🎙️ Behavior-SD

Official repository for our NAACL 2025 paper:
Behavior-SD: Behaviorally Aware Spoken Dialogue Generation with Large Language Models
Sehun Lee*, Kang-wook Kim*, Gunhee Kim (* Equal contribution)

🏆 SAC Award Winner in Speech Processing and Spoken Language Understanding

🔗 Links

• Project Page
• Code

📖 Overview

We explores how to generate natural, behaviorally-rich full-duplex spoken dialogues using large language models (LLMs). We introduce:

• Behavior-SD Dataset: 108K full-duplex dialogues (2,164 hours) with rich speaker-wise behavioral annotations.
• BeDLM: A novel end-to-end LLM-based spoken dialogue generator conditioned on narrative and behavioral traits.

📂 Datasets Info

Split	Num Audios	Storage
train	106,317	176 GB
validation	932	1.6 GB
test	925	1.5 GB

Download

from datasets import load_dataset

# Load the Behavior-SD dataset using streaming mode (recommended for large datasets)
dataset = load_dataset(
    "yhytoto12/behavior-sd",
    split="train", # "validation" or "test"
    streaming=True
)

# Example: Iterate over the dataset
for i, example in enumerate(dataset):
    print(example)
    break

Or you can manually download the dataset repository with Git LFS:

# Make sure you have git-lfs installed
# Installation (if not already installed)
# For Ubuntu: sudo apt-get install git-lfs
# For MacOS: brew install git-lfs
# Then initialize:
git lfs install

# Clone the dataset repository
git clone https://huggingface.co/datasets/yhytoto12/behavior-sd

# The dataset files (audio and metadata) will be downloaded into the behavior-sd/ directory
cd behavior-sd

You can also download only specific chunks.

# Clone only metadata pointers without downloading large files
GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/datasets/yhytoto12/behavior-sd

# Move into the dataset directory
cd behavior-sd

# When you want to download a specific file (e.g., "train/0000.tar"), run:
git lfs pull --include="train/0000.tar"
# Or download a specific split
git lfs pull --include="validation/*"

Description

# "0000000004.json"
{
  "file_name": "0000000004.flac",
  "soda_split": "train",
  "soda_index": 4,
  "narrative": "Cornell knows what Dontrell is thinking...",
  "speakers": ["Cornell", "Dontrell"],
  "behaviors": [
    {"utterance_length": 0, "filler_words": 0, "backchannels": 0, "interruptions": 2},
    {"utterance_length": 0, "filler_words": 2, "backchannels": 0, "interruptions": 0}
  ],
  "num_turns": 10,
  "utterances": [
    {
      "uttr_idx": 0,
      "uttr_type": null,
      "speaker_idx": 1,
      "speaker": "Dontrell",
      "tts_text": "So, I was thinking... um... we should probably plan...",
      "dur_samples": 60672,
      "start_time": 0.0,
      "end_time": 2.75156462585034,
      "backchannels": []
    },
    ...
  ],
  "tts_speaker_ids": ["0001024622_0", "0000805189_1"],
  "tts_genders": ["female", "male"],
  "statistics": {
    "num_utterances": [5, 5],
    "num_turntaking": [5, 4],
    "durations": [5.53, 25.35],
    "num_interruptions": [2, 0],
    "num_backchannels": [0, 0],
    "num_filler_words": [0, 8]
  }
}

📌 Citation

If you find our work useful, please consider citing us:

@inproceedings{lee-and-kim@behaviorsd,
  title     = {Behavior-SD: Behaviorally Aware Spoken Dialogue Generation with Large Language Models},
  author    = {Sehun Lee, Kang-wook Kim, Gunhee Kim},
  booktitle = {Proceedings of the 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics},
  year      = {2025},
  url       = {https://aclanthology.org/2025.naacl-long.484/}
}