license: apache-2.0
dataset_info:
features:
- name: audio
dtype: audio
- name: transcript_annotated
dtype: string
- name: transcript_a
dtype: string
- name: transcript_b
dtype: string
- name: transcript_c
dtype: string
splits:
- name: train
num_bytes: 1343263191.5
num_examples: 4500
- name: validation
num_bytes: 75479207
num_examples: 250
- name: test
num_bytes: 72139425
num_examples: 250
download_size: 1482840572
dataset_size: 1490881823.5
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
DisfluencySpeech Dataset
The DisfluencySpeech Dataset is a single-speaker studio-quality labeled English speech dataset with paralanguage. A single speaker recreates nearly 10 hours of expressive utterances from the Switchboard-1 Telephone Speech Corpus (Switchboard), simulating realistic informal conversations. To aid the development of a text-to-speech (TTS) model that is able to predictively synthesise paralanguage from text without such components, we provide three different transcripts at different levels of information removal (removal of non-speech events, removal of non-sentence elements, and removal of false starts).
Benchmark TTS models for each transcript can be found here.
Dataset Details
All audio files are provided as 22,050 hz .wav files. In the metadata.csv file, 4 different transcripts for each file are provided, each at differing levels of information removal:
- transcript_annotated is a full transcript retaining all non-speech event and disfluency annotations;
- transcript_a contains all textual content recorded, including non-sentence elements and restarts. Only non-speech events such as laughter and sighs are removed from transcript;
- transcript_b is transcript_a but with filled pauses, explicit editing terms, and discourse markers removed. Coordinating conjunctions and asides are left in, as they are non-sentence elements as well, they are often used to convey meaning; and
- transcript_c is transcript_b but with false starts removed. This is the most minimal transcript.
The training set contains 90% of the data, the validation set contains 5% of the data, and the test set contains 5% of the data.
Citation
If you use this dataset, please cite the paper in which it is presented: Kyra Wang, Dorien Herremans, 2024, DisfluencySpeech - Single-Speaker Conversational Speech Dataset with Paralanguage.