babblefish

File structure

app/       - TF training scripts
crying/    - raw crying sample WAVs from source device
white_noise/   - raw white_noise sample WAVs from source device
data/
  crying/  - normalized WAVs for crying label
  white_noise/  - normalized WAVs for white_noise label
  room_empty/  - normalized WAVs for room_empty label

Getting data

Record on rpi with arecord, rsync it to raw/ directory here.

Batchify

Find the valid recordings and use sox to batch up

Cleaning up data

The process-raw-labeled-wav.sh script will:

1. Take raw samples in crying/ and white_noise/ folders
2. Batch them up and splice them evenly into 5-second segments
3. Amplify 45dB, downsample to 22050
4. Store in data/<LABEL> directory.

Learning

Trying to run the training app directly pointed at this app.

Getting:

InvalidArgumentError (see above for traceback): Data too short when trying to read value
         [[Node: DecodeWav = DecodeWav[desired_channels=1, desired_samples=192000, _device="/job:localhost/replica:0/task:0/device:CPU:0"](ReadFile)]]

Seems like the WAV file partitions might be too short? Need exactly SAMPLE_RATE*DURATION_MS samples in WAV files!

Training

python app/train.py --data_url= --data_dir=/Users/andrewhao/workspace/babblefish/data --wanted_words=white_noise,crying --sample_rate=48000 --clip_duration_ms=5000 --start_checkpoint=/tmp/speech_commands_train/conv.ckpt-1300 --how_many_training_steps=1400,300

Freeze

python app/freeze.py --start_checkpoint=/tmp/speech_commands_train/conv.ckpt-1700 --output_file=./graph.pb --clip_duration_ms=5000 --sample_rate=48000 --how_many_training_steps=1400,300 --wanted_words=white_noise,crying --data_dir=/Users/andrewhao/workspace/babblefish/data

Label

python app/label_wav.py --graph=./graph.pb --labels=./conv_labels.txt --wav=data/crying/crying00021.wav

MP3<>WAV utilities

MP3 -> WAV

ffmpeg -i crying/2_years/2019-09-11\ 04:53:03.wav.mp3 -acodec pcm_s16le -r 22050 test.wav