Librosa Trim vs SOX Trim

Description

Help with Librosa trim and understanding it.

Steps/Code to Reproduce

import librosa

trimmed, _ = librosa.effects.trim(signal)
plot_waveform([trimmed[:1000]])
print(librosa.get_duration(signal), librosa.get_duration(trimmed))

Expected Results

Expecting the first ~400 samples to be trimmed.

Actual Results

6.195238095238095 6.195238095238095

Versions

Darwin-17.5.0-x86_64-i386-64bit Python 3.6.5 (default, May 15 2018, 11:11:58) [GCC 4.2.1 Compatible Apple LLVM 9.1.0 (clang-902.0.39.1)] NumPy 1.14.3 SciPy 1.1.0 librosa 0.6.0

[bmcfee]

This is a bit difficult to diagnose without access to the original signal.

For clarification, what trim does is clip off frames at the beginning or end of the clip where the rmse is significantly below the reference value. By default, the reference value is the max over all frames, and the threshold for “significant” is 60dB. The framing parameters are aligned to the library defaults (n=2048, hop=512). In your case, it looks like there’s enough energy in the first frame (2048 samples) to keep it above the threshold. If you want to narrow the analysis, you can drop the frame length (say, 512) and hop length (say 128) and that should give your expected behavior.

[PetrochukM]

@bmcfee Thank you! That worked well. Appreciate your time.