Zaslat záznam e-mailem: Attention-based latent features for jointly trained end-to-end automatic speech recognition with modified speech enhancement