Performance & memory

SoundFile is designed to be efficient for common I/O tasks, but audio can be large. This page outlines the main levers you have.

One-shot vs streaming

• One-shot (sf_read) loads the entire signal into memory.
• Streaming (SoundFile::blocks) lets you process incrementally with bounded memory.

If you are processing long recordings, prefer streaming.

sf_read block size

sf_read() reads data in a loop using blocksize frames per iteration, but still allocates a single output buffer for the full requested range.

blocksize affects performance and temporary working set (not the final result).

[$x, $info] = sf_read('input.wav', blocksize: 16384);

Resampling: chunked mode

The default resampler path is chunked progressive processing (chunkSize is non-null). This avoids repeatedly allocating output arrays per chunk and is safer for large signals.

$y = sf_resample($x, 44100, 16000, chunkSize: 8192);

For small signals, one-shot chunkSize: null is often fine and can be simpler.

DTypes and conversion cost

Writing converts the NDArray dtype to match the chosen SampleFormat.

If you already know what you’ll write, you can pre-convert once in your pipeline to avoid repeated conversions:

use PhpMlKit\NDArray\DType;
use PhpMlKit\SoundFile\Enums\SampleFormat;

$x = $x->astype(SampleFormat::Pcm16->toDtype()); // Int16

Practical tips

• Prefer SoundFile::blocks() + incremental processing for long audio.
• Keep shapes consistent ([frames, channels]) to avoid repeated reshaping.
• When resampling, pick quality intentionally:
  • Best for offline processing
  • Fastest for throughput
  • Linear for cheap previews