new VAD test dataset

[guynich]

I created a VAD test dataset derived from librispeech_asr corpus.

It is now uploaded to HuggingFace datasets.
https://huggingface.co/datasets/guynich/librispeech_asr_test_vad

Running the test splits through Silero-VAD model reveal high AUC scores as expected with clean speech and consistent with results reported in this repo. I put some AUC results on the Dataset Card.

I hope this might be useful to others: such as testing VAD methods with clean speech under additive noise.

[snakers4]

Hi,

Many thanks for the dataset.

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

How is it calculated, which method is used?

[guynich]

Thanks - good feeling to give back.

I don’t think it is a published method. It computes energy in short blocks of the audio signal. Using a lookback window the lowest energy block value is fed to a smoothing function to form an estimate of the noise. Speech is detected when the energy of recent blocks rise above a computed threshold above the noise estimate. It works well here because the source dataset is speech without significant noise. It is not suitable for real world voice activity detection.

[snakers4]

I see! Many thanks for your input!

[guynich]

I added a VAD test set with seven languages.
https://huggingface.co/datasets/guynich/multilingual_librispeech_test_vad
High AUC scores are seen with Silero-VAD model - results.

[snakers4]

Nice!