README.md

---
library_name: transformers
tags:
- audio
- wavcoch
- tokenizer
- neural-codec
---

# WavCochCausalV64k-20ms-babyview

WavCoch causal **tokenizer** (8192-vocab FSQ) trained on the `bad` + `bvd` corpora. Quantizes
16 kHz audio to FSQ code indices and decodes codes back to a cochleagram. This is the tokenizer
only — it does **not** render waveforms. Its codebook is distinct from
`TuKoResearch/WavCochCausalV8192`, so codes are **not** interchangeable.

- Trained checkpoint: `model_best.pt` (step 199000)

```python
from transformers import AutoModel
import torch, torchaudio

m = AutoModel.from_pretrained("TuKoResearch/WavCochCausalV64k-20ms-babyview", trust_remote_code=True).eval()
wav, sr = torchaudio.load("clip.wav")          # resample to 16 kHz first if needed
codes = m.quantize(wav.unsqueeze(0))           # [1, N] FSQ indices
coch = m.decode(codes)                         # [1, T, out_channels] cochleagram
```