Non-contiguous memory layout in stereo input causes corrupted output

[joris-vaneyghen]

When processing stereo audio input where the NumPy array has a non-C-contiguous memory layout (commonly occurring after loading with librosa), the processed output becomes corrupted or garbled. This appears to be a memory layout compatibility issue between the Python/NumPy interface and the underlying C++ processing code.

The C++ processing code expects arrays with C-contiguous memory layout, but librosa-loaded arrays may have different memory stride patterns or non-contiguous layouts that are incompatible.

Suggested Solutions:

Ideal: The process() method should automatically handle memory layout conversion internally
Alternative: Clearly document this requirement with examples in the documentation

Code to reproduce + FIX

import numpy as np
import librosa
import python_stretch as ps

# Load a stereo audio example from librosa (e.g., 'trumpet', 'brahms',...)
audio, sr = librosa.load(librosa.ex('trumpet', hq=True), sr=None, mono=False)

#fix memory layout issue
audio_fixed = np.ascontiguousarray(audio)

print("audio shape:", audio.shape, "dtype:", audio.dtype)
print("audio_fixed shape:", audio_fixed.shape, "dtype:", audio_fixed.dtype)
print("audio flags:", audio.flags)
print("audio_fixed flags:", audio_fixed.flags)
print("Arrays equal:", np.array_equal(audio, audio_fixed))
print("Memory layout same:", audio.flags['C_CONTIGUOUS'] == audio_fixed.flags['C_CONTIGUOUS'])


# Create a Stretch object
stretch = ps.Signalsmith.Stretch()
# Configure using a preset
stretch.preset(audio.shape[0], sr) # numChannels, sampleRate
# Shift up by one octave
stretch.setTransposeSemitones(12)
# Stretch time
stretch.timeFactor = 0.75

# Process
audio_processed = stretch.process(audio)
audio_fixed_processed = stretch.process(audio_fixed)

# Save and listen
import soundfile as sf
sf.write("audio_original.wav", audio.T, sr)
sf.write("audio_processed.wav", audio_processed.T, sr)
sf.write("audio_fixed_processed.wav", audio_fixed_processed.T, sr)

[gregogiudici]

You're absolutely right!

Right now, the safest fix is to convert your array before processing, as you suggested.

# "samples" is your audio input np.array
if samples.ndim == 1:
     samples = samples[np.newaxis, :]

if (samples.ndim == 2 and samples.shape[0] > 1 and not samples.flags.c_contiguous):
     samples = np.ascontiguousarray(samples)

Audiomentations (python-stretch is used as its default pitch/time library) actually does this check and conversion automatically internally, so users never have to think about it.

I was already planning to update the library soon (I hope to find time to work on it next month): my idea is to handle the memory layout stuff inside the library (i.e., inside the .process() method) so people don’t have to worry about it at anymore.

I will leave this issue open until I have updated the library.
If you have ideas for how to make the docs clearer, or you find other bugs, or you have suggestions to improve the library, let me know!