Matt Wood piano sampler — slice raw takes, generate SFZ + Kontakt .nki

[ross-makeacompany]

Context

Matt Wood (composer/musician, cmyk-music.com, Syracuse Architecture classmate of Grant's) recorded all 88 keys of a piano he used to own, one note at a time, but never sampled it into a playable instrument. He's done it manually before in Kontakt and it's too time-consuming. Goal: automate slice → label → SFZ → .nki pipeline so he gets a playable plugin without the hand-mapping tedium.

Intro happened in this channel on 2026-05-26.

Deliverable

Two Kontakt 6 instruments (separate .nki + sample folders), shipped together:

• Piano_Sustain.nki — pedal-down long notes
• Piano_Staccato.nki — short notes

Each backed by an SFZ so it also runs free in sfizz for non-Kontakt users.

V1 ships as two separate instruments (Matt's call). Keyswitch consolidation can come later.

Source material

Matt is sending a zip (Drive invite to ross@bimross.com) of 6 continuous WAV files, each ~25 minutes, 24-bit / 48 kHz stereo. Each file is one pass, chromatic A0 → C8:

File	Articulation	Dynamic	Round-robins
1	Long (pedal-down)	Quiet	RR1
2	Long (pedal-down)	Quiet	RR2
3	Long (pedal-down)	Loud	(single)
4	Short (staccato)	Quiet	RR1
5	Short (staccato)	Quiet	RR2
6	Short (staccato)	Loud	(single)

Total samples after slicing: ~528 (88 × 6), modulo retakes.

Recording notes from Matt

• Gaps between notes are nearly silent. Some room noise from hot mic preamps and audible piano mechanics — wanted, do not remove.
• Within any pass, a note may have been retaken 2× or 3× if the first attempt was bad. Rule: always keep the LAST take of a given pitch (also has the most tail).
• No release tails captured separately.
• Matt's aesthetic: transparent, close, wide. No processing of any kind — no EQ, no compression, no normalization, no denoise, no de-hum, no de-click in the body of the note.

Pipeline stages

0. Slice + label

• Silence-split each of the 6 files (sox / pydub.silence.split_on_silence or equivalent).
• Sample-accurate cuts on zero-crossings to avoid boundary clicks.
• Light pitch detection (FFT peak on first ~200ms, or CREPE for safety) on each slice → identify fundamental → group by detected MIDI note → keep the last slice per note (handles retakes).
• Output per file: 88 labeled WAVs named Piano_<Note>_<sus|stacc>_v<1|2>_rr<1|2>.wav (e.g. Piano_C4_sus_v1_rr1.wav).

1. Edge treatment (ONLY processing allowed)

• Head: ~2-3 ms fade-in to kill boundary pops.
• Tail: ~50-100 ms gentle fade-out, erring long so natural decay is preserved.
• Both configurable; iterate with Matt after first pass.
• Output stays 24-bit / 48 kHz stereo WAV. No resampling, no dither, no re-encode.

2. SFZ generation

Two SFZ files (piano_sustain.sfz, piano_staccato.sfz). Per-instrument structure:

<group> // quiet layer
lokey=A0 hikey=C8
lovel=0 hivel=110
amp_veltrack=100        // 0-40 naturally quietest via velocity curve
seq_length=2            // 2 RR alternate

<group> // loud layer
lovel=111 hivel=127
amp_veltrack=100
seq_length=1            // single RR, no alternation

Per-sample <region> entries with pitch_keycenter, lokey/hikey (interpolated for any holes), sample=..., seq_position=N for RRs.

Release envelope per group; tunable.

3. Kontakt import

Drag SFZ into Kontakt 6 → File → Import → SFZ. Save as .nki. Kontakt builds the instrument and we tweak envelopes / EQ-off / output routing in Kontakt's UI before saving the final .nki.

Licensing / ownership

• Repo is private under BimRoss. Code MIT.
• Samples remain Matt's IP.
• If Matt ever wants to commercially distribute, packaging options range from a free zip (.nki + samples) to NI's Kontakt Player NKS program (expensive, application-based). Out of scope for v1.

Open questions / risks

• Pedal-sustain bleed: if long-quiet notes ring into each other, silence-split may merge them. Fallback: longer silence threshold or pitch-change-based segmentation.
• Retake count uncertainty: Matt doesn't know exactly how many notes were retaken. The "group by detected pitch, keep last" rule handles arbitrary counts.
• Asymmetric RR on loud layers: single RR on loud. Acceptable for v1 per Matt ("ship honest first"). Could fake a second RR later with 5¢ detune + 1 ms offset if needed.

Timeline

First playable .nki to Matt within a week of receiving the audio. Faster if files are clean.

Status

• Repo created
• Audio zip received from Matt
• Repo scaffolded with piano-sampler/ subdir + Python pipeline
• Slice + label working on one file
• All 6 files processed, 528 labeled samples
• SFZ files generated
• First .nki delivered to Matt for review
• Iteration on fades / envelopes / RR feel
• V1 ship

[ross-makeacompany]

Update: release behavior + sustain pedal (CC64)

Adding to the spec per Matt:

Note-off response (no pedal):

• Sustain instrument: ampeg_release ~1.0-1.5s (gentle damper-like taper, tune to taste).
• Staccato instrument: ampeg_release ~0.1-0.3s.

Sustain pedal (CC64):

• SFZ: sustain_sw=on, sustain_cc=64. Standard damper behavior — while pedal ≥ 64, note-off is deferred; sample rings through natural decay until pedal lifts.
• On pedal release (key already up), same ampeg_release taper applies.

Sample-length ceiling:

• The natural decay length recorded in each WAV is the hard ceiling — no synthetic extension.
• Target: 8+ seconds of usable ring while pedal is held. Will measure decay duration per slice after audio lands and flag any registers (likely top octave) that fall short.

Tail-fade clarification:

• The ~50-100ms cosmetic fade-out is applied to the very end of the WAV file (after natural decay → near-silence), only to ensure zero-crossing termination and avoid file-end clicks. Does not truncate audible ring.

[ross-makeacompany]

Pre-implementation audit

Re-reading the spec end-to-end before Matt's audio lands. Surfacing risks, gaps, and decisions worth resolving up front so the first pass doesn't get derailed.

1. Stale fact in spec body

• Body says "Repo is private under BimRoss" — it's now public (flipped at Grant's request so Matt can read this issue without a GH account). Not editing the body to preserve history; flagging here.

2. Slice-boundary strategy needs a real plan (the spec hand-waves it)

The spec says "silence-split each file." That's the easy case. Real risks for the long/quiet passes (pedal-down):

• Notes can ring 10-15s. If Matt didn't wait for full decay before the next note, slices will overlap and silence-split will fail or merge notes.
• "Hot preamps" → elevated noise floor. A fixed silence threshold won't work across files; need per-file adaptive threshold from a measured pre-roll noise sample.
• Pedal/foot noise + piano mechanics can spike above silence threshold mid-decay → false boundaries.

Better plan: hybrid onset detection (librosa.onset.onset_detect or a transient detector) defines slice starts; each slice extends until the next onset (or EOF). Then trim trailing silence per slice with a generous threshold so we keep the full audible decay. This handles overlap and noisy gaps both.

3. Pitch detection — exploit the chromatic prior, don't fight it

The spec proposes "FFT peak on first ~200ms, group by detected note, keep last." Two issues:

• First 200ms includes the attack transient + hammer noise → broadband, terrible for fundamental detection. Should use a window after the attack settles (e.g. 50-300ms post-onset) and/or autocorrelation/YIN instead of FFT peak.
• A0 (27.5 Hz) and low octave — FFT peak is unreliable; will often report a harmonic. CREPE handles low frequencies; use it (or YIN) for the bottom octave at minimum.
• Octave errors are the classic failure mode. Sanity-check via the chromatic-order prior.

Better algorithm — chromatic walk: since the recording order is strictly A0 → C8, the expected pitch for slice N is A0 + N (after retakes are removed). Walk through slices; if detected[N] == expected, advance. If detected[N] == expected - 1, treat as retake (replace previous, don't advance). This is far more robust than independent pitch detection per slice, and gives us a built-in integrity check: each file should end with the walk reaching C8.

4. SFZ → Kontakt 6 import: known limitations

Worth setting expectations now. Kontakt's SFZ importer is partial. Specifically:

• seq_position / seq_length round-robin opcodes — Kontakt's import historically does not wire these up correctly. Often imports as separate overlapping groups that all trigger simultaneously. Likely manual fix in Kontakt's Group Editor post-import.
• sustain_sw=on / sustain_cc=64 — may or may not be honored. CC64 sustain behavior may need manual KSP or Group settings.
• Velocity ranges, key mapping, root keys, ampeg_* — these import reliably.

Plan: ship SFZ import as best-effort, plus a documented "Kontakt fixups" checklist for: (a) wiring RR groups, (b) CC64 sustain, (c) save-as-.nki step. Matt does this once per instrument; the bulk of the tedium (528 samples to map) is still gone. The SFZ also runs cleanly in sfizz without any fixups, which is the fallback if Kontakt fights us.

5. Velocity-split discontinuity at v=111

amp_veltrack=100 on both groups + hard switch at v=110/111 means at v=111 the loud sample plays at the bottom of its velocity-scaled range — potentially quieter than the quiet sample at v=110. Audible duck/dip across the threshold.

Two fixes, pick later:

• Lower amp_veltrack on the loud group (~30-50) so the loud layer stays roughly constant across 111-127.
• True velocity crossfade with xfin_lovel/xfin_hivel etc. around the boundary.

I'll tune amp_veltrack on the loud group for v1, log the duck risk, and let Matt say if it's audible before adding xfades.

6. File naming — drop the #

Piano_C#4_sus_v1_rr1.wav has filesystem and DAW-import gotchas on some platforms. Use either:

• Sharp-letter notation: Cs4, Fs5 (readable, sorts correctly).
• MIDI note numbers: Piano_060_sus_v1_rr1.wav (unambiguous, sorts numerically, no octave-numbering confusion between C3-as-middle-C vs C4-as-middle-C).

Recommending MIDI note numbers internally, with a human-readable note name in a sidecar manifest. Cleaner for the SFZ generator, no naming-convention bikeshedding.

7. Pedal-down sympathetic resonance baked into long samples

Matt recorded the long pass with pedal held down. That means every long sample contains sympathetic ring from other strings — that's the recorded reality and presumably the intended sound. But: when a user plays a single note in the .nki without holding the sampler's CC64, they'll still hear pedal-down resonance baked into the sample. This isn't a bug per Matt's "transparent, what's on tape" ethos — but worth confirming he expects it.

8. Slice "end" is underspecified

The spec defines the head and tail fade, but not where each slice's end-of-audio actually is. With pedal-down ringing, the slice should extend until amplitude drops below (noise_floor + margin) for a sustained window (~500ms). That's the point where the cosmetic 50-100ms fade-out is applied. Without explicit end-detection, we risk either (a) chopping audible decay or (b) including the start of the next note.

9. Gain match across the 6 takes

The 6 passes were recorded across what may have been separate sessions. If gain staging differs even slightly between files, v1 vs v2 (or quiet vs loud) perceived loudness won't match the captured intent. We are correctly forbidden from normalizing, but we should measure per-file RMS / peak and report to Matt before generating the SFZ — gives him a sanity check (and an optional volume= knob per group if he wants to nudge without destructive edits).

10. Stereo + bit-depth invariants

Belt-and-braces: add a hard pre-check that every input file is exactly 24-bit / 48 kHz / stereo. Fail fast with a clear error if not. Output is plain RIFF WAV (fmt + data only, no BWF/non-standard chunks — Kontakt has historically rejected weird WAV chunks).

Zero-crossing detection for slice boundaries: operate on sum(L, R) or max(|L|, |R|) — never per-channel independently (would break stereo coherence).

11. Testing without Matt's audio

So we can validate the pipeline today rather than first-pass on real audio: build a synthetic test that emits a sine-sweep "fake piano take" (88 chromatic sine notes with envelopes, plus a few intentional retakes, plus injected room-noise) → run pipeline → assert 88 correctly-labeled outputs. Cheap, catches obvious bugs, gives Matt confidence the pipeline isn't speculative when he sees the first .nki.

12. Proposed repo layout

For when the audio lands, planned structure:

audio-plugins/
└── piano-sampler/
    ├── README.md
    ├── pyproject.toml
    ├── src/piano_sampler/
    │   ├── slice.py          # onset/silence detection, slicing
    │   ├── pitch.py          # CREPE/YIN + chromatic walk
    │   ├── edge_treatment.py # head/tail fades
    │   ├── sfz.py            # SFZ generator
    │   └── cli.py            # `piano-sampler build <input-dir> <output-dir>`
    ├── tests/
    │   ├── synthetic_take.py # builds a fake-piano test fixture
    │   └── test_pipeline.py
    └── kontakt-fixups.md     # the "what Matt does in Kontakt after import" checklist

Samples themselves won't be committed (large + Matt's IP). Output directory referenced by relative path from the .nki.

13. Decisions I'll make if Matt doesn't weigh in

To keep momentum once audio lands, defaults I'll ship with on first pass — Matt can countermand any of these:

• MIDI note numbers in filenames.
• Onset-based slicing with adaptive per-file silence floor.
• CREPE for pitch on bottom octave, YIN above.
• ampeg_release: 1.2s sustain, 0.2s staccato.
• amp_veltrack: 100 on quiet group, 40 on loud group (mitigates v=111 dip).
• Hard switch at v=110/111 (no xfade in v1).
• Tail fade-out: 80ms linear, applied after decay reaches noise floor.
• Head fade-in: 2ms linear.

All configurable via CLI flags / config file; nothing baked in.

14. Status checklist updates

Adding to the bottom of the checklist:

• Synthetic test fixture working end-to-end
• Per-file noise floor + RMS report generated for Matt before SFZ
• Kontakt fixups doc written + tested on one instrument

[ross-makeacompany]

Instrument name: Bubby's Piano (per Matt, 2026-05-26).

Naming updates:

• Deliverables: BubbysPiano_Sustain.nki + BubbysPiano_Staccato.nki
• Sample folder: BubbysPiano/Samples/
• SFZ files: bubbys_piano_sustain.sfz, bubbys_piano_staccato.sfz
• Pipeline tooling subdir stays generic (piano-sampler/).

[ross-makeacompany]

Name provenance: The piano belonged to Carolyn's Bubby (Matt's wife's grandmother). The instrument carries her piano forward — name honors the source.

[ross-makeacompany]

Spec update (Matt, 2026-05-26): velocity crossfade instead of hard switch at v=110/111.

• Xfade band: v=100 → v=120 (tunable).
• Quiet group: `lovel=0 hivel=120`, `xfout_lovel=100 xfout_hivel=120`.
• Loud group: `lovel=100 hivel=127`, `xfin_lovel=100 xfin_hivel=120`.
• Same xfade applied to both Sustain and Staccato instruments.

Risk reminder: Kontakt SFZ importer historically doesn't honor xfin_*/xfout_* opcodes — falls under the same "Kontakt fixups" checklist as RR + CC64. Will validate the xfade in sfizz first (where xfade opcodes work cleanly) before deciding whether Kontakt needs manual Group Editor setup.

Supersedes earlier "Hard switch at v=110/111 (no xfade in v1)" default in the pre-implementation audit.

[ross-makeacompany]

Tuning rule (Matt, 2026-05-26): Bubby's may be a few cents off A440 globally (e.g. A439). That's the piano — preserve it.

• Pitch detector tolerance: ±50 cents per note when assigning detected pitch to nearest MIDI note. Comfortably handles stretch tuning at keyboard extremes too.
• Chromatic walk uses relative semitone steps, so global tuning offset is invisible to the labeler.
• No detuning in the SFZ. pitch_keycenter names the nearest MIDI note so the instrument plays C4 when you press C4 on the keyboard — but the recorded sample plays at its native pitch. The slightly-off intonation is part of the instrument's character, not a bug to fix.

[ross-makeacompany]

Noise-vs-character distinction (Matt, 2026-05-26):

• Piano-internal sounds (mechanics, pedal noise, sympathetic resonance, mic-preamp hiss, room air) → character, preserve.
• Human/incidental noises (chair creak, movement, breathing, off-mic sounds) → noise, remove where safely possible.

Pipeline handling:

1. Between-note creaks: dropped by adaptive trailing-silence trim. Threshold set a few dB above measured per-file noise floor — eats chair creak in the gap, not the decay tail of the preceding note.
2. Mid-body creaks: not auto-processed (no-body-processing rule). Pipeline emits a suspect_slices.md report flagging slices with unusual mid-body transients / RMS spikes / low-frequency thumps. Matt spot-checks the list; contaminated samples can be re-recorded (Matt sends N replacement notes, pipeline drops them in via the same chromatic-walk dedup logic) or accepted.

This keeps the "no destructive processing" rule honest while giving Matt a finite list of slices to review instead of all 528.

[ross-makeacompany]

Constraint (Matt, 2026-05-26): Re-recording is not an option. All workarounds must use existing samples.

Updated handling for contaminated slices (supersedes earlier "re-record N notes" path):

Substitution priority:

1. Other RR, same note/dynamic/articulation. If RR1 is contaminated but RR2 is clean, RR2 plays in both slots. Cost: one note loses true round-robin variation.
2. Same note, other dynamic (cross-borrow quiet ↔ loud). Last resort, changes velocity character. Flag and ask Matt before applying.
3. Accept and ship. If the noise is minor, leave it. Documented in the deliverable.

Explicitly out:

• Pitch-shift from neighbor notes (violates no-body-processing rule, sounds bad).
• Spectral denoise / RX-style scrubbing on the body (same rule).

M3 deliverable change: suspect_slices.md becomes a decision sheet — one row per flagged slice with the substitution options surfaced. Matt picks; pipeline applies the chosen substitution in the SFZ generator (via per-region overrides rather than touching the WAVs).

[ross-makeacompany]

M1 complete — pipeline alive on synthetic data.

Commit: b146ea7

On a 6-file synthetic input set (88 chromatic notes × 6 passes, 2 retakes per file + injected chair-creak transient), the build command produces 528 correctly-labeled per-note WAVs and two valid SFZ files (Sustain + Staccato) in ~11 seconds.

Sanity-check output from build_summary.json on the synthetic data:

• All 88 notes covered (A0–C8) across every input file
• Retakes correctly resolved (chromatic walk keeps the last take per detected pitch)
• Chair-creak transient flagged as non_pitched and rejected from the sample set
• Bubby's global -4 cent detune (simulated) detected and preserved (no auto-correction applied)
• Noise floor measured at -71 dBFS per file (percentile-based, robust to no pre-roll)

Pipeline runtime is well within budget — should comfortably process Matt's ~1.8 GB of real audio in a few minutes.

Next: M2 — download Short_Quiet_RR 1.wav (smallest real file, 166 MB), run the pipeline against it, hand Matt 88 labeled WAVs to spot-check.