Method for adding noise to a signal

madmom/audio/signal.py

@@ -322,6 +322,38 @@ def trim(signal, where='fb'):
     return signal[first:last]
 
 
+def add_noise(signal, snr=10.):
+    """
+    Add Additive White Gaussian Noise (AWGN) to a signal.
+
+    Parameters
+    ----------
+    signal : numpy array
+        Original signal.
+    snr : float, optional
+        The Signal-to-noise ratio determines how strong the noise should be
+        relative to the signal. A lower value results in a stronger noise.
+
+    Returns
+    -------
+    numpy array
+        Signal with added noise.
+
+    """
+    # make sure SNR is in allowed range
+    if snr <= 0:
+        raise ValueError("Invalid value for SNR, must be greater than zero.")
+    # generate gaussian random noise
+    noise = np.random.normal(0, 1, signal.shape)
+    # compute power for signal and noise
+    power_signal = root_mean_square(signal)
+    power_noise = root_mean_square(noise)
+    # scale noise to match SNR
+    noise *= (power_signal / power_noise) / snr
+    # return signal with added noise
+    return signal + noise.astype(signal.dtype)
+
+
 def energy(signal):
     """
     Compute the energy of a (framed) signal.

tests/test_audio_signal.py

@@ -546,6 +546,66 @@ def test_values(self):
         self.assertTrue(np.allclose(result, [5, 4, 3, 2, 1]))
 
 
+class TestAddNoiseFunction(unittest.TestCase):
+
+    def test_types(self):
+        # mono signals
+        result = add_noise(sig_1d)
+        self.assertTrue(type(result) == type(sig_1d))
+        self.assertTrue(result.ndim == sig_1d.ndim)
+        self.assertTrue(result.dtype == sig_1d.dtype)
+        signal = Signal(sample_file)
+        result = add_noise(signal)
+        self.assertIsInstance(result, Signal)
+        self.assertIsInstance(result, np.ndarray)
+        self.assertTrue(result.ndim == signal.ndim)
+        self.assertTrue(result.dtype == np.int16)
+        # multi-channel signals
+        result = trim(sig_2d)
+        self.assertTrue(type(result) == type(sig_2d))
+        self.assertTrue(len(result) == len(sig_2d))
+        self.assertTrue(result.ndim == sig_2d.ndim)
+        self.assertTrue(result.dtype == sig_2d.dtype)
+        signal = Signal(stereo_sample_file)
+        result = trim(signal)
+        self.assertIsInstance(result, Signal)
+        self.assertIsInstance(result, np.ndarray)
+        self.assertTrue(result.ndim == signal.ndim)
+        self.assertTrue(result.dtype == np.int16)
+
+    def test_errors(self):
+        with self.assertRaises(ValueError):
+            add_noise(sig_1d, 0)
+        with self.assertRaises(ValueError):
+            add_noise(sig_1d, -1)
+
+    def test_values(self):
+        # fixed random seed
+        np.random.seed(0)
+        # mono signals
+        result = add_noise(sig_1d, snr=10.)
+        self.assertTrue(np.allclose(result, sig_1d, atol=0.11))
+        result = add_noise(sig_1d, snr=20.)
+        self.assertTrue(np.allclose(result, sig_1d, atol=0.06))
+        # same with int dtype
+        result = add_noise(sig_1d.astype(np.int), 10.)
+        self.assertTrue(np.allclose(result, sig_1d))
+        result = add_noise(sig_1d.astype(np.int), 0.5)
+        self.assertTrue(np.allclose(result, [0, 1, 2, 0, 0, 1, -2, 0, 1]))
+        # multi-channel signals
+        result = add_noise(sig_2d, snr=10.)
+        self.assertTrue(np.allclose(result, sig_2d, atol=0.13))
+        result = add_noise(sig_2d, snr=20.)
+        self.assertTrue(np.allclose(result, sig_2d, atol=0.08))
+        # same with int dtype
+        result = add_noise(sig_2d.astype(np.int), 10.)
+        self.assertTrue(np.allclose(result, sig_2d))
+        result = add_noise(sig_2d.astype(np.int), 0.5)
+        self.assertTrue(np.allclose(result, [[0, 2], [0, 1], [1, 1],
+                                             [0, 2], [0, 1], [3, -1],
+                                             [-1, 2], [-1, 2], [1, 1]]))
+
+
 class TestEnergyFunction(unittest.TestCase):
 
     def test_types(self):