DiffSoundStream: Efficient Speech Tokenization via Diffusion Decoding