nikitastheo/speech-babylm-checkpoints

Speech BabyLM checkpoints

Mirrors of third-party speech model checkpoints used for zero-shot evaluation (ABX, BabySLM) in the Speech BabyLM project, so they can be fetched with huggingface_hub instead of from their original, less reliable hosts. None of these models were trained by us; all are MIT-licensed, and each folder keeps its project's license.

File	Model	Original source	Original SHA-256
cpc/cpc_checkpoint_106.safetensors	CPC (CPC_audio)	cpc_checkpoint_106.pt in the ZeroSpeech internal_tools tutorial models.zip (Google Drive id 129Fkg_bQpVB_yN-YT5MVK6ulZS_zjhH6)	42c80034…d57d3db
melhubert/melhubert_960_stage2_20ms.safetensors	MelHuBERT, 20 ms frames (MelHuBERT)	melhubert_960_stage2_20ms.ckpt, same models.zip	0c100c34…f76f65aa
slam/mhubert_base_25hz_cp_mls_cv_sp_fisher_L11_km500.bin	k-means (500 units) on layer 11 of mHuBERT-base-25Hz, the speech tokenizer of TWIST and Slam (textlesslib)	https://dl.fbaipublicfiles.com/textless_nlp/twist/speech_tokenizer/mhubert_base_25hz_cp_mls_cv_sp_fisher_L11_km500.bin	03cc04a9…7cf8d28ad (unchanged)

Changes from the originals

• CPC, MelHuBERT: only the model weights are kept, as safetensors: the original checkpoint's gEncoder (CPC) or model (MelHuBERT) state dict, with every tensor unchanged. Optimizer state and training arguments are dropped.
• k-means: the original file, unchanged. It is a scikit-learn MiniBatchKMeans saved with joblib (a pickle), as slamkit loads it.

Usage

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

state_dict = load_file(hf_hub_download("nikitastheo/speech-babylm-checkpoints", "cpc/cpc_checkpoint_106.safetensors"))