Why is speed perturbation treated as different speakers?

[TranThanh96]

Hi, thanks for releasing this great project!

I noticed in the code that when speed_perturb is enabled, the number of classes is multiplied by 3 (line 130, train.py):

    if configs['data_type'] != 'feat' and configs['dataset_args'][
            'speed_perturb']:
        # diff speed is regarded as diff spk
        configs['projection_args']['num_class'] *= 3
        if configs.get('do_lm', False):
            logger.info(
                'No speed perturb while doing large margin fine-tuning')
            configs['dataset_args']['speed_perturb'] = False

This seems to treat each speed-perturbed version of an utterance as if it were from a different speaker.
I would expect speed perturbation to keep the same speaker label (since the identity doesn’t change, only the speaking rate).

Could you clarify the motivation or provide references for this design choice?

Is there a specific paper or benchmark showing that treating speed-perturbed audio as different speakers improves performance?

Wouldn’t this risk confusing the model by artificially inflating the number of classes?

Thanks a lot!

[JiJiJiang]

The implementation of our speed perturb would also change the pitch of audio, which sounds differently from the original speaker. You can select some samples and listen by yourself.
Of course, other data augmentation of speed perturb would keep the pitch and only change the speed of audio, then the speaker id should be kept the same.

[JiJiJiang]

Check the difference between:

1. sox input.wav speed_1.1x_pitch_fixed.wav speed 1.1 -p
2. sox input.wav speed_0.8x_pitch_down.wav speed 0.8