Reproducibility of X-codec1 metrics in Table 1 #32
Description
Hello, I am trying to reproduce the objective metrics listed in Table 1 of the Llasa paper. I was able to achieve similar results for models like X-Codec2 and MimiCodec in terms of WER, SPK-SIM, PESQ-WB, and PESQ-NB (though the UT-MOS scores differ slightly). However, for X-Codec1 (nq1 and nq2), my results are significantly different.
For X-Codec1, I used the official implementation from GitHub, specifically the xcodec_hubert_librispeech model, and reconstructed the audio using the provided inference.py script. Could you clarify if there were any additional settings or adjustments used in your experiments?
