TurnGuide

June 15, 2026 · View on GitHub

arXiv Hugging Face 2:1 Hugging Face 3:1

News

  • 🎉🎉🎉 TurnGuide has been accepted to the Interspeech 2026 Long Paper Track!

Github repository for paper: TurnGuide: Enhancing Meaningful Full Duplex Spoken Interactions via Dynamic Turn-Level Text-Speech Interleaving

Explore the demo!

This work introduces an end-to-end full-duplex speech language model and strengthens its capabilities to generate coherent, meaningful full duplex dialogues. As an initial release, we provide the test splits from the Fisher and Candor datasets to support fair and straightforward benchmarking for future research.

Inference code

This repository includes the TurnGuide inference demo together with the GLM-4-Voice code modules needed to run it:

  • turnguide_inference.py: the current TurnGuide inference script.
  • turnguide_inference_reproducible.py: a legacy/reproducible inference variant.
  • flow_inference.py, speech_tokenizer/, cosyvoice/, and third_party/Matcha-TTS/: supporting code from GLM-4-Voice and its decoder stack.

The model weights are not included in this repository. Download the GLM-4-Voice decoder separately and pass model paths explicitly when needed.

Model assets:

The two TurnGuide checkpoints differ only in the training loss ratio between text tokens and speech tokens:

  • qqjz/turnguide_loss_2_1: text:speech token loss ratio = 2:1
  • qqjz/turnguide_loss_3_1: text:speech token loss ratio = 3:1

Both checkpoints can be used with the same inference script by changing --model-path.

Environment

The recommended environment is exported from the tested torch250cu121 setup used for TurnGuide inference. It uses Python 3.10, PyTorch 2.5.0, torchaudio 2.5.0, and CUDA 12.1.

conda env create -f environment.yml
conda activate turnguide

The environment pins mkl=2021.4.0 because newer MKL builds can break PyTorch 2.5.0 imports with an iJIT_NotifyEvent symbol error. It also includes tiktoken for the TurnGuide tokenizer remote code and ruamel.yaml==0.18.6 for GLM-4-Voice decoder config loading through HyperPyYAML.

To check the core runtime versions:

python -c "import torch, torchaudio, transformers; print(torch.__version__); print(torch.version.cuda); print(torchaudio.__version__); print(transformers.__version__); print(torch.cuda.is_available())"

Expected core versions:

torch 2.5.0
CUDA 12.1
torchaudio 2.5.0
transformers 4.44.1

If you already have a CUDA 12.1 PyTorch environment, you can install only the pip-side dependencies with:

pip install -r requirements.txt

Run inference

git clone https://huggingface.co/zai-org/glm-4-voice-decoder

python turnguide_inference.py \
  --input-audio examples/audio/fe_03_05844_first_2min_left.wav \
  --model-path qqjz/turnguide_loss_2_1 \
  --tokenizer-path zai-org/glm-4-voice-tokenizer \
  --flow-path ./glm-4-voice-decoder \
  --output-dir ./turnguide_demo_output

To use the 3:1 checkpoint, replace --model-path qqjz/turnguide_loss_2_1 with --model-path qqjz/turnguide_loss_3_1. You can also pass a local checkpoint directory.

This repository includes examples/audio/fe_03_05844_first_2min_left.wav as a small example input for the inference command above. You can replace it with any mono user-audio WAV file.

Data

Examples:

  • Fisher: fe_03_11632_60_180.wav
    • ID: fe_03_11632
    • Time segment: 60-180s
  • Candor: 46f8e9b8-f80a-48cf-90a0-2e29908202c0_420.0_540.0.wav
    • ID: 46f8e9b8-f80a-48cf-90a0-2e29908202c0
    • Time segment: 420-540s

Download Fisher and Candor datasets.

Acknowledgements and license

This code release builds on GLM-4-Voice. The included GLM-4-Voice code is licensed under Apache-2.0. GLM-4-Voice model weights are governed by their own model license and must be downloaded separately.

Citation

@article{turnguide2026,
  title={TurnGuide: Enhancing Meaningful Full Duplex Spoken Interactions via Dynamic Turn-Level Text-Speech Interleaving},
  author={Cui, Wenqian and Zhu, Lei and Li, Xiao-Hui and Guo, Zhihan and Bai, Haoli and Hou, Lu and King, Irwin},
  journal={arXiv preprint arXiv:2508.07375},
  year={2026}
}