Add read_wav and list_wav_info support

[yongtang]

This PR adds read_wav and list_wav_info support so that it is possible to read wav file into a tensor.
Since WAV file is splittable, read_wav could take a start and count parameter so that only a slice of the wav file is read.

Read a chunk of the wav file is related to #49 (comment)

This is also part of the effort to rework on Dataset to move to primitive ops. See See #382 and #366 for related discussions.

Signed-off-by: Yong Tang yong.tang.github@outlook.com

[yongtang]

/cc @faroit for reading a chunk of WAV file. You can use read_wav and pass start=start, count=count to only read a chunk of the samples.

Note start stand for the starting index of the sample.

The total number of samples in WAV could be obtained through:
spec, rate = list_wav_info(filename) where spec is a TensorSpec which gives you the shape, and dtype of the tensor. Shape is [n, c] where n is the total number of samples and c is the channels.

The dtype could be int8 or int16 for 8-bit and 16-bit.

I tried to see if I could also add 24bit, though I need a sample 24bit WAV file to see the actual memory layout (to figure out how to fit 24bit into an int32)

[faroit]

@yongtang This is great!

count sounds a bit wobbly, I would propose to use either

• frames or num_frames (samples*channel)
• length (duration in samples)

or to set the end sample in samples

• stop
• end

[faroit]

@yongtang before I try it: is the wavdataset form this PR compatible to the tf 2.x beta?

[faroit]

Looks great, thanks