Potential reference of a promising transformer-only streaming single-codec model

[hbwu-ntu]

TS3-Codec: Transformer-Based Simple Streaming Single Codec:
https://arxiv.org/abs/2411.18803

[jishengpeng]

TS3-Codec: Transformer-Based Simple Streaming Single Codec: https://arxiv.org/abs/2411.18803

Thank you very much for your attention. We took note of TS3-Codec as soon as it was uploaded to arXiv. It is an excellent piece of work. The models listed in WavChat are currently updated only up to early November. If WavChat is eventually accepted, we will include TS3-Codec in the final version of the paper. Additionally, we will consider comparing TS3-Codec with our method in future codec-related work.

Furthermore, we noticed that you provided results from Table 3. As the first author of WavTokenizer, I would like to share some personal observations. It appears that Table 3 does not specify the audio sampling rates used for each codec, particularly for 16,000 Hz and 24,000 Hz sampling rates, which are critical for determining the actual compression ratios. From our experiments, we found that this has a significant impact on the number of tokens per second. Specifically, the 50-token metric of WavTokenizer at a 16,000 Hz sampling rate is equivalent to the 75-token metric at a 24,000 Hz sampling rate.

[hbwu-ntu]

Thanks for taking note on the TS3-Codec. I will also give some discussions about the sampling rate in the TS3-Codec paper.