(1, n_samples) is not accepted as a mono waveform

Description Librosa typically operates on one-dimensional, one-channel signals whose shape is (n_samples,). In some cases, stereo signals are accepted in “time-major” format, e.g. of shape (2, n_samples). But the analogous single-channel signal, whose shape is (1, n_samples), is rejected – at least by numpy.resample()

To Reproduce

import librosa
import numpy as np

# Conventional mono
mono_signal = np.random.randn(1000)
resampled_mono = librosa.resample(mono_signal, 1000, 2000)

# Conventional stereo
stereo_signal = np.random.randn(2, 1000)
resampled_stereo = librosa.resample(stereo_signal, 1000, 2000)

# Stereo-like single channel
onechan_signal = np.random.randn(1, 1000)
resampled_onechan = librosa.resample(onechan_signal, 1000, 2000)

Expected behavior resampled_mono == resampled_onechan[0]

Actual behavior

---------------------------------------------------------------------------
ParameterError                            Traceback (most recent call last)
<ipython-input-7-3776517cab8b> in <module>()
      1 onechan_signal = np.random.randn(1, 1000)
----> 2 resampled_onechan = librosa.resample(onechan_signal, 1000, 2000)
      3 print(resampled_onechan.shape)

1 frames
/usr/local/lib/python3.7/dist-packages/librosa/util/utils.py in valid_audio(y, mono)
    304     elif y.ndim == 2 and y.shape[0] < 2:
    305         raise ParameterError(
--> 306             "Mono data must have shape (samples,). " "Received shape={}".format(y.shape)
    307         )
    308 

ParameterError: Mono data must have shape (samples,). Received shape=(1, 1000)

Software versions

Linux-5.4.104+-x86_64-with-Ubuntu-18.04-bionic
Python 3.7.10 (default, May  3 2021, 02:48:31) 
[GCC 7.5.0]
NumPy 1.19.5
SciPy 1.4.1
librosa 0.8.1
INSTALLED VERSIONS
------------------
python: 3.7.10 (default, May  3 2021, 02:48:31) 
[GCC 7.5.0]

librosa: 0.8.1

audioread: 2.1.9
numpy: 1.19.5
scipy: 1.4.1
sklearn: 0.22.2.post1
joblib: 1.0.1
decorator: 4.4.2
soundfile: 0.10.3
resampy: 0.2.2
numba: 0.51.2

numpydoc: None
sphinx: 1.8.5
sphinx_rtd_theme: None
sphinxcontrib.versioning: None
sphinx-gallery: None
pytest: 3.6.4
pytest-mpl: None
pytest-cov: None
matplotlib: 3.2.2
presets: None

[bmcfee]

Interesting. I wasn’t suggesting that we’d validate (via valid_audio) empty arrays with no time samples, like (1, 2, 3, 4, 0). Only those with nonzero time samples and zero channels (whatever non-last dimension you assign to channel): like, (1, 2, 3, 0, 256).

So an empty dimension is okay as long as it’s not in the time position? I guess the shape check would then be y.shape[-1] != 0?

I also agree on the padding problem, which would work in zero mode but not in reflect mode. So in my opinion, np.array([]) should not pass valid_audio because we can’t guarantee that a downstream STFT with pad_mode=reflect (the default!) will also pass NumPy’s np.pad.

Or, we could switch the default padding behavior to constant/zeros. It’s going into a major revision, after all. 😁

Back to my original question, i think it would be nice if the following assertions could hold:

x = librosa.mfcc(np.zeros((2, 1024)))
y = librosa.mfcc(np.zeros((0, 1024)))

assert x.dtype == y.dtype
assert x.shape[1:] == y.shape[1:]

Yeah, I think I agree with that. The lingering question for me is whether there are legitimate use cases for a length-0 signal. I can’t think of one right now, but there might be some weird edge cases that pop up, eg, in the cqt module where it might be nice to have a graceful fallback when the signal is sliced too thin.

[lostanlen]

Valid audio should check finitude and dtype Shape can be anything If the trailing dimension is not the largest, throw a warning

sounds good to me