Pushing the performances of ASR models on English and Spanish accents