File objects work with librosa.load but not librosa.stream

Description

Loading a WAV file via tf.gfile.GFile works with librosa.load but not librosa.stream. Interestingly the same error appears regardless of if it is a bucket path or a local path.

Steps/Code to Reproduce

import librosa as lr
import tensorflow as tf

path = 'gs://my-bucket/example.wav'

# Loading the entire file works as expected.
with tf.gfile.GFile(path, 'rb') as f:
    lr.load(f, 48000)

# The streaming method crashes.
with tf.gfile.GFile(path, 'rb') as f:
    assert f.seekable() is True
    for waveforms in lr.stream(f, 128, 1024, 512):
        print(waveforms.shape)

Expected Results

No errors.

Actual Results

RuntimeError: Error opening : File contains data in an unknown format.

Versions

import platform; print(platform.platform())...
Linux-4.9.0-9-amd64-x86_64-with-debian-9.9
Python 3.5.3 (default, Sep 27 2018, 17:25:39) 
[GCC 6.3.0 20170516]
NumPy 1.17.0
SciPy 1.3.0
librosa 0.7.0

[carlthome]

And super curiously, this will work fine

import soundfile as sf

with tf.gfile.GFile(path, 'rb') as f:
    for waveform in sf.blocks(f, blocksize=1024):
        pass

so perhaps it’s something with the overlapping blocks.

[bmcfee]

Thanks for tracking this down!

• Require a sample rate to be input into librosa.stream and remove the sf.info call.

I don’t like this option, because it connotes that an automatic sampling rate conversion would be invoked. If you give the wrong sampling rate in the call, there’s no obvious recourse.

• Check if path is a file object and call path.seek(0) after sf.info (hacky).

I agree that this is hacky, but I think it might be the best option.

• Update PySoundFile’s sf.info function to reset the pointer or defensively copy its argument (perhaps unwanted behaviour, but cannot think of reasons why).

That’s a tricky one. Maybe it would be worth discussing with the soundfile devs?