This paper presents the Androids Corpus, a new benchmark for speech-based automatic depression detection. The corpus is a collection of 228 recordings uttered by 118 native Italian speakers, including 64 who were diagnosed with depression by professional psychiatrists. Out of the 228 recordings, 112 contain read speech (all speakers read the same text) and 116 contain spontaneous speech (all speakers answer the same questions posed by an interviewer). For 110 speakers, including 58 diagnosed with depression, the corpus includes both read and spontaneous speech samples. Overall, the total duration of the material is 1 hour, 33 minutes and 49 seconds for read speech and 7 hours, 24 minutes and 22 seconds for spontaneous speech. Besides the data, the corpus includes experimental protocols that can be replicated, thus ensuring reproducibility of the experiments performed over it and comparison of the results.