sound_rs v0.2.2

07 · Guides

NumPy and arrays

Two directions: arrays in for playback, arrays out for analysis. Both use float32 samples in the range −1.0 to 1.0.

Shapes

CallShapeMeaning
play_array(mono, rate)(frames,)1D, one channel
play_array(stereo, rate)(2, frames)2D, channel first
load(path)(frames, channels)2D, frames first
The two directions have opposite layoutsplay_array takes (channels, frames) because that is how a matrix of channels is usually built; load returns (frames, channels) so audio[i] is one frame and audio[:, 0] is the left channel. Round-tripping needs a transpose: sound_rs.play_array(audio.T, rate).

Playing arrays

import numpy as np
import sound_rs

rate = 48000
seconds = 3
t = np.arange(rate * seconds) / rate

left = (0.2 * np.sin(2 * np.pi * 220 * t)).astype(np.float32)
right = (0.2 * np.sin(2 * np.pi * 330 * t)).astype(np.float32)

sound_rs.play_array(np.stack([left, right]), rate)

The array is converted to the device layout once, before playback starts: channel folding, then one FFT resample if the rates differ. Playing a 48 kHz array on a 44.1 kHz device costs a few milliseconds and sounds correct — a naive nearest-neighbour resampler would alias.

Decoding to arrays

import sound_rs

audio, rate = sound_rs.load("track.mp3")
print(audio.shape, audio.dtype, rate)        # (9437184, 2) float32 44100
print(audio[:, 0].max())                     # peak of the left channel

load returns the whole file. Ten minutes of stereo float32 is 211 MB, so for long files prefer streaming playback or decode one channel at a time from your own pipeline.

Batch decoding

import glob
import sound_rs

paths = sorted(glob.glob("dataset/*.flac"))
clips = sound_rs.load_many(paths)             # uses every core
print(len(clips), clips[0][0].shape)
Threadsload_many(paths, threads=2) limits the pool. The default is one worker per core. Ordering matches the input list, and the first failing file raises.

Pitfalls