This paper presents XLSR which learns cross-lingual speech representations by\npretraining a single model from the raw waveform of speech in multiple\nlanguages. We build on wav2vec 2.0 which is trained by solving a contrastive\ntask over masked latent speech representations and jointly learns a\nquantization of the latents shared across languages. The resulting model is\nfine-tuned on labeled data and experiments show that cross-lingual pretraining\nsignificantly outperforms monolingual pretraining. On the CommonVoice\nbenchmark, XLSR shows a relative phoneme error rate reduction of 72% compared\nto the best known results. On BABEL, our approach improves word error rate by\n16% relative compared to a comparable system. Our approach enables a single\nmultilingual speech recognition model which is competitive to strong individual\nmodels. Analysis shows that the latent discrete speech representations are\nshared across languages with increased sharing for related languages. We hope\nto catalyze research in low-resource speech understanding by releasing XLSR-53,\na large model pretrained in 53 languages.\n