Mojodex can learn only 245 tokens as "vocabulary" for transcription. We should find a way to lift the restriction.