Merge pull request #182 from hbredin/chore/remove-librosa

chore: remove librosa dependency in favor of torchaudio

.github/workflows/test_requirements.txt

@@ -4,7 +4,6 @@ torch==1.11.0
 torchaudio==0.11.0
 audioread>=2.1.8
 julius>=0.2.3,<0.3
-librosa==0.9.1
 py-cpuinfo>=7.0.0
 pytest==5.3.4
 pytest-cov==2.8.1

README.md

@@ -217,6 +217,10 @@ classification. It was successfully applied in the paper
 
 * Add new transforms: `Mix`, `Padding`, `RandomCrop` and `SpliceOut`
 
+### Changes
+
+* Remove `librosa` dependency in favor of `torchaudio`
+
 ## [v0.11.2] - 2025-01-09
 
 ### Fixed

environment.yml

@@ -19,7 +19,6 @@ dependencies:
   - black==23.12.1
   - coverage==5.3
   - julius>=0.2.3,<0.3
-  - librosa==0.9.1
   - pandas==1.1.4
   - py-cpuinfo==7.0.0
   - pytest==7.4.4

scripts/demo.py

@@ -3,7 +3,6 @@
 import time
 from pathlib import Path
 
-import librosa
 import numpy as np
 import torch
 from scipy.io import wavfile
@@ -30,6 +29,7 @@
 from torch_audiomentations.augmentations.splice_out import SpliceOut
 from torch_audiomentations.core.transforms_interface import ModeNotSupportedException
 from torch_audiomentations.utils.object_dict import ObjectDict
+from torch_audiomentations.utils.io import Audio
 
 SAMPLE_RATE = 44100
 
@@ -88,14 +88,14 @@ def __exit__(self, type, value, traceback):
     random.seed(43)
 
     filenames = ["perfect-alley1.ogg", "perfect-alley2.ogg"]
-    samples1, _ = librosa.load(
-        os.path.join(TEST_FIXTURES_DIR, filenames[0]), sr=SAMPLE_RATE, mono=False
-    )
-    samples2, _ = librosa.load(
-        os.path.join(TEST_FIXTURES_DIR, filenames[1]), sr=SAMPLE_RATE, mono=False
-    )
-    samples = np.stack((samples1, samples2), axis=0)
-    samples = torch.from_numpy(samples)
+    audio = Audio(SAMPLE_RATE, mono=True)
+    samples1 = audio(os.path.join(TEST_FIXTURES_DIR, filenames[0]))
+    _, num_samples1 = samples1.shape
+    samples2 = audio(os.path.join(TEST_FIXTURES_DIR, filenames[1]))
+    _, num_samples2 = samples2.shape
+    num_samples = min(num_samples1, num_samples2)
+    samples = torch.stack([samples1[:, :num_samples], samples2[:, :num_samples]], dim=0)
+
 
     modes = ["per_batch", "per_example", "per_channel"]
     for mode in modes:

scripts/measure_convolve_execution_time.py

@@ -1,4 +1,3 @@
-import librosa
 import numpy as np
 import torch
 from cpuinfo import get_cpu_info
@@ -8,14 +7,14 @@
 from scripts.demo import TEST_FIXTURES_DIR, timer
 from scripts.plot import show_horizontal_bar_chart
 from torch_audiomentations.utils.convolution import convolve as torch_convolve
+from torch_audiomentations.utils.io import Audio
 
 if __name__ == "__main__":
     file_path = TEST_FIXTURES_DIR / "acoustic_guitar_0.wav"
     sample_rate = 48000
-    samples, _ = librosa.load(file_path, sr=sample_rate)
-    ir_samples, _ = librosa.load(
-        TEST_FIXTURES_DIR / "ir" / "impulse_response_0.wav", sr=sample_rate
-    )
+    audio = Audio(sample_rate, mono=True)
+    samples = audio(file_path).numpy()
+    ir_samples = audio(TEST_FIXTURES_DIR / "ir" / "impulse_response_0.wav").numpy()
 
     is_cuda_available = torch.cuda.is_available()
     print("Is torch CUDA available:", is_cuda_available)

setup.py

@@ -38,7 +38,6 @@ def find_version(*file_paths):
     ),
     install_requires=[
         "julius>=0.2.3,<0.3",
-        "librosa>=0.6.0",
         "torch>=1.7.0",
         "torchaudio>=0.9.0",
         "torch-pitch-shift>=1.2.2",

tests/test_background_noise.py

@@ -13,27 +13,18 @@
 
 from torch_audiomentations import AddBackgroundNoise
 from torch_audiomentations.utils.dsp import calculate_rms
-from torch_audiomentations.utils.file import load_audio
 from .utils import TEST_FIXTURES_DIR
-
+from torch_audiomentations.utils.io import Audio
 
 class TestAddBackgroundNoise(unittest.TestCase):
     def setUp(self):
         self.sample_rate = 16000
         self.batch_size = 16
         self.empty_input_audio = torch.empty(0, 1, 16000)
         # TODO: use utils.io.Audio
-        self.input_audio = (
-            torch.from_numpy(
-                load_audio(
-                    TEST_FIXTURES_DIR / "acoustic_guitar_0.wav",
-                    sample_rate=self.sample_rate,
-                )
-            )
-            .unsqueeze(0)
-            .unsqueeze(0)
-        )
 
+        audio = Audio(self.sample_rate, mono=True)
+        self.input_audio = audio(TEST_FIXTURES_DIR / "acoustic_guitar_0.wav")[None]
         self.input_audios = torch.cat([self.input_audio] * self.batch_size, dim=0)
 
         self.bg_path = TEST_FIXTURES_DIR / "bg"

tests/test_convolution.py

@@ -1,21 +1,20 @@
-import librosa
 import torch
 from numpy.testing import assert_almost_equal
 from scipy.signal import convolve as scipy_convolve
 
 from tests.utils import TEST_FIXTURES_DIR
 from torch_audiomentations.utils.convolution import convolve as torch_convolve
-
+from torch_audiomentations.utils.io import Audio
 
 class TestConvolution:
     def test_convolve(self):
         sample_rate = 16000
 
         file_path = TEST_FIXTURES_DIR / "acoustic_guitar_0.wav"
-        samples, _ = librosa.load(file_path, sr=sample_rate)
-        ir_samples, _ = librosa.load(
-            TEST_FIXTURES_DIR / "ir" / "impulse_response_0.wav", sr=sample_rate
-        )
+        audio = Audio(sample_rate, mono=True)
+        samples = audio(file_path).numpy()
+        ir_samples = audio(TEST_FIXTURES_DIR / "ir" / "impulse_response_0.wav").numpy()
+
 
         expected_output = scipy_convolve(samples, ir_samples)
         actual_output = torch_convolve(

tests/test_impulse_response.py

@@ -4,7 +4,7 @@
 import torch
 
 from torch_audiomentations import ApplyImpulseResponse
-from torch_audiomentations.utils.file import load_audio
+from torch_audiomentations.utils.io import Audio
 from .utils import TEST_FIXTURES_DIR
 
 
@@ -15,16 +15,8 @@ def sample_rate():
 
 @pytest.fixture
 def input_audio(sample_rate):
-    return (
-        torch.from_numpy(
-            load_audio(
-                os.path.join(TEST_FIXTURES_DIR, "acoustic_guitar_0.wav"),
-                sample_rate=sample_rate,
-            )
-        )
-        .unsqueeze(0)
-        .unsqueeze(0)
-    )
+    audio = Audio(sample_rate, mono=True)
+    return audio(os.path.join(TEST_FIXTURES_DIR, "acoustic_guitar_0.wav"))[None]
 
 
 @pytest.fixture

tests/test_mix.py

@@ -4,32 +4,16 @@
 
 from torch_audiomentations.augmentations.mix import Mix
 from torch_audiomentations.utils.dsp import calculate_rms
-from torch_audiomentations.utils.file import load_audio
+from torch_audiomentations.utils.io import Audio
 from .utils import TEST_FIXTURES_DIR
 
 
 class TestMix(unittest.TestCase):
     def setUp(self):
         self.sample_rate = 16000
-        self.guitar = (
-            torch.from_numpy(
-                load_audio(
-                    TEST_FIXTURES_DIR / "acoustic_guitar_0.wav",
-                    sample_rate=self.sample_rate,
-                )
-            )
-            .unsqueeze(0)
-            .unsqueeze(0)
-        )
-        self.noise = (
-            torch.from_numpy(
-                load_audio(
-                    TEST_FIXTURES_DIR / "bg" / "bg.wav", sample_rate=self.sample_rate
-                )
-            )
-            .unsqueeze(0)
-            .unsqueeze(0)
-        )
+        audio = Audio(self.sample_rate, mono=True)
+        self.guitar = audio(TEST_FIXTURES_DIR / "acoustic_guitar_0.wav")[None]
+        self.noise = audio(TEST_FIXTURES_DIR / "bg" / "bg.wav")[None]
 
         common_num_samples = min(self.guitar.shape[-1], self.noise.shape[-1])
         self.guitar = self.guitar[:, :, :common_num_samples]

torch_audiomentations/utils/dsp.py

@@ -24,14 +24,6 @@ def calculate_desired_noise_rms(clean_rms, snr):
     return noise_rms
 
 
-def resample_audio(audio, orig_sr, target_sr):
-    # TODO: We can probably remove this function and call resample directly where needed
-    """Resamples the audio to a new sampling rate."""
-    import librosa
-
-    return librosa.resample(audio, orig_sr=orig_sr, target_sr=target_sr)
-
-
 def convert_decibels_to_amplitude_ratio(decibels):
     return 10 ** (decibels / 20)
 

torch_audiomentations/utils/file.py

@@ -2,9 +2,6 @@
 from pathlib import Path
 from typing import List, Union
 
-import soundfile
-
-from .dsp import resample_audio
 
 SUPPORTED_EXTENSIONS = (".wav",)
 
@@ -64,16 +61,3 @@ def find_audio_files(
             break
 
     return file_paths
-
-
-def load_audio(audio_file_path, sample_rate=None, start=0, stop=None):
-    # TODO: Clarify whether start/stop is in samples or in seconds, and whether or not it
-    #  relates to the original or the resampled audio.
-    """Loads the audio given the path of an audio file."""
-    audio, source_sample_rate = soundfile.read(audio_file_path, start=start, stop=stop)
-
-    if sample_rate:
-        audio = resample_audio(audio, source_sample_rate, sample_rate)
-
-    # TODO: return sample rate as well
-    return audio

torch_audiomentations/utils/io.py

@@ -1,8 +1,6 @@
-import warnings
 from pathlib import Path
 from typing import Text, Union
 
-import librosa
 import torch
 import torchaudio
 from torch import Tensor
@@ -155,18 +153,7 @@ def downmix_and_resample(self, samples: Tensor, sample_rate: int) -> Tensor:
 
         # resample
         if self.sample_rate != sample_rate:
-            samples = samples.numpy()
-            if self.mono:
-                # librosa expects mono audio to be of shape (n,), but we have (1, n).
-                samples = librosa.core.resample(
-                    samples[0], orig_sr=sample_rate, target_sr=self.sample_rate
-                )[None]
-            else:
-                samples = librosa.core.resample(
-                    samples.T, orig_sr=sample_rate, target_sr=self.sample_rate
-                ).T
-
-            samples = torch.tensor(samples)
+            samples = torchaudio.functional.resample(samples, sample_rate, self.sample_rate)
 
         return samples