APM Quality Generator, noise generator and evaluation score workers factory + echo noise generator

webrtc/modules/audio_processing/test/py_quality_assessment/quality_assessment/noise_generation.py

Index: webrtc/modules/audio_processing/test/py_quality_assessment/quality_assessment/noise_generation.py
diff --git a/webrtc/modules/audio_processing/test/py_quality_assessment/quality_assessment/noise_generation.py b/webrtc/modules/audio_processing/test/py_quality_assessment/quality_assessment/noise_generation.py
index ef1eec99aa159a9ac1b65e8562d2a6fa8a90a535..a6a8a0653c7800e3efda9be678565dc278041605 100644
--- a/webrtc/modules/audio_processing/test/py_quality_assessment/quality_assessment/noise_generation.py
+++ b/webrtc/modules/audio_processing/test/py_quality_assessment/quality_assessment/noise_generation.py
@@ -7,25 +7,32 @@
 # be found in the AUTHORS file in the root of the source tree.
 
 """Noise generators producing pairs of signals intended to be used to test the
-   APM module. Each pair consists of a noisy and a reference signal. The former
-   is used as input for APM, and it is generated by adding noise to a signal.
-   The reference is the expected APM output when using the generated input.
-
-   Throughout this file, the following naming convention is used:
-    - input signal: the clean signal (e.g., speech),
-    - noise signal: the noise to be summed up to the input signal (e.g., white
-      noise, Gaussian noise),
-    - noisy signal: input + noise.
-   The noise signal may or may not be a function of the clean signal. For
-   instance, white noise is independently generated, whereas reverberation is
-   obtained by convolving the input signal with an impulse response.
+ APM module. Each pair consists of a noisy and a reference signal. The former
+ is used as input for APM, and it is generated by adding noise to a signal.
+ The reference is the expected APM output when using the generated input.
+
+ Throughout this file, the following naming convention is used:
+  - input signal: the clean signal (e.g., speech),
+  - noise signal: the noise to be summed up to the input signal (e.g., white
+    noise, Gaussian noise),
+  - noisy signal: input + noise.
+ The noise signal may or may not be a function of the clean signal. For
+ instance, white noise is independently generated, whereas reverberation is
+ obtained by convolving the input signal with an impulse response.
 """
 
 import logging
 import os
+import sys
+
+try:
+  import scipy.io
+except ImportError:
+  logging.critical('Cannot import the third-party Python package scipy')
+  sys.exit(1)
 
 from . import data_access
-from .signal_processing import SignalProcessingUtils
+from . import signal_processing
 
 class NoiseGenerator(object):
   """Abstract class responsible for the generation of noisy signals.
@@ -52,10 +59,12 @@ class NoiseGenerator(object):
 
   @classmethod
   def register_class(cls, class_to_register):
-    """ Decorator to automatically register the classes that extend
-        NoiseGenerator.
+    """Register an NoiseGenerator implementation.
+
+    Decorator to automatically register the classes that extend NoiseGenerator.
     """
     cls.REGISTERED_CLASSES[class_to_register.NAME] = class_to_register
+    return class_to_register
 
   @property
   def config_names(self):
@@ -76,8 +85,9 @@ class NoiseGenerator(object):
   def generate(
       self, input_signal_filepath, input_noise_cache_path, base_output_path):
     """Generate a set of noisy input and reference audiotrack file pairs.
-       This method initializes an empty set of pairs and calls the _generate()
-       method implemented in a concrete class.
+
+    This method initializes an empty set of pairs and calls the _generate()
+    method implemented in a concrete class.
     """
     self.clear()
     return self._generate(
@@ -96,7 +106,7 @@ class NoiseGenerator(object):
 
   def _add_noise_snr_pairs(self, base_output_path, noisy_mix_filepaths,
                            snr_value_pairs):
-    """ Add noisy-reference signal pairs.
+    """Adds noisy-reference signal pairs.
 
     Args:
       base_output_path: noisy tracks base output path.
@@ -142,9 +152,9 @@ class NoiseGenerator(object):
 # Identity generator.
 @NoiseGenerator.register_class
 class IdentityGenerator(NoiseGenerator):
-  """
-  Generator that adds no noise, therefore both the noisy and the reference
-  signals are the input signal.
+  """Generator that adds no noise.
+
+  Both the noisy and the reference signals are the input signal.
   """
 
   NAME = 'identity'
@@ -165,8 +175,7 @@ class IdentityGenerator(NoiseGenerator):
 
 @NoiseGenerator.register_class
 class WhiteNoiseGenerator(NoiseGenerator):
-  """
-  Additive white noise generator.
+  """Additive white noise generator.
   """
 
   NAME = 'white'
@@ -189,12 +198,16 @@ class WhiteNoiseGenerator(NoiseGenerator):
   def _generate(
       self, input_signal_filepath, input_noise_cache_path, base_output_path):
     # Load the input signal.
-    input_signal = SignalProcessingUtils.load_wav(input_signal_filepath)
-    input_signal = SignalProcessingUtils.normalize(input_signal)
+    input_signal = signal_processing.SignalProcessingUtils.load_wav(
+        input_signal_filepath)
+    input_signal = signal_processing.SignalProcessingUtils.normalize(
+      input_signal)
 
     # Create the noise track.
-    noise_signal = SignalProcessingUtils.generate_white_noise(input_signal)
-    noise_signal = SignalProcessingUtils.normalize(noise_signal)
+    noise_signal = signal_processing.SignalProcessingUtils.generate_white_noise(
+        input_signal)
+    noise_signal = signal_processing.SignalProcessingUtils.normalize(
+        noise_signal)
 
     # Create the noisy mixes (once for each unique SNR value).
     noisy_mix_filepaths = {}
@@ -207,11 +220,12 @@ class WhiteNoiseGenerator(NoiseGenerator):
       # Create and save if not done.
       if not os.path.exists(noisy_signal_filepath):
         # Create noisy signal.
-        noisy_signal = SignalProcessingUtils.mix_signals(
+        noisy_signal = signal_processing.SignalProcessingUtils.mix_signals(
             input_signal, noise_signal, snr)
 
         # Save.
-        SignalProcessingUtils.save_wav(noisy_signal_filepath, noisy_signal)
+        signal_processing.SignalProcessingUtils.save_wav(
+            noisy_signal_filepath, noisy_signal)
 
       # Add file to the collection of mixes.
       noisy_mix_filepaths[snr] = noisy_signal_filepath
@@ -230,8 +244,7 @@ class WhiteNoiseGenerator(NoiseGenerator):
 # TODO(alessiob): remove comment when class implemented.
 # @NoiseGenerator.register_class
 class NarrowBandNoiseGenerator(NoiseGenerator):
-  """
-  Additive narrow-band noise generator.
+  """Additive narrow-band noise generator.
   """
 
   NAME = 'narrow_band'
@@ -247,8 +260,7 @@ class NarrowBandNoiseGenerator(NoiseGenerator):
 
 @NoiseGenerator.register_class
 class EnvironmentalNoiseGenerator(NoiseGenerator):
-  """
-  Additive environmental noise generator.
+  """Additive environmental noise generator.
   """
 
   NAME = 'environmental'
@@ -258,6 +270,7 @@ class EnvironmentalNoiseGenerator(NoiseGenerator):
   _NOISE_TRACKS_PATH = os.path.join(os.getcwd(), 'noise_tracks')
 
   # TODO(alessiob): allow the user to have custom noise tracks.
+  # TODO(alessiob): exploit NoiseGeneratorFactory.GetInstance().
   _NOISE_TRACKS = [
       'city.wav'
   ]
@@ -277,12 +290,26 @@ class EnvironmentalNoiseGenerator(NoiseGenerator):
 
   def _generate(
       self, input_signal_filepath, input_noise_cache_path, base_output_path):
+    """Generate environmental noise.
+
+    For each noise track and pair of SNR values, the following 2 audio tracks
+    are created: the noisy signal and the reference signal. The former is
+    obtained by mixing the (clean) input signal to the corresponding noise
+    track enforcing the target SNR.
+
+    Args:
+      input_signal_filepath: (clean) input signal file path.
+      input_noise_cache_path: path for the cached noise track files.
+      base_output_path: base output path.
+    """
     # Init.
     snr_values = set([snr for pair in self._SNR_VALUE_PAIRS for snr in pair])
 
     # Load the input signal.
-    input_signal = SignalProcessingUtils.load_wav(input_signal_filepath)
-    input_signal = SignalProcessingUtils.normalize(input_signal)
+    input_signal = signal_processing.SignalProcessingUtils.load_wav(
+        input_signal_filepath)
+    input_signal = signal_processing.SignalProcessingUtils.normalize(
+        input_signal)
 
     noisy_mix_filepaths = {}
     for noise_track_filename in self._NOISE_TRACKS:
@@ -294,8 +321,10 @@ class EnvironmentalNoiseGenerator(NoiseGenerator):
         logging.error('cannot find the <%s> noise track', noise_track_filename)
         continue
 
-      noise_signal = SignalProcessingUtils.load_wav(noise_track_filepath)
-      noise_signal = SignalProcessingUtils.normalize(noise_signal)
+      noise_signal = signal_processing.SignalProcessingUtils.load_wav(
+          noise_track_filepath)
+      noise_signal = signal_processing.SignalProcessingUtils.normalize(
+          noise_signal)
 
       # Create the noisy mixes (once for each unique SNR value).
       noisy_mix_filepaths[noise_track_name] = {}
@@ -307,11 +336,12 @@ class EnvironmentalNoiseGenerator(NoiseGenerator):
         # Create and save if not done.
         if not os.path.exists(noisy_signal_filepath):
           # Create noisy signal.
-          noisy_signal = SignalProcessingUtils.mix_signals(
+          noisy_signal = signal_processing.SignalProcessingUtils.mix_signals(
               input_signal, noise_signal, snr)
 
           # Save.
-          SignalProcessingUtils.save_wav(noisy_signal_filepath, noisy_signal)
+          signal_processing.SignalProcessingUtils.save_wav(
+              noisy_signal_filepath, noisy_signal)
 
         # Add file to the collection of mixes.
         noisy_mix_filepaths[noise_track_name][snr] = noisy_signal_filepath
@@ -321,19 +351,128 @@ class EnvironmentalNoiseGenerator(NoiseGenerator):
         base_output_path, noisy_mix_filepaths, self._SNR_VALUE_PAIRS)
 
 
-# TODO(alessiob): remove comment when class implemented.
-# @NoiseGenerator.register_class
+@NoiseGenerator.register_class
 class EchoNoiseGenerator(NoiseGenerator):
-  """
-  Echo noise generator.
+  """Echo noise generator.
   """
 
   NAME = 'echo'
 
-  def __init__(self):
+  _IMPULSE_RESPONSES = {
+      'lecture': 'air_binaural_lecture_0_0_1.mat',  # Long echo.
+      'booth': 'air_binaural_booth_0_0_1.mat',  # Short echo.
+  }
+  _MAX_IMPULSE_RESPONSE_LENGTH = None
+
+  # Each pair indicates the clean vs. noisy and reference vs. noisy SNRs.
+  # The reference (second value of each pair) always has a lower amount of noise
+  # - i.e., the SNR is 5 dB higher.
+  _SNR_VALUE_PAIRS = [
+      [3, 8],  # Smallest noise.
+      [-3, 2],  # Largest noise.
+  ]
+
+  _NOISE_TRACK_FILENAME_TEMPLATE = '{0}.wav'
+  _NOISY_SIGNAL_FILENAME_TEMPLATE = '{0}_{1:d}_SNR.wav'
+
+  def __init__(self, aechen_ir_database_path):
     NoiseGenerator.__init__(self)
+    self._aechen_ir_database_path = aechen_ir_database_path
 
   def _generate(
       self, input_signal_filepath, input_noise_cache_path, base_output_path):
-    # TODO(alessiob): implement.
-    pass
+    """Generates echo noise.
+
+    For each impulse response, one noise track is created. For each impulse
+    response and pair of SNR values, the following 2 audio tracks are
+    created: the noisy signal and the reference signal. The former is
+    obtained by mixing the (clean) input signal to the corresponding noise
+    track enforcing the target SNR.
+
+    Args:
+      input_signal_filepath: (clean) input signal file path.
+      input_noise_cache_path: path for the cached noise track files.
+      base_output_path: base output path.
+    """
+    # Init.
+    snr_values = set([snr for pair in self._SNR_VALUE_PAIRS for snr in pair])
+
+    # Load the input signal.
+    input_signal = signal_processing.SignalProcessingUtils.load_wav(
+        input_signal_filepath)
+
+    noisy_mix_filepaths = {}
+    for impulse_response_name in self._IMPULSE_RESPONSES:
+      noise_track_filename = self._NOISE_TRACK_FILENAME_TEMPLATE.format(
+          impulse_response_name)
+      noise_track_filepath = os.path.join(
+          input_noise_cache_path, noise_track_filename)
+      noise_signal = None
+      try:
+        # Load noise track.
+        noise_signal = signal_processing.SignalProcessingUtils.load_wav(
+            noise_track_filepath)
+      except IOError:  # File not found.
+        # Generate noise track by applying the impulse response.
+        impulse_response_filepath = os.path.join(
+            self._aechen_ir_database_path,
+            self._IMPULSE_RESPONSES[impulse_response_name])
+        noise_signal = self._generate_noise_track(
+            noise_track_filepath, input_signal, impulse_response_filepath)
+      assert noise_signal is not None
+
+      # Create the noisy mixes (once for each unique SNR value).
+      noisy_mix_filepaths[impulse_response_name] = {}
+      for snr in snr_values:
+        noisy_signal_filepath = os.path.join(
+            input_noise_cache_path,
+            self._NOISY_SIGNAL_FILENAME_TEMPLATE.format(
+                impulse_response_name, snr))
+
+        # Create and save if not done.
+        if not os.path.exists(noisy_signal_filepath):
+          # Create noisy signal.
+          noisy_signal = signal_processing.SignalProcessingUtils.mix_signals(
+              input_signal, noise_signal, snr, bln_pad_shortest=True)
+
+          # Save.
+          signal_processing.SignalProcessingUtils.save_wav(
+              noisy_signal_filepath, noisy_signal)
+
+        # Add file to the collection of mixes.
+        noisy_mix_filepaths[impulse_response_name][snr] = noisy_signal_filepath
+
+    # Add all the noise-SNR pairs.
+    self._add_noise_snr_pairs(base_output_path, noisy_mix_filepaths,
+                              self._SNR_VALUE_PAIRS)
+
+  def _generate_noise_track(self, noise_track_filepath, input_signal,
+                            impulse_response_filepath):
+    """Generates noise track.
+
+    Generate a signal by convolving input_signal with the impulse response in
+    impulse_response_filepath; then save to noise_track_filepath.
+
+    Args:
+      noise_track_filepath: output file path for the noise track.
+      input_signal: (clean) input signal samples.
+      impulse_response_filepath: impulse response file path.
+    """
+    # Load impulse response.
+    data = scipy.io.loadmat(impulse_response_filepath)
+    impulse_response = data['h_air'].flatten()
+    if self._MAX_IMPULSE_RESPONSE_LENGTH is not None:
+      logging.info('truncating impulse response from %d to %d samples',
+                   len(impulse_response), self._MAX_IMPULSE_RESPONSE_LENGTH)
+      impulse_response = impulse_response[:self._MAX_IMPULSE_RESPONSE_LENGTH]
+
+    # Apply impulse response.
+    processed_signal = (
+        signal_processing.SignalProcessingUtils.apply_impulse_response(
+            input_signal, impulse_response))
+
+    # Save.
+    signal_processing.SignalProcessingUtils.save_wav(
+        noise_track_filepath, processed_signal)
+
+    return processed_signal