FDlalala/test

---
license: cc-by-nc-4.0
task_categories:
- audio-classification
- automatic-speech-recognition
language:
- en
tags:
- conversational
- emotions
- dialogues
- conversations
pretty_name: Deep Dialogue (XTTS-v2)
size_categories:
- 100K<n<1M
dataset_info:
- config_name: all
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 208162364
num_examples: 39055
download_size: 0
dataset_size: 208162364
- config_name: default
features:
- name: conversation_id
dtype: string
- name: model_dir
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: turn_index
dtype: int64
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio_path
dtype: string
- name: segment_audio_path
dtype: string
- name: audio_segment_id
dtype: float64
- name: audio_model
dtype: string
- name: audio_actor
dtype: string
- name: audio_original_text
dtype: string
- name: audio_substituted_text
dtype: string
- name: audio_cleaned_text
dtype: string
- name: audio_dialogue_emotion
dtype: string
- name: audio_ravdess_emotion
dtype: string
- name: audio_ref_audio
dtype: string
- name: audio_ref_transcript
dtype: string
- name: audio_start_time
dtype: float64
- name: audio_end_time
dtype: float64
- name: audio_duration
dtype: float64
- name: audio_following_silence
dtype: float64
- name: audio_generation_time
dtype: float64
- name: audio_realtime_factor
dtype: float64
splits:
- name: train
num_bytes: 261914837
num_examples: 243295
download_size: 80172060
dataset_size: 261914837
- config_name: dialogues_cohere7B
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 2197335
num_examples: 673
download_size: 0
dataset_size: 2197335
- config_name: dialogues_cohere7B_gemma3-4B
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 4659388
num_examples: 1144
download_size: 0
dataset_size: 4659388
- config_name: dialogues_gemma3-27B
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 17243981
num_examples: 3218
download_size: 0
dataset_size: 17243981
- config_name: dialogues_gemma3-4B
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 5194294
num_examples: 1206
download_size: 0
dataset_size: 5194294
- config_name: dialogues_llama3-70B
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 20470475
num_examples: 3636
download_size: 0
dataset_size: 20470475
- config_name: dialogues_llama3-70B_qwen2.5-72B
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 21600583
num_examples: 3791
download_size: 0
dataset_size: 21600583
- config_name: dialogues_llama3-8B
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 7223668
num_examples: 1582
download_size: 0
dataset_size: 7223668
- config_name: dialogues_llama3-8B_cohere7B
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 7466351
num_examples: 1644
download_size: 0
dataset_size: 7466351
- config_name: dialogues_llama3-8B_gemma3-4B
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 9629300
num_examples: 1986
download_size: 0
dataset_size: 9629300
- config_name: dialogues_phi4-14B
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 7642499
num_examples: 1622
download_size: 0
dataset_size: 7642499
- config_name: dialogues_phi4-14B_gemma3-27B
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 19799165
num_examples: 3455
download_size: 0
dataset_size: 19799165
- config_name: dialogues_phi4-mini
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 3764587
num_examples: 1014
download_size: 0
dataset_size: 3764587
- config_name: dialogues_qwen2.5-32B
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 18944118
num_examples: 3294
download_size: 0
dataset_size: 18944118
- config_name: dialogues_qwen2.5-32B_gemma3-27B
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 18619665
num_examples: 3493
download_size: 0
dataset_size: 18619665
- config_name: dialogues_qwen2.5-32B_phi4-14B
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 21506478
num_examples: 3529
download_size: 0
dataset_size: 21506478
- config_name: dialogues_qwen2.5-72B
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: timestamp
dtype: string
- name: model1
dtype: string
- name: model2
dtype: string
- name: configuration
dtype: string
- name: conversation
sequence:
- name: speaker
dtype: string
- name: text
dtype: string
- name: emotion
dtype: string
- name: full_audio
dtype: audio
- name: segments
sequence: audio
- name: segment_metadata
sequence:
- name: segment_id
dtype: string
- name: filename
dtype: string
- name: speaker
dtype: string
- name: model
dtype: string
- name: actor
dtype: string
- name: original_text
dtype: string
- name: substituted_text
dtype: string
- name: cleaned_text
dtype: string
- name: dialogue_emotion
dtype: string
- name: ravdess_emotion
dtype: string
- name: ref_audio
dtype: string
- name: ref_transcript
dtype: string
- name: start_time
dtype: float32
- name: end_time
dtype: float32
- name: duration
dtype: float32
- name: following_silence
dtype: float32
- name: generation_time
dtype: float32
- name: realtime_factor
dtype: float32
splits:
- name: train
num_bytes: 22200484
num_examples: 3768
download_size: 0
dataset_size: 22200484
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---
# DeepDialogue-xtts

**DeepDialogue-xtts** is a large-scale multimodal dataset containing 40,150 high-quality multi-turn dialogues spanning 41 domains and incorporating 20 distinct emotions with coherent emotional progressions.
This repository contains the XTTS-v2 variant of the dataset, where speech is generated using [XTTS-v2](https://huggingface.co/coqui/XTTS-v2) with explicit emotional conditioning.

[](https://arxiv.org/abs/2505.19978)

## 🚨 Important

This dataset is large (~180GB) due to the inclusion of high-quality audio files. When cloning the repository, ensure you have sufficient disk space and a stable internet connection.

## 💬 Dataset Overview

DeepDialogue pairs 9 different language models (4B-72B parameters) to generate dialogues with emotionally coherent trajectories. Each conversation includes:

- Multi-turn dialogues (3-10 turns) between two AI agents
- Domain-specific content across 41 topics
- Emotional annotations for each utterance (20 distinct emotions)
- High-quality synthesized speech with explicit emotion conditioning
- Paired audio-text data suitable for speech and dialogue research

### Emotional Speech Generation

The XTTS-v2 variant uses reference audio samples from the [RAVDESS dataset](https://zenodo.org/records/1188976) to explicitly condition the speech synthesis on specific emotions, creating natural-sounding emotional expressions in the spoken dialogues.

## 📦 Installation

The dataset contains large audio files and uses Git LFS. To properly clone the repository:

```bash
# Install Git LFS if you haven't already
git lfs install

# Clone the repository (be prepared for a large download, ~180GB)
git clone https://huggingface.co/datasets/SALT-Research/DeepDialogue-xtts
cd DeepDialogue-xtts
```

You can also access specific files through the Hugging Face web interface if you don't need the entire dataset.

## 🗂️ Dataset Structure

The dataset is organized as follows:

```
data/
├── dialogues_[model_combination]/ # Folders grouped by model pairs
│ ├── [dialogue_id].json # JSON representation of the dialogue
│ └── [dialogue_id]/ # Folder containing audio files
│ ├── [dialogue_id]_full.wav # Complete dialogue audio
│ ├── metadata.tsv # Metadata for the audio segments
│ └── segments/ # Individual utterance audio files
│ └── [segment_id]_[speaker]_[emotion].wav
└── train-00000-of-00001.parquet # Metadata for all dialogues
```

### Model Combinations

The dataset includes dialogues from the following model combinations:

```
dialogues_cohere7B dialogues_llama3-70B_qwen2.5-72B dialogues_phi4-14B_gemma3-27B
dialogues_cohere7B_gemma3-4B dialogues_llama3-8B dialogues_phi4-mini
dialogues_gemma3-27B dialogues_llama3-8B_cohere7B dialogues_qwen2.5-32B
dialogues_gemma3-4B dialogues_llama3-8B_gemma3-4B dialogues_qwen2.5-32B_gemma3-27B
dialogues_llama3-70B dialogues_phi4-14B dialogues_qwen2.5-32B_phi4-14B
dialogues_qwen2.5-72B
```

### Domains

The dataset covers 41 distinct domains:

```python
topics = [
"art", "books", "cars", "celebrities", "coding", "cooking", "education", "events",
"fashion", "finance", "fitness", "food", "gaming", "gardening", "health", "history",
"hobbies", "holidays", "home", "languages", "makeup", "movies", "music", "nature",
"news", "pets", "philosophy", "photography", "podcasts", "politics", "relationships",
"science", "shopping", "social_media", "spirituality", "sports", "technology",
"traditions", "travel", "weather", "work"
]
```

### Metadata Structure

The metadata in `metadata.tsv` files includes:

| Field | Description |
|-------|-------------|
| segment_id | Unique identifier for the audio segment |
| filename | Filename of the audio segment |
| speaker | Speaker identifier (LLM1 or LLM2) |
| model | LLM model that generated this turn |
| actor | Voice actor ID from RAVDESS |
| original_text | Original text generated by the LLM |
| substituted_text | Text after any processing |
| cleaned_text | Text after cleaning for TTS |
| dialogue_emotion | Emotion label from dialogue generation |
| ravdess_emotion | Mapped emotion from RAVDESS |
| ref_audio | Reference audio file used for emotion conditioning |
| ref_transcript | Transcript of the reference audio |
| start_time | Start time in the full audio (seconds) |
| end_time | End time in the full audio (seconds) |
| duration | Duration of the segment (seconds) |
| following_silence | Silence after the segment (seconds) |
| generation_time | Time taken to generate the audio (seconds) |
| realtime_factor | Ratio of audio duration to generation time |

## 📊 Usage Examples

### 1. Load the Full Conversations for a Single Model

```python
import os
import json
import pandas as pd
from glob import glob

def load_conversations_for_model(model_dir):
"""Load all conversations for a specific model combination."""
model_path = os.path.join("data", model_dir)

if not os.path.exists(model_path):
print(f"Model directory {model_dir} not found.")
return None

conversations = []

# Get all JSON files in the model directory
json_files = glob(os.path.join(model_path, "*.json"))

for json_file in json_files:
with open(json_file, 'r') as f:
conversation = json.load(f)

# Add model information
conversation['model_dir'] = model_dir
conversations.append(conversation)

print(f"Loaded {len(conversations)} conversations from {model_dir}")
return conversations

# Example usage:
conversations = load_conversations_for_model("dialogues_llama3-70B")
```

### 2. Load the Full Conversations for a Single Topic/Domain

```python
import os
import json
import pandas as pd
from glob import glob

def load_conversations_for_domain(domain, base_path="data"):
"""Load all conversations for a specific domain."""

# First, we'll use the parquet file to find conversations in this domain
parquet_path = os.path.join(base_path, "train-00000-of-00001.parquet")

if os.path.exists(parquet_path):
# Use parquet file for efficient filtering
df = pd.read_parquet(parquet_path)
domain_convs = df[df['domain'] == domain]['conversation_id'].unique()

print(f"Found {len(domain_convs)} conversations in domain '{domain}'")

# Load each conversation JSON
conversations = []
for conv_id in domain_convs:
# Find the model directory for this conversation
model_dir = df[df['conversation_id'] == conv_id]['model_dir'].iloc[0]
json_path = os.path.join(base_path, model_dir, f"{conv_id}.json")

if os.path.exists(json_path):
with open(json_path, 'r') as f:
conversation = json.load(f)
conversations.append(conversation)

return conversations
else:
# Fallback: search through all model directories
print("Parquet file not found, searching through all model directories...")
all_model_dirs = [d for d in os.listdir(base_path) if d.startswith("dialogues_")]

conversations = []
for model_dir in all_model_dirs:
model_path = os.path.join(base_path, model_dir)
json_files = glob(os.path.join(model_path, "*.json"))

for json_file in json_files:
with open(json_file, 'r') as f:
conv = json.load(f)
if conv.get('domain') == domain:
# Add model directory information
conv['model_dir'] = model_dir
conversations.append(conv)

print(f"Found {len(conversations)} conversations in domain '{domain}'")
return conversations

# Example usage:
music_conversations = load_conversations_for_domain("music")
```

### 3. Load All Full Conversations

```python
import os
import json
from glob import glob

def load_all_conversations(base_path="data"):
"""Load all conversations from all model directories."""

# Get all model directories
model_dirs = [d for d in os.listdir(base_path) if d.startswith("dialogues_")]

all_conversations = []
for model_dir in model_dirs:
model_path = os.path.join(base_path, model_dir)
json_files = glob(os.path.join(model_path, "*.json"))

for json_file in json_files:
with open(json_file, 'r') as f:
conversation = json.load(f)

# Add model information
conversation['model_dir'] = model_dir
all_conversations.append(conversation)

print(f"Loaded {len(all_conversations)} conversations from all model directories")
return all_conversations

# Example usage:
all_conversations = load_all_conversations()
```

### 4. Load the Segments of a Full Conversation

```python
import os
import pandas as pd
from IPython.display import Audio
import matplotlib.pyplot as plt
import librosa
import librosa.display
import numpy as np

def load_conversation_segments(conversation_id, model_dir, base_path="data"):
"""Load all segments of a specific conversation with metadata."""

# Path to the conversation directory
conv_dir = os.path.join(base_path, model_dir, conversation_id)

if not os.path.exists(conv_dir):
print(f"Conversation directory not found: {conv_dir}")
return None

# Load metadata
metadata_path = os.path.join(conv_dir, "metadata.tsv")
if os.path.exists(metadata_path):
metadata = pd.read_csv(metadata_path, sep='\t')
else:
print(f"Metadata file not found: {metadata_path}")
return None

# Path to segments directory
segments_dir = os.path.join(conv_dir, "segments")

# Full audio path
full_audio_path = os.path.join(conv_dir, f"{conversation_id}_full.wav")

result = {
'conversation_id': conversation_id,
'model_dir': model_dir,
'metadata': metadata,
'segments_dir': segments_dir,
'full_audio_path': full_audio_path
}

return result

def play_segment(segment_info, index):
"""Play a specific segment from a conversation."""
if segment_info is None:
return

metadata = segment_info['metadata']
if index >= len(metadata):
print(f"Segment index {index} out of range. Max index: {len(metadata)-1}")
return

filename = metadata.iloc[index]['filename']
segment_path = os.path.join(segment_info['segments_dir'], filename)

if os.path.exists(segment_path):
print(f"Playing segment {index+1}/{len(metadata)}: {filename}")
print(f"Text: \"{metadata.iloc[index]['cleaned_text']}\"")
print(f"Emotion: {metadata.iloc[index]['dialogue_emotion']}")
return Audio(segment_path)
else:
print(f"Segment file not found: {segment_path}")

def visualize_segment_waveform(segment_info, index):
"""Visualize the waveform of a specific segment."""
if segment_info is None:
return

metadata = segment_info['metadata']
if index >= len(metadata):
print(f"Segment index {index} out of range. Max index: {len(metadata)-1}")
return

filename = metadata.iloc[index]['filename']
segment_path = os.path.join(segment_info['segments_dir'], filename)

if os.path.exists(segment_path):
# Load the audio file
y, sr = librosa.load(segment_path)

# Create a figure and plot the waveform
plt.figure(figsize=(12, 4))
librosa.display.waveshow(y, sr=sr)
plt.title(f"Waveform: {filename} | Emotion: {metadata.iloc[index]['dialogue_emotion']}")
plt.xlabel("Time (s)")
plt.ylabel("Amplitude")
plt.tight_layout()
plt.show()
else:
print(f"Segment file not found: {segment_path}")

# Example usage:
segment_info = load_conversation_segments("music_85_9", "dialogues_llama3-70B")

# Play a specific segment (e.g., the first one)
if segment_info:
play_segment(segment_info, 0)
visualize_segment_waveform(segment_info, 0)

# Print all segments in this conversation
print("\nAll segments in conversation:")
for i, row in segment_info['metadata'].iterrows():
print(f"{i+1}. Speaker: {row['speaker']} | Emotion: {row['dialogue_emotion']} | Text: \"{row['cleaned_text']}\"")
```

## 🔄 Related Resources

- [DeepDialogue-orpheus](https://huggingface.co/datasets/SALT-Research/DeepDialogue-orpheus): The companion dataset using Orpheus TTS instead of XTTS for speech synthesis
- [Project Website](https://salt-research.github.io/DeepDialogue): Additional information and resources

### 🔗 Links
- **TTS model**: [coqui/XTTS-v2](https://huggingface.co/coqui/XTTS-v2)
- **Emotion source**: [RAVDESS Dataset on Zenodo](https://zenodo.org/records/1188976)

## 📜 Citation

If you use this dataset in your research, please cite our [paper](https://arxiv.org/abs/2505.19978):

```
@misc{koudounas2025deepdialoguemultiturnemotionallyrichspoken,
title={DeepDialogue: A Multi-Turn Emotionally-Rich Spoken Dialogue Dataset},
author={Alkis Koudounas and Moreno La Quatra and Elena Baralis},
year={2025},
eprint={2505.19978},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2505.19978},
}
```

## 📃 License

This dataset is licensed under the [CC BY-NC-SA 4.0 International License](https://creativecommons.org/licenses/by-nc-sa/4.0/).