Method __call__

data/waveform_mixers.py:19–62

(self, waveforms)

        }

    def __call__(self, waveforms):
        
        batch_size = waveforms.shape[0]

        data_dict = {
            'segment': [],
            'mixture': [],
        }

        for n in range(0, batch_size):

            segment = waveforms[n].clone()

            # create zero tensors as the background template
            noise = torch.zeros_like(segment)

            mix_num = random.randint(2, self.max_mix_num)
            assert mix_num >= 2

            for i in range(1, mix_num):
                next_segment = waveforms[(n + i) % batch_size]
                rescaled_next_segment = dynamic_loudnorm(audio=next_segment, reference=segment, **self.loudness_param)
                noise += rescaled_next_segment

            # randomly normalize background noise
            noise = dynamic_loudnorm(audio=noise, reference=segment, **self.loudness_param)

            # create audio mixyure
            mixture = segment + noise

            # declipping if need be
            max_value = torch.max(torch.abs(mixture))
            if max_value > 1:
                segment *= 0.9 / max_value
                mixture *= 0.9 / max_value

            data_dict['segment'].append(segment)
            data_dict['mixture'].append(mixture)

        for key in data_dict.keys():
            data_dict[key] = torch.stack(data_dict[key], dim=0)

        # return data_dict
        return data_dict['mixture'], data_dict['segment']


def rescale_to_match_energy(segment1, segment2):