README.md

Data Prep

Data preparation can be done using the script provided in data_prep/make_tfrecords.py, which accepts

1. .csv manifest with the following structure

   files,labels
   "path to wav/flac file","label_string"

   Wav/flac files should be of desired sample rate and channels. (For speechcommands, 16000 kHz, mono). The label string comprises of list of label(s) delimited by delimiter. For eg, for a multilabel dataset it could be "up,down" tags, the delimiter being ,
2. .json label map which maps individual labels to integer values

   {
     "backward": 0, 
     "bed": 1, 
     "bird": 2,
      .....
   }
   

3. The following command can then be used to create the tfrecords

   python make_tfrecords.py --manifest_path train.csv --labels_map lbl_map.json --output_dir $BASE_OUTPUT_DIR --split_name train --multiproc_threads 6 --files_per_record 2048 --desired_duration 1. --clip_larger_files --compression "ZLIB"

   More options can be found in the script. The above command was used to generate speechcommands v2 tfrecords
4. Once this is done, final step is to make .csv files with the following structure

   files
   $BASE_DIR/$SPLIT_NAME/file_00000-00020_bytes_compressed.tfrec
   

   which are used in model configs such as recipes/speechcommands/configs/resnet18.py