Speech Watermarking with Discrete Intermediate Representations