jepa-q2d2

jepa-q2d2

a 1.6 kbps neural speech codec. a jepa (joint-embedding predictive architecture) encoder feeds q2d2, a geometry-aware 2d rhombic-lattice quantizer, and a hifi-gan-style decoder. the main codec is trained without any adversarial loss (a frozen wavlm perceptual loss is used while training the decoder).

• code: https://github.com/anant-004/jepa-q2d2
• paper: extended abstract in the github repo (paper/sltc2026/main.pdf)

models

each folder has pytorch_model.pt, model.safetensors, and config.json.

quality is on a fixed 50-utterance librilight set. estoi is extended stoi, comparable only within this paper's pipeline. the main cd64 model is ahead of encodec at a comparable rate and ahead of mimi on pesq, behind mimi on estoi. the sigreg model is the co-design point: in our runs, the codec at the aggressive 25 hz / cd32 setting collapsed unless the encoder distribution was gaussianized with sigreg. unlike the main model, it was trained with a gan discriminator. the teacher is the higher-rate fsq codec (128 dims with 8 levels each, so 4.8 kbps). a separate cd64 run was distilled from it; the released cd64 model was trained without the distillation loss.

emergent properties

the codec is trained on english read speech only and never sees language or emotion labels. the encoder features still pick up structure it was not supervised on.

cross-lingual: a 5-nn probe on frozen pre-quantization features separates 6 fleurs languages at 0.85 (chance 0.167), ahead of mimi 0.63, encodec 0.61, dac 0.45 and the multilingual ssl models we tried (xls-r 0.75, mhubert-147 0.61). only uncompressed english wav2vec 2.0 base is higher (0.93). cluster nmi against language is higher too (0.26 vs encodec 0.10, dac 0.07). the utterance-disjoint number is the same (0.85).

style and emotion: utterance embeddings separate by speaking style on a hindi 5-class set (angry, excited, neutral, sad, whisper). whisper splits off cleanly and the rest form distinct regions for the jepa encoders, while encodec and dac mix them.

usage

import torch, soundfile as sf
from koe.fast.hf_codec import load_codec_from_hf   # from the github repo

model, info = load_codec_from_hf("main", device="cpu")   # downloads weights from here
print(info)

wav, sr = sf.read("in.wav")                              # mono, resampled to 24 khz internally
x = torch.from_numpy(wav).float().view(1, 1, -1)
z_q = model.encode(x)[0]                                 # discrete q2d2 tokens
recon = model.decode(z_q)                                # 24 khz waveform

model code lives in the github repo. the released checkpoints already include the fine-tuned encoder, so inference needs only these files plus the repo code. the files also hold the stage 1 ema target encoder and predictor (about 121m parameters), which inference does not use.

training data

librilight, english read speech, 24 khz. the cross-lingual structure above is evaluated zero-shot on fleurs.

citation

@misc{shukla2026jepaq2d2,
  title        = {JEPA-Q2D2: A Low-Bitrate Speech Codec with Emergent Cross-Lingual Structure},
  author       = {Shukla, Anant and Anand, Aman and Shakya, Suryansh and Bharti, Vatsal},
  year         = {2026},
  howpublished = {\url{https://github.com/anant-004/jepa-q2d2}},
}

license

cc-by-4.0, released for research use.