SpeakerKit: The last turn is attributed to the wrong speaker

[robertobrega]

File: Sources/SpeakerKit/Pyannote/SpeakerSegmenterModel.swift
(predict(...), in the per-chunk worker, just before
AudioProcessor.padOrTrimAudio).

Symptom. The last ~10 s of a recording
are reproducibly split off into a phantom extra speaker, even for a
single-voice recording on one channel.

Root cause. The segmenter's CoreML model has a fixed [480000]
(30 s) input, so the final short chunk is zero-padded to 30 s
(padOrTrimAudio). The embedder then runs its preprocessor over the
whole padded 30 s chunk (SpeakerEmbedderModel.processChunk,
SpeakerEmbedderPreprocessorInput(waveforms:)). A global op in that
frontend lets the trailing zeros attenuate the embeddings of the
chunk's real-audio windows — their L2 norm roughly halves
(~3.7 → ~2.1, and as low as ~1.1) even when the window's speaker mask
lies entirely inside real audio. Because the tail windows of a recording are only ever
emitted from this padded final chunk, the last paragraph gets the
phantom speaker almost every time.

Fix. Fill the final short chunk's tail by mirror-reflecting the
real audio (alternating reversed/forward copies) instead of
zero-padding. This keeps the preprocessor's global statistics in a
speech regime, so the real-window embeddings keep their normal norm and
stay in the correct cluster. waveformLength is still computed from the
real sample count, so bounded() and downstream turn-clamping are
unchanged — only the model's input padding content differs.

The diff
Two changes, both in SpeakerSegmenterModel.swift:

1. New reflect-pad block:

var srcWaveform = chunk.waveform
if srcWaveform.count < maxChunkLength, !srcWaveform.isEmpty {
var ext = srcWaveform
var forward = false // alternate reversed / forward copies = mirror reflect
while ext.count < maxChunkLength {
let block = forward ? srcWaveform : Array(srcWaveform.reversed())
ext.append(contentsOf: block.prefix(maxChunkLength - ext.count))
forward.toggle()
}
srcWaveform = ext
}

2. One-line redirect — feed the reflected buffer into the model instead of the raw chunk:

guard let audioSamples = AudioProcessor.padOrTrimAudio(
fromArray: chunk.waveform,
fromArray: srcWaveform,
startAt: 0,
toLength: maxChunkLength
) else {