README.md

---
language:
- en
pretty_name: librispeech_asr_test_vad
tags:
- speech
license: cc-by-4.0
task_categories:
- text-classification
---

Voice Activity Detection (VAD) Test Dataset

This dataset is based on the `test.clean` and `test.other` splits from the
[librispeech_asr](https://huggingface.co/datasets/openslr/librispeech_asr)
corpus. It includes two binary labels:

- **speech**: Indicates presence of speech ([0, 1]), computed using a dynamic threshold method with background noise estimation and smoothing.

- **confidence**: A post-processing flag to optionally correct transient dropouts in speech. It is set to 1 by default, but switches to 0 for up to ~0.1 seconds (3 chunks of audio) following a transition from speech to silence.  Approximately 7% of the `speech` labels in this dataset are `confidence` 0.  The remaining 93% are `confidence` 1 enabling VAD testing.

The dataset has minimal background noise, making it suitable for mixing with
external noise samples to test VAD robustness.


## Example data

A plot for an example showing audio samples and the `speech` feature.

<img src="assets/test_other_item_02.png" alt="Example from test.other"/>

The example below shows brief dropouts in the `speech` feature during natural
short pauses of quiet in the talker's speech.  Since some VAD models may react
more slowly, the `confidence` feature offers a way to optionally ignore these
transient droputs when evaluating performance.

<img src="assets/test_clean_item_02.png" alt="Example from test.other"/>

# Example usage of dataset

The VAD model under test must support processing a chunk size of 512 audio
samples at 16000 Hz generating a prediction for each `speech` feature.

```console
import datasets
import numpy as np
from sklearn.metrics import roc_auc_score

dataset = datasets.load_dataset("guynich/librispeech_asr_test_vad")

audio = dataset["test.clean"][0]["audio"]["array"]
speech = dataset["test.clean"][0]["speech"]

# Compute voice activity probabilities
speech_probs = vad_model(audio)

# Add test code here
roc_auc = roc_auc_score(speech, speech_probs)
```
In practice you would run the AUC computation across the entire test split.

## Ignore transient dropouts

The `confidence` values can be used to filter the data. Removing zero confidence
values excludes 6.8% of the dataset and causes numerical increase in
computed precision.  This compensates for slower moving voice activity decisions
as encountered in real-world applications.

```console
confidence = dataset["test.clean"][0]["confidence"]

speech_array = np.array(speech)
speech_probs_array = np.array(speech_probs)

roc_auc_confidence = roc_auc_score(
    speech_array[np.array(confidence) == 1],
    speech_probs_array[np.array(confidence) == 1],
)
```

# Model evaluation example

Example AUC plots computed for
[Silero VAD](https://github.com/snakers4/silero-vad?tab=readme-ov-file)
model with `test.clean` split.

<img src="assets/roc_test_clean.png" alt="Example from test.clean with Silero-VAD"/>

Precision values are increased when data is sliced by `confidence` values.
These low-confidence `speech` labels are flagged rather than removed, allowing
users to either exclude them (as shown here) or handle them with other methods.

<img src="assets/roc_test_clean_exclude_low_confidence.png" alt="Example from test.clean with Silero-VAD"/>

# License Information

This derivative dataset retains the same license as the source dataset
[librispeech_asr](https://huggingface.co/datasets/openslr/librispeech_asr).

[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)


# Citation Information

Labels contributed by Guy Nicholson were added to the following dataset.
```
@inproceedings{panayotov2015librispeech,
  title={Librispeech: an ASR corpus based on public domain audio books},
  author={Panayotov, Vassil and Chen, Guoguo and Povey, Daniel and Khudanpur, Sanjeev},
  booktitle={Acoustics, Speech and Signal Processing (ICASSP), 2015 IEEE International Conference on},
  pages={5206--5210},
  year={2015},
  organization={IEEE}
}
```