go-speak

Speech recognition challenge with TensorFlow and Go.

Your challenge, should you choose to accept it is to use the TensorFlow Go package to import a trained model and apply it to a wav file.

Prerequisites:

• Docker
• optional: Go - we will have go running inside the docker container

*We will NOT be installing TensorFlow locally in this exercise but use a docker container instead.

Setup

In your terminal (from inside the root directory of this repository):

docker-compose run speech_recognition /bin/bash  

Your terminal should now be running bash from inside the docker container we've setup for this task.

Train the model

To train our model we will be following the first part of the TensorFlow tutorial for audio recognition.

The Python scripts are already inside the container. Run:

cd /go/src/github.com/tensorflow/tensorflow/tensorflow/examples/speech_commands
python train.py \
--data_dir=$APP_DIR/tmp/speech_dataset/ \
--summaries_dir=$APP_DIR/tmp/retrain_logs \
--train_dir=$APP_DIR/tmp/speech_commands_train

If the training stopped in the middle you can check your speech_commands_train directory for the latest checkpoint and rerun the train.py script with --start_checkpoint=$APP_DIR/tmp/speech_commands_train/conv.ckpt-[last-checkpoint-number]

####Save the model (graph)

python freeze.py \
--start_checkpoint=$APP_DIR/tmp/speech_commands_train/conv.ckpt-1000 \
--output_file=$APP_DIR/tmp/speech_recognition_graph.pb

####Run the (failing) tests

cd $APP_DIR
go test

Implement the code until tests pass

####Run your code to see the results

go run main.go