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 6d2f6e99e21e135210121a70bb0634451ca17ad0..fa6d6ada1cc048c99023f57e90f22dd2da20f550 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
@@ -9,8 +9,10 @@
 import logging
 import os
 
+import scipy.io

[kjellander_webrtc, 2017/03/09 07:49:31]

This is not part of the standard library. Is it possible to pull it down via
DEPS to avoid relying on system-installed libraries?

Please implement a user-friendly error message if the module is missing on
import.

[AleBzk, 2017/03/09 11:31:20]

True. And in another file I also use another 3rd part package (pydub).
If you could point me on a guideline to add 3rd party libs in chromium or
webrtc, I will do that in two separate CLs (one for scipy and one for pydub).
I have done this before on another repository, so I know that I should take care
of licensing stuff.
As a reference I guess I can look at
https://chromium.googlesource.com/chromium/src/third_party/jinja2/+/master
Feel free to contact me privately to advise on this.

As for the user-friendly message, I added a try-except block around import.

[kjellander_webrtc, 2017/03/09 14:05:42]

This is fine. If this never is going to run on any bot, we don't need to DEPS
them in.
Otherwise you'll have to create a mirror of the source code at
https://chromium.googlesource.com/ and pull that in via DEPS.
It can be auto-mirrored from github.com repos, but if it's not there, we might
rely on manually importing to a deps repo, assuming the license of these
libraries allows redistribution in source form. If they have native code as
well, it becomes an even harder problem (I assume these have?).

+
 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.
@@ -45,6 +47,7 @@ class NoiseGenerator(object):
         NoiseGenerator.
     """
     cls.REGISTERED_CLASSES[class_to_register.NAME] = class_to_register
+    return class_to_register

[AleBzk, 2017/03/06 11:29:53]

Bug fix: the class decorator MUST return the class

 
   @property
   def config_names(self):
@@ -175,12 +178,16 @@ 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 = {}
@@ -193,11 +200,12 @@ 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(
             noise_signal, input_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
@@ -245,6 +253,7 @@ class EnvironmentalNoiseGenerator(NoiseGenerator):
       os.path.dirname(os.path.realpath(__file__)), 'noise_tracks')
 
   # TODO(alessiob): allow the user to have custom noise tracks.
+  # TODO(alessiob): exploit NoiseGeneratorFactory.GetInstance().
   _NOISE_TRACKS = [
       'city.wav'
   ]
@@ -271,8 +280,10 @@ class EnvironmentalNoiseGenerator(NoiseGenerator):
     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:
@@ -283,8 +294,10 @@ class EnvironmentalNoiseGenerator(NoiseGenerator):
       if not os.path.exists(noise_track_filepath):
         logging.error('cannot find the <%s> noise track', noise_track_filename)
 
-      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] = {}
@@ -296,11 +309,12 @@ 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(
               noise_signal, input_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
@@ -310,8 +324,7 @@ 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.
@@ -319,10 +332,105 @@ class EchoNoiseGenerator(NoiseGenerator):
 
   NAME = 'echo'
 
-  def __init__(self):
+  _IMPULSE_RESPONSES = {

[kjellander_webrtc, 2017/03/09 07:49:31]

I know nothing about the signal processing involved here, so I suggest you pick
a more suitable reviewer for that. I'm only reviewing style and basic logic from
here on.

[AleBzk, 2017/03/09 11:31:20]

Yup

+      '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.
+  # Since the implementation below only changes the gain of the noise, the
+  # values indicate the noise-to-signal ratio. Therefore a higher value means
+  # larger amount of noise.
+  # 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, -2],  # Largest noise.
+      [-3, -8],
+  ]
+
+  _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

[AleBzk, 2017/03/06 11:29:53]

@kjellander: the same comment I got from you about the POLQA path also applies
to the AIR database path, I also corrected this

 
   def _generate(

[kjellander_webrtc, 2017/03/09 07:49:31]

Add docstring

[AleBzk, 2017/03/09 11:31:19]

Done.

       self, input_signal_filepath, input_noise_cache_path, base_output_path):
-    # TODO(alessiob): implement.
-    pass
+    # 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(
+              noise_signal, input_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):
+    """
+    Generate a signal by convolving input_signal with the impulse response in
+    impulse_response_filepath; then save to noise_track_filepath.

[kjellander_webrtc, 2017/03/09 07:49:31]

Add docstring for arguments

[AleBzk, 2017/03/09 11:31:20]

Done.

+    """
+    # 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