Datasets:
license: cc-by-nc-4.0
language:
- en
pretty_name: CBT-Audio
size_categories:
- 1K<n<10K
tags:
- audio
- speech
- mental-health
- cognitive-behavioral-therapy
- distress-estimation
- audio-language-models
CBT-Audio
CBT-Audio is a dataset for studying patient-side distress intensity estimation in CBT-style educational recordings.
The dataset contains 1,802 patient-side turns from 96 publicly available CBT educational recordings. Each turn includes a transcript, preceding conversational context, source-video timestamp metadata, and a 1–5 distress-intensity reference label based on a PANAS-adapted label scale.
This release contains text, labels, timestamps, and metadata only. Raw audio is not redistributed.
Files
| File | Rows | Description |
|---|---|---|
data/labeled_patient_turns.csv |
1,802 | Patient-side turns with transcripts, context, timestamps, and reference labels |
data/source_recordings.csv |
96 | Source recording metadata, including URLs and titles |
labeled_patient_turns.csv
Each row corresponds to one patient-side turn.
| Column | Type | Description |
|---|---|---|
global_id |
string | Unique turn identifier, formatted as NNN_NNN |
video_id |
string | Zero-padded source recording ID. Joins with source_recordings.csv |
turn_id |
integer | Turn index within the source recording |
reference_label |
integer | Distress-intensity reference label from 1 to 5 |
patient_transcript |
string | Transcript of the patient-side utterance |
context |
string | Up to three preceding conversational turns used as context |
start_time_sec |
float | Turn start time in seconds, relative to the source recording |
end_time_sec |
float | Turn end time in seconds, relative to the source recording |
source_recordings.csv
Each row corresponds to one source recording.
| Column | Type | Description |
|---|---|---|
video_id |
string | Zero-padded source recording ID |
youtube_url |
string | Public URL of the source recording |
original_title |
string | Source recording title |
Audio reconstruction
Raw audio and reconstructed clips are not included in this Hugging Face dataset.
The released metadata includes source URLs and timestamps so that users can reconstruct clips locally from the original recordings. Users are responsible for accessing the original source recordings lawfully, following YouTube's terms, respecting the original creators' rights, and not redistributing downloaded audio or reconstructed clips.
Intended use
CBT-Audio is intended for research on audio language models, multimodal evaluation, and distress-intensity estimation in CBT-style educational recordings.
It is not a clinical diagnostic dataset. The labels should not be treated as clinical ground truth or used to support individual-level clinical assessments.
Scope
CBT-Audio is based on publicly available English-language CBT educational recordings hosted on Youtube. The source material consists of educational content, including demonstrations and case-style recordings, rather than private clinical therapy sessions.
The dataset is focused on patient-side turns and a 1–5 distress-intensity scale. The labels are intended for research evaluation and should be interpreted as reference labels for this task, not as clinical ground truth.
License
The CBT-Audio dataset files released here are licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). This applies to dataset files created by the authors, including labels, metadata, timestamps.
Raw audio is not redistributed. The original source recordings remain subject to the terms and rights of the original platforms and creators. This release does not grant permission to redistribute the original videos, downloaded audio, or reconstructed audio clips.