpus of realistic urban sound scenes with controlled acoustic properties. In 173rd Meeting of the Acoustical Society of America and the 8th Forum Acusticum (Acoustics' 17). [2] Gloaguen, J. R., Can, A., Lagrange, M., & Petiot, J. F. (2018), Road traffic sound level from realistic urban sound mixtures by Non-negative Matrix Factorization, submitted for publication