Grapheme-level pronunciation quality detection