ltuncay

/

Audio-JEPA

like 2

Hi Ludovic,

As agreed by email after your reply of July 15, 2026, here is the promised
contribution to round off the repo.

Three files added or replaced:

1. config.json : a machine-readable description of the architecture and
   the audio pipeline, with the confirmed training configuration
   (n_mels=128, target_time_bins=256, patch 16 x 16 giving a grid
   of 8 x 16 = 128 patches, sample rate 32000). It also includes an
   explicit note on the known speech gap and a redirection to BEST-RQ-2
   for speech-dominant use cases.

2. inference_example.py : a standalone CPU inference script (about 200
   lines) with a shim for flash_attn.modules.mha.MHA matching the
   qkv and proj naming of the checkpoint. This allows use without
   building the flash-attn CUDA kernels. Strict load succeeds with
   0 missing and 0 unexpected keys, tested on a 5 s waveform at
   RTF around 0.014 on a 20-core x86 CPU.

3. README.md : a proper HuggingFace page with a YAML frontmatter
   (tags, license, BibTeX citation), an honest "domain fit" section
   covering both strengths (generic audio) and limits (speech-only),
   and a minimal usage example. The current README being essentially
   empty, this PR replaces it in full.

Feel free to adjust or reject any part, this is your call for the team.
The README.md in particular deserves a careful re-read to make sure
it reflects your position.

Thanks again for your availability,
Michel-Marie