The analysis pipeline

Every rendered or loaded sample goes through the same six-stage pipeline before the exporter turns the results into target-specific preset files. This page walks through what each stage does and what knobs in your config affect it.

flowchart LR
    IN[Sample audio] --> T[Trim]
    T --> E[Envelope]
    E --> P[Pitch check]
    P --> L[Loop detect]
    L --> N[Normalize]
    N --> OUT[Ready for export]

1. Trim

Trims leading and trailing silence. Walks a mono-downmixed RMS envelope from each end and cuts at the first frame whose peak crosses −60 dB.

Trim never cuts into an authored loop. If the input WAV has a smpl chunk with loop start/end (see Honoring authored loops below), trim is constrained to stay outside those bounds.

Off with: analysis.trim: false in the config.

2. Envelope detection

The most subtle stage. Classifies each sample as one of:

  • pluck — no meaningful sustain after the attack
  • sustained — flat sustain
  • sustained_with_modulation — sustained with a periodic (LFO/tremolo) or free-running (filter sweep) timbral movement

The classification decides whether looping is even attempted, and — for VST captures — is what scan-from-probe uses to auto-pick a profile.

How it works

  1. Compute the RMS envelope on a 512-sample hop.
  2. Find the peak RMS frame. Walk backward looking for a plateau (frames within 90% of peak); if there’s a plateau of ≥ 2 s, use its start as sustain_start. Otherwise the peak is the sustain start.
  3. Walk forward from the peak until RMS drops below 40% of peak. That’s sustain_end, extended forward if a stable plateau (coefficient of variation ≤ 0.25) continues for ≥ 0.5 s beyond the drop.
  4. Pluck check. Measure RMS in a window near note-off (for VST/CLAP where note-off is known) or in [peak + 1.5s, peak + 3.0s] clamped to [sustain_start, sustain_end] (for library samples). If that RMS is < 10% of peak, it’s a pluck.
  5. Modulation check. If the RMS variance over the sustain is ≥ 40% of peak, run autocorrelation and look for periodicity in [0.35s, 8s]. If a period is found and it snaps to a nearby BPM subdivision (quarter, eighth…) within ±12%, mark as sustained_with_modulation.
  6. Spectral LFO override. Compute the MFCC centroid over the held region. If it has high coefficient of variation (≥ 0.10) AND autocorrelation shows a slow period in [1s, 12s], override to sustained_with_modulation even if the RMS variance test didn’t trip.

The tempo used for BPM-snapping comes from capture.tempo_bpm in the config (default 120).

3. Pitch verification

Verifies that each sample plays at the note its filename (or the VST render setup) claims. Uses autocorrelation on a 1-second window taken from the first quarter of the audio, looking in the range [sr/4000 Hz, sr/20 Hz].

The detected pitch is folded into the nearest octave of the expected pitch before comparison — autocorrelation regularly errors by octaves, and for the purpose of “did I render/label the right note” that’s not really an error. The remaining cent difference is checked against analysis.pitch_tolerance_cents (default: 50, i.e. ±quarter-tone).

The VST quirk

Pitch verification runs only for library sources — never for VST/CLAP. A plugin that’s told to play MIDI 60 will play MIDI 60; there’s nothing to verify. Library WAVs, on the other hand, might be mislabeled, octave-shifted, or de-tuned by the author on purpose. The drums profile also disables pitch verification (pitch_verify: false) — pitch on a drum hit is often not the fundamental you’d expect.

Out-of-tolerance samples are flagged, not rejected. The generated preset still includes them; the flag is a nudge to double-check.

4. Loop detection

Given a sustained-region audio buffer, find two sample indices (start, end) such that playing audio[start:end] on repeat is imperceptible from continuing to hold the original note.

patch-press runs three candidate hunters (chroma-based pitch matching, tempo-synced quarter-note multiples, period-aligned brute force), gates the results on seam quality and timbre continuity, scores the survivors with length and placement rewards, and picks the winner. An adaptive crossfade — sized in fundamental periods per note and scaled by the residual seam discontinuity — smooths any leftover seam.

If nothing passes the gates, a fallback takes the central 50% of the detected sustain and snaps to zero-crossings.

For the full picture — the candidate hunters explained in depth, what to do when a preset doesn’t loop, how to hand-edit a bad loop point, what loop_quality_threshold is for — see the loop detection page.

Honoring authored loops

Some libraries (Samples From Mars is one) bake loop points into the WAV file’s RIFF smpl chunk — a standard RIFF sub-chunk that stores loop start/end frame indices.

When patch-press reads a WAV that has a smpl chunk with at least one loop, it uses those authored loop points verbatim and skips loop detection entirely. Trim is constrained to stay outside the loop bounds. The author’s intent is treated as ground truth.

The Deluge itself doesn’t read the smpl chunk — the Deluge exporter writes loop points into the XML. So patch-press only consumes smpl on input; whether an exporter re-writes one on output is up to that specific target.

5. Normalize

Applies gain so the sample set peaks at −1 dB. Three modes:

Mode Behavior
per_set (default for synth/pad/pluck) All samples in the set share the same gain, chosen so the loudest hits −1 dB. Preserves relative velocity dynamics and round-robin variation.
per_sample (default for drums) Every sample is independently normalized. Loses velocity dynamics on purpose — each drum hit gets its own headroom.
none No gain applied. Useful when you’re feeding into another normalization stage.
This site uses Just the Docs, a documentation theme for Jekyll.