fix: properly clear all internal buffers to prevent audio artifacts after seeking

[rosostolato]

Problem

When reusing SoundTouch instances after seeking (calling clear()), audio artifacts (buzz, pops, crackling) would persist for 1-2 seconds. This made it impractical to reuse instances, forcing developers to recreate SoundTouch objects on every seek, which causes GC pressure in real-time audio applications.

Root Cause

The clear() method had three issues:

1. FifoSampleBuffer.clear() only reset pointers, not the actual Float32Array data. Stale audio samples remained in memory and bled into subsequent processing.

2. RateTransposer had no clear() method. The interpolation state (prevSampleL, prevSampleR, slopeCount) persisted, causing discontinuities when processing resumed.

3. Stretch.clearMidBuffer() had two problems:

   • Only cleared midBuffer when midBufferDirty was true
   • Never cleared refMidBuffer (correlation reference buffer), causing the WSOLA algorithm to find incorrect overlap positions

Solution

FifoSampleBuffer.js

clear() {
  this._vector.fill(0);  // Zero actual data, not just pointers
  this._position = 0;
  this._frameCount = 0;
}

RateTransposer.js

clear() {
  super.clear();
  this.reset();  // Reset interpolation state
}

Stretch.js

clearMidBuffer() {
  this.midBufferDirty = false;
  this.midBuffer = null;  // Trigger proper reinitialization in process()
  if (this.refMidBuffer) {
    this.refMidBuffer.fill(0);  // Clear correlation buffer
  }
  this.skipFract = 0;  // Reset skip fraction
}

Why midBuffer = null instead of fill(0)?

Setting midBuffer = null triggers the initialization block in process():

if (this.midBuffer === null) {
  this.midBuffer = new Float32Array(this.overlapLength * 2);
  this._inputBuffer.receiveSamples(this.midBuffer, this.overlapLength);
}

This ensures midBuffer is populated with actual audio samples from the new position, rather than zeros. When midBuffer contains zeros, the cross-correlation calculations produce arbitrary results, leading to incorrect overlap positions and audio artifacts.

Testing

Tested in a multi-track audio player with pitch shifting:

• Play from start with pitch shift enabled ✓
• Seek to random positions ✓
• Rapid seeking back and forth ✓
• Persistent buzz/artifacts eliminated ✓
• Brief warmup period (~1 second) still occurs due to WSOLA algorithm needing time to find optimal overlap positions - this is expected behavior

Breaking Changes

None. This is a bug fix that makes clear() work as expected.

[cutterbl]

@rosostolato Thanks for your PR. And thank you for digging in to work it out. Great explanation of actual vs expected behavior, and cause and solution. I will review everything over the weekend, and we'll see if we can't get this in play. (no pun intended ;) )

[rosostolato]

This was all claude's help but worked to me

[rosostolato]

There's another bug still going on that makes the audio pop when seeking and playing but I can't find what it is. Idk if it's in the lib or my implementation