|
--- |
|
license: mit |
|
--- |
|
|
|
# EDEN ASR Dataset |
|
A subset of this data was used to support the development of empathetic feedback modules in [EDEN](https://arxiv.org/abs/2406.17982) and [its prior work](https://arxiv.org/abs/2404.13764). |
|
|
|
The dataset contains audio clips of native Mandarin speakers. The speakers conversed with a chatbot hosted on an [English practice platform](https://dl.acm.org/doi/abs/10.1145/3491140.3528329?casa_token=ER-mfy0xauQAAAAA:FyDgmH0Y0ke7a6jpOnuycP1HRfeV1B5qaq5JWM5OV5dB9fLFL_vzVRUacZ4fUMRBDl71UeWMIA9Z). |
|
3081 audio clips from 613 conversations and 163 users remained after filtering. |
|
The filtering process removes audio clips containing only Mandarin, duplicates, and a subset of self-introductions from the users. |
|
Each audio clip ranges from one second to two minutes. We did not collect demographic information to protect user identities. |
|
|
|
In our original work, we directly transcribed the speech with Whisper Medium. However, since the audio clips are accented speech, these transcripts have instances of ASR error. We have **manually verified** the original transcripts to ensure they are high-quality. |
|
|
|
Thanks to my mentee, Brittney Lilly, for doing the crux of the ASR transcript verification work! |
|
|
|
## Dataset Columns |
|
* **audio_url**: The URL to the audio clips; you can download them using wget or urllib to your local machines (see the code snippet below). |
|
* **emotion_label**: We manually labeled a subset of the clips as **Neutral** (neutral emotions), **Negative** (the speaker is displaying negative emotions), or **Pauses** (there are a lot of pauses in the speech, potentially signaling language anxiety). The process is documented in [EDEN's prior work](https://arxiv.org/abs/2404.13764). When this field is empty, it means that the clip was not labeled. |
|
* **corrected_whisper_transcript**: The high-quality transcript we have verified; the person performing verification is a native English speaker. |
|
|
|
## Intended Use |
|
Accented ASR research! |
|
|
|
## Code Example |
|
|
|
```python |
|
from datasets import load_dataset |
|
dataset = load_dataset("sylviali/EDEN_ASR_Data", split="train") |
|
|
|
print(dataset) |
|
|
|
# Download the audio to a local file |
|
import urllib |
|
urllib.request.urlretrieve(dataset[0]["audio_url"], "audio.wav") |
|
|
|
# Extract the ASR transcript |
|
print(dataset[0]["corrected_whisper_transcript"]) |
|
|
|
# Extract the emotion label |
|
print(dataset[0]["emotion_label"]) |
|
|
|
# Check audio clips labeled as having negative emotions |
|
negative_emotion_clips = dataset.filter(lambda example: example["emotion_label"] == "Negative") |
|
print(len(negative_emotion_clips)) |
|
print(negative_emotion_clips[0]) |
|
``` |
|
|
