Set fix input frame size to simplify code, added benchmark script to estimate inference efficiency

benchmark/benchmark.py

@@ -0,0 +1,50 @@
+# Copyright 2022 David Scripka. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Imports
+import openwakeword
+import numpy as np
+from pathlib import Path
+from collections import defaultdict
+
+# Define benchmark to assess inference speed of models at different audio chunk sizes
+# Smaller chunk sizes may increase model performance, at the cost of inference efficiency
+def run_benchmark():
+    # Load models
+    model_paths = [str(i) for i in Path("openwakeword/resources/models").glob("*.onnx") \
+                   if "embedding" not in str(i) and "melspectrogram" not in str(i)]
+    M = openwakeword.Model(
+        wakeword_model_paths=model_paths,
+        input_sizes=[16]
+    )
+
+    # Create random data to use for benchmarking
+    clip = np.random.random(16000*10).astype(np.float32)
+
+    # Run the benchmark
+    step_size = 1280
+    preprocessing_times = []
+    model_times = defaultdict(list)
+    for i in range(0, clip.shape[0]-step_size, step_size):
+        pred, timing_dict = M.predict(clip[i:i+step_size], timing=True)
+        preprocessing_times.append(timing_dict["preprocessor"])
+        for mdl_name in M.models.keys():
+            model_times[mdl_name].append(timing_dict["models"][mdl_name])
+
+    print(f"Average of {np.mean(preprocessing_times)} for audio preprocessing with a frame size of {step_size/16000} seconds")
+    for mdl_name in M.models.keys():
+        print(f"Average of {np.mean(model_times[mdl_name])} for model \"{mdl_name}\"", "\n\n")
+
+if __name__ == "__main__":
+   run_benchmark()

openwakeword/detect.py

@@ -18,6 +18,7 @@ import numpy as np
 import os
 from openwakeword.utils import AudioFeatures
 from typing import List
+import time
 
 class Model():
     def __init__(self, wakeword_model_paths: List[str], input_sizes: List[int], **kwargs):
@@ -36,16 +37,40 @@ class Model():
         # Create AudioFeatures object
         self.preprocessor = AudioFeatures(**kwargs)
 
-    def predict(self, x):
+    def predict(self, x, timing=False):
         """Predict with all of the wakeword models on the input audio frames"""
+        # Get audio features
+        if timing:
+            timing_dict = {}
+            timing_dict["models"] = {}
+            feature_start = time.time()
+        
         self.preprocessor(x)
+
+        if timing:
+            feature_end = time.time()
+            timing_dict["preprocessor"] = feature_end - feature_start
+
+        # Get predictions from model(s)
         predictions = {}
         for mdl in self.models.keys():
             input_name = self.models[mdl].get_inputs()[0].name
+
+            if timing:
+                model_start = time.time()
+
             predictions[mdl] = self.models[mdl].run(
                                     None,
                                     {input_name: self.preprocessor.get_features(self.model_inputs[mdl])}
                                 )[0][0][0]
-        return predictions
+
+            if timing:
+                model_end = time.time()
+                timing_dict["models"][mdl] = model_end - model_start
+
+        if timing:
+            return predictions, timing_dict
+        else:
+            return predictions
 
         

openwakeword/utils.py

@@ -12,17 +12,19 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+
 # Imports
 import os
 import onnxruntime as ort
 import numpy as np
+import pathlib
 from collections import deque
 
 # Base class for computing audio features using Google's speech_embedding model (https://tfhub.dev/google/speech_embedding/1)
 class AudioFeatures():
     def __init__(self,
-            melspec_onnx_model_path=os.path.join("openwakeword", "resources", "models", "melspectrogram.onnx"),
-            embedding_onnx_model_path=os.path.join("openwakeword", "resources", "models", "embedding_model.onnx"),
+            melspec_onnx_model_path=os.path.join(pathlib.Path(__file__).parent.resolve(), "resources", "models", "melspectrogram.onnx"),
+            embedding_onnx_model_path=os.path.join(pathlib.Path(__file__).parent.resolve(), "resources", "models", "embedding_model.onnx"),
             sr=16000,
             ncpu=1
         ):
@@ -80,22 +82,17 @@ class AudioFeatures():
             self.melspectrogram_buffer = self.melspectrogram_buffer[-self.melspectrogram_max_len:, :]
     
     def _streaming_features(self, x):
-        assert len(x) == 1280*3
+        if len(x) != 1280:
+            raise ValueError(f"You must provide input samples in frames of 1280 samples @ 1600khz. Received a frame of {len(x)} samples.")
         self._streaming_melspectrogram(x)
-        if self.melspectrogram_buffer.shape[0] < 76:
-            pass
-        for i in range(-3,0,1):
-            start = -76+8*i
-            end = start + 76
-            x = self.melspectrogram_buffer[start:end].astype(np.float32)[None,:,:,None]
-            if x.shape[1] != 76:
-                continue
+        x = self.melspectrogram_buffer[-76:].astype(np.float32)[None,:,:,None]
+        if x.shape[1] == 76:
             self.feature_buffer = np.vstack((self.feature_buffer, self.embedding_model.run(None, {'input_1': x})[0].squeeze()))
             
         if self.feature_buffer.shape[0] > self.feature_buffer_max_len:
             self.feature_buffer = self.feature_buffer[-self.feature_buffer_max_len:, :]
 
-    def get_features(self, x, n_feature_frames=16):
+    def get_features(self, n_feature_frames=16):
         return self.feature_buffer[-n_feature_frames:, :][None,].astype(np.float32)
             
     def __call__(self, x):

tests/test_models.py

@@ -11,6 +11,20 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+# Copyright 2022 David Scripka. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 
 # Imports
 import openwakeword
@@ -30,7 +44,7 @@ class TestModels:
             input_sizes=[16]
         )
 
-        step_size = 1280*3
+        step_size = 1280
         predictions = []
         for i in range(0, hey_jane_clip.shape[0]-step_size, step_size):
             predictions.append(model.predict(hey_jane_clip[i:i+step_size])["hey_jane"])