How to write a wav to a tfrecord and then read it back

Question

I'm trying to write an encoded wav to a tfrecord and then read it back. I know I can write the wav as a normal tensor, but am trying to save space.

I'd like to do something like the following, but am unsure how to fill in the ellipses. In particular, I don't know if I should save as an int64 feature or a bytes feature.

def wav_feature(wav):
    value = tf.audio.encode_wav(wav, 44100)
    return tf.train.Feature(...)

example = tf.train.Example(features=tf.train.Features(feature={
    'foo': wav_feature(wav),
}))

with tf.io.TFRecordWriter(outpath) as writer:
    writer.write(example.SerializeToString())

# In parser

features = tf.io.parse_single_example(
            serialized=proto,
            features={'foo': tf.io.FixedLenFeature([], ...)})

decoded, sr = tf.audio.decode_wav(features['foo'])

Answer 1

It looks like encode_wav returns a string tensor, so using a bytes feature is best:

def _bytes_feature(value):                                                      
  """Returns a bytes_list from a string / byte."""                              
  if isinstance(value, type(tf.constant(0))):                                   
    value = value.numpy() # BytesList won't unpack a string from an EagerTensor.
  return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))         

# Convert to a string tensor.
wav_encoded = tf.audio.encode_wav(wav, 44100)

feature = {'foo': _bytes_feature(wav_encoded)}         
example = tf.train.Example(features=tf.train.Features(feature=feature))      

Then, in the parser:

features = tf.io.parse_single_example(
        example.SerializeToString(),                 
        features={'foo': tf.io.FixedLenFeature([], tf.string)})               
# wav_encoded will be a string tensor. 
wav_encoded = features['foo']

Definition of _bytes_feature is here.