"Let's Go Real Talk: Spoken Dialogue Model for Face-to-Face Conversation", accepted to ACL 2024 (oral presentation).

**Audio files have been newly processed and re-uploaded on 7/11/2024. Please download the files again for an updated version.

Homepage: https://multidialog.github.io
Paper: https://arxiv.org/pdf/2406.07867
Audio Dataset: https://huggingface.co./datasets/IVLLab/MultiDialog (this repository)
Video Dataset: https://drive.google.com/drive/u/1/folders/1RPMwVHU34yX0R_HbxAWmxF2EHy961HA3

Dataset Description

This dataset includes manually annotated metadata linking audio files to transcriptions, emotions, and other attributes. For access to video files of MultiDialog, download them here.

Dataset Statistics

	train	valid_freq	valid_rare	test_freq	test_rare	Total
# dialogues	7,011	448	443	450	381	8,733
# utterance	151,645	8,516	9,556	9,811	8,331	187,859
avg # utterance/dialogue	21.63	19.01	21.57	21.80	21.87	21.51
avg length/utterance (s)	6.50	6.23	6.40	6.99	6.49	6.51
avg length/dialogue (min)	2.34	1.97	2.28	2.54	2.36	2.33
total length (hr)	273.93	14.74	17.00	19.04	15.01	339.71

Example Usage

There are 'train', 'test_freq', 'test_rare', 'valid_freq', and 'valid_rare' splits. Below is an example usage.

from datasets import load_dataset

MultiD = load_dataset("IVLLab/MultiDialog", "valid_freq", use_auth_token=True)

# see structure
print(MultiD)

# load audio sample on the fly
audio_input = MultiD["valid_freq"][0]["audio"]  # first decoded audio sample
transcription = MultiD["valid_freq"][0]["value"]  # first transcription

Supported Tasks

multimodal dialogue generation : The dataset can be used to train an end-to-end multimodal
automatic-speech-recognition: The dataset can be used to train a model for Automatic Speech Recognition (ASR).
text-to-speech: The dataset can also be used to train a model for Text-To-Speech (TTS).

Languages

Multidialog contains audio and transcription data in English.

Gold Emotion Dialogue Subset

We provide a gold emotion dialogue subset in the MultiDialog dataset, a more reliable resource for studying emotional dynamics in conversations. We classify dialogues from actors that exhibit emotion accuracy above 40% as gold emotion dialogue. Please use dialogues from actors with the following ids: a, b, c, e, f, g, i, j, and k.

Dataset Structure

Data Instances

{
    'file_name': 't_ffa55df6-114d-4b36-87a1-7af6b8b63d9b/t_ffa55df6-114d-4b36-87a1-7af6b8b63d9b_0k.wav'
    'conv_id': 't_ffa55df6-114d-4b36-87a1-7af6b8b63d9b', 
    'utterance_id': 0,
    'from': 'gpt', 
    'audio': 
        {
            # in streaming mode 'path' will be 't_ffa55df6-114d-4b36-87a1-7af6b8b63d9b/t_ffa55df6-114d-4b36-87a1-7af6b8b63d9b_0k.wav'
            'path': '/home/user/.cache/huggingface/datasets/downloads/extracted/cache_id/t_ffa55df6-114d-4b36-87a1-7af6b8b63d9b/t_ffa55df6-114d-4b36-87a1-7af6b8b63d9b_0k.wav, 
            'array': array([0.0005188 , 0.00085449, 0.00012207, ..., 0.00125122, 0.00076294, 0.00036621], dtype=float32), 
            'sampling_rate': 16000
        },
    'value': 'Are you a football fan?', 
    'emotion': 'Neutral', 
    'original_full_path': 'valid_freq/t_ffa55df6-114d-4b36-87a1-7af6b8b63d9b/t_ffa55df6-114d-4b36-87a1-7af6b8b63d9b_0k.wav'
}

Data Fields

file_name (string) - relative file path to the audio sample in the specific split directory.
conv_id (string) - unique identifier for each conversation.
utterance_id (float) - uterrance index.
from (string) - who the message is from (human, gpt).
audio (Audio feature) - a dictionary containing the path to the audio, the decoded audio array, and the sampling rate. In non-streaming mode (default), the path point to the locally extracted audio. In streaming mode, the path is the relative path of an audio. segment inside its archive (as files are not downloaded and extracted locally).
value (string) - transcription of the utterance.
emotion (string) - the emotion of the utterance.
original_full_path (string) - the relative path to the original full audio sample in the original data directory.
speaker_id can be obtained from the last letter of 'file_name' excluding '.wav' (e.g. 'k' in the above example)

Emotion is assigned from the following labels: "Neutral", "Happy", "Fear", "Angry", "Disgusting", "Surprising", "Sad"