Improving grapheme-to-phoneme conversion by learning pronunciations from speech recordings