Update (2026-06-25): A Transformers-native version of Xcodec2 has been released here!

Update (2025-02-13): Add Llasa finetune instruction.

Update (2025-02-07): Our paper has been released!

Paper

LLaSA: Scaling Train Time and Inference Time Compute for LLaMA based Speech Synthesis

Codec Does Matter: Exploring the Semantic Shortcoming of Codec for Audio Language Model (AAAI 2025, xcodec 1.0)

Getting Started with XCodec2 on Hugging Face

XCodec2 is a speech tokenizer that offers the following key features:

1. Single Vector Quantization
2. 50 Tokens per Second
3. Multilingual Speech Semantic Support and High-Quality Speech Reconstruction

To use xcodec2, ensure you have it installed. You can install it using the following command:

conda create -n xcodec2 python=3.9
conda activate xcodec2
pip install xcodec2  (Use `xcodec2==0.1.5` for codec inference and llasa fine-tuning. I’ve removed unnecessary dependencies, and it works fine in my testing. However,  I’m not sure if other problems may arise. If you prefer more stability, I recommend using `xcodec2==0.1.3` which accurately aligns during my codec training.)

Then,

import torch
import soundfile as sf
from transformers import AutoConfig

from xcodec2.modeling_xcodec2 import XCodec2Model

model_path = "HKUSTAudio/xcodec2"

model = XCodec2Model.from_pretrained(model_path)
model.eval().cuda()

wav, sr = sf.read("test.wav")
wav_tensor = torch.from_numpy(wav).float().unsqueeze(0)  # Shape: (1, T)

with torch.no_grad():
   # Only 16khz speech
   # Only supports single input. For batch inference, please refer to the link below.
    vq_code = model.encode_code(input_waveform=wav_tensor)
    print("Code:", vq_code )

    recon_wav = model.decode_code(vq_code).cpu()       # Shape: (1, 1, T')

sf.write("reconstructed.wav", recon_wav[0, 0, :].numpy(), sr)
print("Done! Check reconstructed.wav")

If you want to train your own xcodec2, batch inference, or large-scale code extraction, the code is released here.