Preserving spoken content in voice anonymisation with character-level vocoder conditioning