How to evaluate on VOiCES dataset?

[RamakrishnaChaitanya]

Hi, can you please share the procedure to prepare the VOiCES dataset for evaluating the VAD models? Is there any documentation already available that explains how to generate the .wav files and the corresponding labels (0/1's)? At the higher level, the folder was like this for the VOiCES dataset.

[YannisStyl]

Hi!

To prepare the VOiCES dataset for evaluation, we used the development set, which provides both the clean 'source' audio and the corresponding 'distant' (noisy/reverberant) recordings.

We used the Montreal Forced Aligner to align the clean source audio with the provided transcripts. This generates the start and end timestamps for the speech segments. Since the distant recordings are synchronized playbacks of the source audio, we used the timestamps from the clean alignment as the ground truth for the noisy files. We then converted these timestamps into frame-level to match the frame rate of our VAD models.

I am not sure if this fully answers your question but for more information on the VOiCES dataset please refer to https://iqtlabs.github.io/voices/Lab41-SRI-VOiCES_README/