road2018 / aecnn-rpi

Real-time Python framework for speech enhancement using DNNs

Real-time Python framework for speech enhancement using AECNNs

Fotios Drakopoulos, Deepak Baby and Sarah Verhulst. Real-time audio processing on a Raspberry Pi using deep neural networks, 23rd International Congress on Acoustics, 9 to 13 September 2019 in Aachen, Germany.

This work received funding from the European Research Council (ERC) under the Horizon 2020 Research and Innovation Programme (grant agreement No 678120 RobSpear)

---

The Keras framework for the implementation of the AECNN models is adapted from here. The necessary scripts can be found in the AECNN folder.

Preparation of data

1. The dataset of Valentini et al. was used for the training and testing of the models. The download_dataset.sh script can be used to download the dataset (it requires sox to downsample the data to 16 kHz).

   ./download_dataset.sh

2. The data need to be segmented in the training and testing sets depending on the desired window size (the input/output size of the AECNN model). For this, the variable opts ['window_size'] needs to be defined in the prepare_data.py script.

   python prepare_data.py

Training the model

python run_aecnn.py

The opts variable needs to be edited to modify the architecture configurations of the AECNN model.

Running the trained model in real-time

python audio_processing.py -m model_directory -n model_input_size -f keras

The directory of the trained model needs to be defined with the -m argument as well as the input/output size of the model with the -n argument. Keras or Tensorflow can be used as the frontend (-f) and 0% or 50% overlap (-o) or frame buffering (-b) can be applied.

---

Benchmarking a model

A trained model can be benchmarked within the current framework in terms of execution time, in order to get the time constrains for different settings. A .wav file needs to be provided (-i) and the (parent) directory containing the model folder(s) needs to be defined (-d). This way, multiple models can be benchmarked with this script.

python model_benchmark.py -i wav_file -d parent_directory -f keras -it 1

Converting a Keras model to protobuf format (.pb)

A trained model can be converted to protobuf format for inference in Tensorflow:

python tensorflow_converter.py -m model_directory

Measure the complexity of a Keras model

The number of parameters and floating-point operations of a trained model can be computed:

python measure_complexity.py -m model_directory