RandomResample

audiomentations/__init__.py

@@ -21,6 +21,7 @@
 from .augmentations.peaking_filter import PeakingFilter
 from .augmentations.pitch_shift import PitchShift
 from .augmentations.polarity_inversion import PolarityInversion
+from .augmentations.random_resample import RandomResample
 from .augmentations.resample import Resample
 from .augmentations.reverse import Reverse
 from .augmentations.shift import Shift

audiomentations/augmentations/random_resample.py

@@ -0,0 +1,78 @@
+import random
+import librosa
+import numpy as np
+from audiomentations.core.transforms_interface import BaseWaveformTransform
+
+
+class RandomResample(BaseWaveformTransform):
+    """
+    Resample signal using librosa.core.resample in one direction than resample back to original shape.
+    """
+
+    supports_multichannel = True
+
+    def __init__(self, min_sample_rate=8000, max_sample_rate=44100, methods=('kaiser_fast', 'fft'), p=0.5):
+        """
+        :param min_sample_rate: int, Minimum sample rate
+        :param max_sample_rate: int, Maximum sample rate
+        :param methods: int, Resample type. If several added it will be chosen randonly.
+        All possible values see in: librosa.core.audio.resample
+        :param p: The probability of applying this transform
+        """
+        super().__init__(p)
+        assert min_sample_rate <= max_sample_rate
+        self.min_sample_rate = min_sample_rate
+        self.max_sample_rate = max_sample_rate
+        self.methods = methods
+
+    def randomize_parameters(self, samples, sample_rate):
+        super().randomize_parameters(samples, sample_rate)
+        if self.parameters["should_apply"]:
+            self.parameters["target_sample_rate"] = random.randint(
+                self.min_sample_rate, self.max_sample_rate
+            )
+            self.parameters["target_method"] = random.choice(
+                self.methods
+            )
+
+    def apply(self, samples, sample_rate):
+        orig_shape = samples.shape
+        samples = librosa.core.resample(
+            samples,
+            orig_sr=sample_rate,
+            target_sr=self.parameters["target_sample_rate"],
+            res_type=self.parameters["target_method"],
+        )
+        samples = librosa.core.resample(
+            samples,
+            orig_sr=self.parameters["target_sample_rate"],
+            target_sr=sample_rate,
+            res_type=self.parameters["target_method"],
+        )
+        new_shape = samples.shape
+        # print(self.parameters["target_sample_rate"], self.parameters["target_method"])
+        if orig_shape[-1] < new_shape[-1]:
+            # print('Shape less! {} != {}'.format(orig_shape, new_shape))
+            samples = samples[..., :orig_shape[-1]]
+        elif orig_shape[-1] > new_shape[-1]:
+            # print('Shape more! {} != {}'.format(orig_shape, new_shape))
+            samples = np.pad(samples, ((0, 0), (0, orig_shape[-1] - new_shape[-1])), 'constant')
+        return samples
+
+
+"""
+Time statistics for different methods:
+Aug: RandomResample Type: kaiser_best Time: 56.39 sec Per sample: 0.563885 sec
+Aug: RandomResample Type: kaiser_fast Time: 15.02 sec Per sample: 0.150230 sec
+Aug: RandomResample Type: scipy/fft Time: 12.38 sec Per sample: 0.123800 sec
+Aug: RandomResample Type: linear Time: 2.48 sec Per sample: 0.024790 sec
+Aug: RandomResample Type: zero_order_hold Time: 1.83 sec Per sample: 0.018273 sec
+Aug: RandomResample Type: sinc_best Time: 70.59 sec Per sample: 0.705902 sec
+Aug: RandomResample Type: sinc_medium Time: 19.38 sec Per sample: 0.193758 sec
+Aug: RandomResample Type: sinc_fastest Time: 10.09 sec Per sample: 0.100885 sec
+Aug: RandomResample Type: soxr_vhq Time: 5.58 sec Per sample: 0.055791 sec
+Aug: RandomResample Type: soxr_hq Time: 2.59 sec Per sample: 0.025902 sec
+Aug: RandomResample Type: soxr_mq Time: 2.42 sec Per sample: 0.024174 sec
+Aug: RandomResample Type: soxr_lq Time: 2.36 sec Per sample: 0.023578 sec
+Aug: RandomResample Type: soxr_qq Time: 1.82 sec Per sample: 0.018217 sec
+"""