Clarification on which PGSLM model is used in experiments

[dlion168]

Hi, thank you for releasing this great work!

I have a question regarding the PGSLM model used in your experiments.

In the default configuration of the repo, the PGSLM tokenizer section specifies:

"tokenizer": {
    "dense_model_name": "mhubert-base-25hz",
    "quantizer_model_name": "kmeans",
    "encoder_vocab_size": 500,
    "deduplicate": true,
    "need_f0": true,
    "f0_func": "parselmouth",
    "log_f0": true,
    "mean_f0": false,
    "scale_f0": false,
    "f0_bins_path": "/cs/labs/oabend/avishai.elma/pgslm_data_my_bins/f0_bins.pt"
}

However, the official PGSLM release from [fairseq’s textless_nlp examples](https://github.com/facebookresearch/fairseq/tree/main/examples/textless_nlp/pgslm) only includes a model with 100 units and uses HuBERT-base-ls960 as the dense model.

Could you please clarify:

• Which PGSLM model was actually used in your experiments?
• If it’s a custom model, could you share more details or a pointer to where it can be obtained?

Thank you very much for your time and for sharing your work!