Attention-based Model for Keywords Spotting

Paper: https://arxiv.org/abs/1803.10916

Data: http://download.tensorflow.org/data/speech_commands_v0.01.tar.gz

Docker

Build container:

./docker/build.sh <container>

Run container:

./docker/run.sh <container> <port>

Stop container:

./docker/stop.sh <container>

Model utilization

Init project module:

./scripts/init_module.sh

Download training data:

./scripts/download_data.sh

Download model checkpoint:

./scripts/download_model.sh

Start training process:

./scripts/train_model.sh

Model inference, this one is configured to process example.wav file. Output probabilities are plotted in example_probs.jpg file:

./scripts/test_model.sh

Results

Here is an example of model probs for example.wav. This audio contains the following spoken text: "Hello, dear Sheila! This message was recorded to you by Marvin himself".