# Time-Stretch / Pitch-Shift Research Synthesis
Date: 2026-02-28
## Scope Notes
- The patent IDs requested in the prompt appear to include at least one mismatch:
- `US8805696` is an ExxonMobil process patent (not audio DSP).
- `US9263026` was not found as a valid/public audio patent record in the queried sources.
- Closest directly relevant patent family members found (Disch/Nagel/Walther lineage): `US8837750`, `US9230557`, `US9240196`, plus earlier phase-lock/partition work (`EP1918911A1` / `US20050010397A1`).
- When a detail below is marked **Inference**, it is reconstructed from multiple sources rather than explicitly stated.
### Requested Source Coverage (Prompt Checklist)
- `US 8,805,696`: checked; not audio DSP (mismatch).
- `US 9,263,026`: requested ID did not resolve to a matching/public audio record in searched sources.
- zplane/Disch/Wolff/Nagel/Walther patent lineage analyzed through: `US8837750B2`, `US9230557B2`, `US9240196B2`, `EP1918911A1`, `US20050010397A1`, `US10283130B2`.
- Laroche & Dolson (1999): covered.
- Dolson (1986): covered.
- Bonada (2000): covered via survey/toolbox lineage due limited direct full text access.
- Röbel (2003): covered from DAFx paper.
- Driedger/Müller/Disch and Driedger/Müller/Ewert (2014): covered via DAFx toolbox + SPL/HPSS hybrid references.
- Driedger & Müller (2016): covered in depth (review synthesis).
- Verhelst & Roelands (1993): covered.
- Grofit & Lavner (2008): covered.
- Magalhães & de Góes (2013): covered via cited selective phase-locking method summary.
- Moinet & Dutoit (2011): covered from DAFx PVSOLA paper.
---
## 1. Patent Analysis
### 1.1 Patent Inventory (Relevant)
- `US8837750B2` — transient-event handling in spectral manipulation with padded-block strategy.
- `US9230557B2` — continuation/refinement of transient-event handling and routing/recombination.
- `US9240196B2` — handling transient events under replay-speed/pitch modification.
- `EP1918911A1` / `US20050010397A1` — phase-locking with Bark-scale spectral partition and transient-focused processing.
- `US10283130B2` — vertical phase correction in subbands for phase-vocoder-style processing.
### 1.2 Per-Patent Architecture Extraction
| Patent | Signal Flow / Block Diagram | Transient Detection | Decomposition | Per-Component Processing | Recombination | Parameters / Values | Phase Coherence Strategy | Real-Time / Latency Claims |
|---|---|---|---|---|---|---|---|---|
| `US8837750B2` | Input -> transient detector -> adaptive pre/post processing around transient -> modified spectral path -> output | High-frequency-energy change around event time; transition-state selection around transient | Transient vs non-transient sub-regions via state machine | At transient windows: no time-stretch or alternate manipulation; non-transient regions stretch normally | Region-wise OLA with special transient insertion | Explicit example: process block length ~716 samples, time-stretch block length ~1024, ratio guidance ~1.4x | Preserve or reinsert transient-local phase/state while allowing spectral manipulation elsewhere | Designed for streaming blocks with bounded decisions per block |
| `US9230557B2` | Similar to `US8837750` with explicit subband routing and adjusted overlap skipping during transient | Same family: detects transient segment and routes around it | Spectral subbands + transient event path | Non-transient: stretch/pitch manipulation; transient bins/segments bypass or reduced manipulation | Re-add skipped overlap for output-length consistency | Window and skip/reinsert logic described at block boundaries | Avoids severe phase inconsistency by preserving original transient segments | Blockwise method suitable for real-time frame processing |
| `US9240196B2` | Audio manipulation pipeline with transient-aware correction during replay-speed/pitch change | Detect transient events, maintain event timing during manipulation | Event vs background separation | Event path preserved/adjusted distinctly from manipulated background | Event-aligned recombination | Event timing and frame-alignment constraints emphasized | Maintains event-local coherence, reduces smearing | Explicitly targets replay speed/pitch use-cases (practical real-time scenario) |
| `EP1918911A1` / `US20050010397A1` | STFT -> adaptive bark-like band partition -> phase-lock groups -> synthesis | Uses “dominant line” and/or abrupt spectral structure change (incl. transients) | Multiple frequency partitions (low/mid/high regions by spectral structure) | Different phase-lock strengths/peak-grouping by band | Weighted overlap-add synthesis | Bark-scale partition + varying lock weights; references 50%/75% overlap trade-offs in related work | Identity/peak-linked phase correction to reduce phasiness | Framed as efficient transform-domain processing |
| `US10283130B2` | Multi-band vocoder path with vertical phase correction in selected bands | Not primarily a detector patent; complements detector-based stacks | Subband decomposition | Per-band phase correction to retain vertical coherence | Standard synthesis recombination | Subband-local correction rules (no fixed FFT-size claim in extracted snippet) | Explicit vertical coherence correction | Focus on reducing artifacts without major extra compute |
### 1.3 zplane / Inventor Search Outcome
- Primary inventor trails for Sascha Disch in public records are heavily linked to Fraunhofer-assigned transient-handling and phase-vocoder patents that match Elastique-like behavior.
- Explicit “zplane development GmbH assigned” records were sparse in this specific subtopic search; most technically aligned records were in the Fraunhofer/Disch/Nagel lineage.
- **Inference:** product architecture likely combines patented transient-event bypass/reinsertion with modern multi-band phase-locking and robust real-time buffer management.
---
## 2. Academic Technique Synthesis
## 2.1 Phase Vocoder Foundations
### Dolson (1986) — Phase Vocoder Tutorial
- Core: STFT analysis/synthesis, phase unwrapping, instantaneous frequency estimation, overlap-add.
- Complexity: `O(T * N log N)` for FFT frames.
- Failure modes: transient smearing, phasiness from weak vertical phase coherence.
- Implementation-critical: consistent analysis/synthesis windows and COLA behavior.
### Laroche & Dolson (1999) — Improved PV / Phase Locking
- Core: preserve vertical coherence by locking neighboring bins to spectral peaks (identity phase locking).
- Complexity: PV + peak detection / bin-group assignment (near-linear in bins per frame).
- Artifacts reduced: metallic/phasey character on harmonic/polyphonic material.
- Tradeoff: peak assignment errors can cause local warble.
## 2.2 Transient Preservation
### Bonada (2000) (as summarized in Driedger/Müller survey lineage)
- Core direction: transient-aware frequency-domain handling to approach near-lossless TSM quality.
- Typical strategy: detect/segment transient regions and avoid naive phase propagation through attacks.
- Known issue addressed: blurred attacks and temporal envelope damage.
### Röbel (2003) — New Transient Processing in PV
- Core: bin-level transient detection/processing using spectral-bin criteria and transient position relative to window center.
- Important result: no strict need to force local ratio=1 if phase reinitialization is aligned near optimal transient position.
- Complexity: PV + bin-level transient logic and local criteria.
- Failure modes: detector mistakes still leak artifacts, but improved over broad-band transient gating.
### Driedger, Müller, Ewert (2014 SPL / 2014 DAFx toolbox context)
- Core: HPSS-based hybrid TSM:
- Harmonic component -> phase-vocoder + phase locking.
- Percussive component -> OLA/WSOLA-style path.
- Benefit: preserves both stable harmonics and transient punch better than single-method approaches.
## 2.3 Hybrid / Modern System Patterns
### Driedger & Müller (2016) Review
- Taxonomy: time-domain, frequency-domain, model-based, and hybrid combinations.
- Practical takeaway: no single algorithm dominates all signals; hybrid routing + transient-aware logic is strongest general strategy.
- Recommended evaluation includes objective + listening metrics, content-dependent presets.
### Nagel & Walther transient handling (AES 2009 citation chain)
- Core direction: explicit transient management within time-stretch algorithms to avoid attack corruption.
- **Inference:** similar design objective to transient bypass/reinsert patent family.
## 2.4 Time-Domain Methods
### Verhelst & Roelands (1993) — WSOLA
- Core: overlap-add with waveform similarity search to align local periodic structure.
- Complexity: `O(T * search_window * frame_len)` worst-case without optimization.
- Strengths: strong local transient/time-domain preservation for speech/percussive events.
- Weaknesses: stutter/drift artifacts when search picks poor matches; less robust on dense polyphony.
### Grofit & Lavner (2008) — Enhanced WSOLA
- Core: WSOLA + transient management (PST detection, selective handling of transients vs steady sections).
- Stronger quality on music than uniform WSOLA.
## 2.5 Phase Coherence Extensions
### Magalhães & de Góes (2013) — Selective Phase Locking
- Core idea: selective (not uniform) locking to preserve key spectral relationships while avoiding overconstraint.
- Benefit: better balance between coherence and flexibility.
### Moinet & Dutoit (2011) — PVSOLA
- Core: periodically reset PV phase coherence by inserting synchronization points selected with SOLA-like cross-correlation.
- Hybrid temporal/spectral correction loop.
- Strength: notable phasiness reduction without full-time phase-lock overhead.
- Cost: additional cross-correlation and reset scheduling logic.
---
## 3. Open-Source Implementation Study
## 3.1 Rubber Band (R3/Finer)
- Multi-resolution analysis bands (long/mid/short FFT ranges).
- Harmonic/Percussive/Residual classification using horizontal/vertical median filters.
- Segmentation guides adaptive FFT-band assignment and phase-lock bands.
- Guided phase advance with:
- peak tracking,
- channel-aware locking,
- kick/pre-kick logic,
- high-frequency unlock at extreme ratios.
- Real-time architecture:
- fixed-capacity ring buffers,
- explicit start pad/start delay reporting,
- bounded hop/ratio update logic.
- Elastique likely differs by:
- stronger proprietary event handling and psychoacoustic tuning,
- deeper per-content heuristics/presets,
- potentially tighter commercial QA across edge cases.
## 3.2 SoundTouch
- WSOLA-like time-domain core with auto sequence/seek window adaptation by tempo.
- Strong simplicity and speed.
- Limited transform-domain coherence controls; more artifact-prone under large/modulated ratios.
## 3.3 Paulstretch
- Extreme-stretch spectral method with randomized phases.
- Not suitable as general DJ-grade low-latency method.
## 3.4 librosa
- Educational/reference PV implementation.
- Explicitly not transient-robust; good baseline, not production quality target.
---
## 4. Taxonomy of Approaches
| Category | Strengths | Weaknesses | Best Use |
|---|---|---|---|
| Time-domain (OLA/WSOLA) | attack sharpness, low conceptual complexity | drift/stutter, weaker harmonic handling | speech/percussive mild ranges |
| Frequency-domain (PV + phase locking) | harmonic consistency, large ratio flexibility | transient smear if unmanaged | polyphonic/tonal material |
| HPSS hybrid | best broad quality trade-off | more modules + tuning complexity | full-mix music, DJ workflows |
| Model-based (sinusoidal+noise etc.) | high controllability | model mismatch risk, complexity | offline high-end processing |
---
## 5. Elastique Architecture Reconstruction (Best-Effort)
**Inference-based reconstruction:**
1. Multi-feature transient detector (flux/energy/phase + band emphasis).
2. Event-aware segmentation with lookahead.
3. Split processing:
- transient/event path (time-domain or minimally manipulated reinsertion),
- steady-state path (multi-resolution PV + adaptive phase locking),
- possible residual/noise path with relaxed locking.
4. Event-aligned recombination with adaptive crossfades.
5. Dynamic quality/latency mode controls and robust realtime buffering.
6. Pitch shift path includes envelope/formant protection and transient-aware event handling.
7. Stereo is one instance with linked analysis, not two mono instances — the V3 SDK documentation states this outright, for both quality ("the content of both channels is taken into account for the analysis, and the stereo processing is linked between both channels") and performance. An explicit M/S input mode exists (`kMSMode`).
8. Output blocks are capped at 1024 frames, and a DirectAPI exists whose stated purpose is spreading each block's cost across small audio callbacks ("performance peaks ... at low latencies"). Pro is documented as about 2× the CPU of Efficient; Efficient about 2× the mobile mode. (SDK docs read 2026-09-02; see Sources.)
---
## 6. Critical Implementation Details (What Separates Good vs Great)
- Analysis/synthesis window pair must preserve overlap-add normalization under chosen hop ratios.
- Phase locking should be selective and peak-informed, not globally rigid.
- Transient detection must include adaptive thresholding and local timing refinement.
- Event reinsertion/bypass should be aligned to window-center criteria (Röbel-style insight).
- Multi-resolution bands improve quality but require robust guide logic at ratio extremes.
- Cross-channel coherence handling (stereo mid/side or shared guidance) is essential for DJ use.
---
## 7. Transient Handling Deep Dive
- Detector features used in high-quality systems:
- weighted spectral flux,
- frame energy slope,
- phase deviation / group-delay cues,
- low/high-band specific checks for kick/hat discrimination.
- Decision mechanics:
- adaptive thresholds (median/MAD or mean/sigma),
- minimum onset spacing,
- lookahead confirmation (avoid false spikes),
- event strength scoring for path routing.
- Processing:
- transient path via WSOLA/direct-event insertion,
- steady path via PV,
- optional HPSS residual handling.
- Recombination:
- raised-cosine crossfades at boundaries,
- overlap compensation to keep exact final timeline.
---
## 8. Phase Coherence Strategies
- Horizontal coherence (frame-to-frame bin continuity): standard PV unwrapped phase propagation.
- Vertical coherence (across bins in one frame): identity/selective phase locking around spectral peaks.
- Channel coherence: shared peak/channel-lock policy at low-mid frequencies.
- Practical pattern: use stronger locking in low bands, relax in highs at high stretch ratios.
---
## 9. Real-Time Constraints and Budgets
Target (DJ): `<10 ms` perceived/control-latency envelope where feasible.
Implementation constraints:
- fixed-capacity ring buffers,
- zero-allocation in steady audio callback path,
- no locks/syscalls in callback,
- bounded per-block loops,
- preplanned FFT/resampler allocations,
- SIMD or vectorized hot paths for OLA/windowing/magnitude/phase loops.
### 9.1 Latency budgets in shipping DJ products (2026-09-02)
Three budgets, only one of which DJ software exposes to the user:
1. **Audio interface buffer** — the number DJs tune. Traktor and
rekordbox guidance is 128–256 samples; Ableton performers run 64–128
for instruments. At 44.1 kHz: ~3–6 ms per buffer, ~10–15 ms round
trip through driver input and output stages.
2. **Gesture-to-audio** — jog nudge, pitch bend, hot-cue jump, scratch.
Buffer plus whatever the stretcher delays. Pioneer publishes no
master-tempo delay for the CDJ-3000; what it advertises is jog touch
latency halved versus the NXS2.
3. **Stretcher-internal delay** — hidden on the timeline by every DJ
app (playhead, waveform, cues, sync all read ahead). Only budget 2
exposes it.
Reported stretcher delays:
| Engine | Delay at 44.1 kHz | Source |
|---|---|---|
| Rubber Band R2 short / default / long window | 512 / 1024 / 2048 frames (12 / 23 / 46 ms) | breakfastquay integration guide |
| Rubber Band R3 short / default | 1280 / 2048 frames (29 / 46 ms) | same |
| Elastique Pro / Efficient V3 | no fixed figure; dynamic via `GetFramesNeeded`, initial unused frames discarded; output block ≤ 1024 frames | zplane SDK docs |
| Elastique Pitch plugin (Pro engine + resampler + formant path) | 163 ms at 44.1 kHz, 150 ms at 48 kHz | zplane product page |
| timestretch-rs Keylock / WideKeylock | 560 frames (12.7 ms) / 0 ms (source-side lookahead) | README |
Reading: the 560-frame (12.7 ms) Keylock contract is stricter than
anything shipping and equal to R2's short window. 2048 frames is the
class R3 and, by its block sizes, Elastique operate in — and at 44.1
or 48 kHz it is the FFT size that resolves bass (46 ms window, 21.5 Hz
bins). ROADMAP's two-budget policy follows from this table: gestures on
the 12.7 ms profile, corrected steady playback on a 46 ms profile,
timeline-compensated exactly as DJ apps do.
Sample rate caveat. The table is in frames because the vendors quote
frames, and their window sizes do not scale with rate: Rubber Band's
2048-frame default is 21 ms at 96 kHz, with 47 Hz bins. Our budgets
are stated in time instead. Keylock already scales its lag with the
rate (`keylock_latency_frames`: 560 frames at 44.1 kHz, ≈1219 at
96 kHz, 12.7 ms either way); the wide PV head does not (`WIDE_FFT` is a
fixed 2048), so at Halo's 96 kHz it runs half the frequency resolution
the bass argument above assumes. The Stage 26 profile therefore sizes
its FFT from the rate — 2048 at 44.1/48 kHz, 4096 at 88.2/96 kHz — and
is deliberately stricter than the vendors at high rates.
---
## 10. Quality Benchmarking Guidance
Objective metrics to track:
- transient timing MAE (ms) at expected event anchors,
- STFT magnitude distortion vs reference (`L1/L2` normalized),
- phase-coherence stability (pairwise harmonic phase-diff variance),
- unexpected-band energy ratio (artifact leakage proxy),
- loudness drift (LUFS delta).
Subjective checks:
- kick/snare attack sharpness,
- bass solidity (no flanging/cancellation),
- vocal clarity/formant stability,
- phasiness/metallic tails on pads and cymbals.
---
## 11. In-Repo Empirical Findings (2026-03 Autoresearch, Legacy Streaming Engine)
Provenance: an automated research loop (March 2026) optimized the old
`StreamProcessor` streaming path — PV-only, later deleted in the Stage 9 engine
cutover (d44bb7e, 2026-07-15). Its scores were measured against the also-deleted
`qa/streaming_quality.rs` benchmark (0–1000 composite; plateaued at ~969.7).
The full experiment logs and exhaustive parameter sweeps live in git history at
`autoresearch.md` / `autoresearch.ideas.md` (last version: commit b732b6f). The
sweeps tuned a gain-compensation mechanism that no longer exists and are
intentionally not carried forward. The durable findings below were re-checked
against the current engine (SOLA + varispeed live path; batch PV only as the
offline wide-ratio fallback, `src/engine/offline.rs`) in August 2026.
### Verified against the current engine (2026-08)
- **PV energy loss is phase-modification destructive interference, not a
normalization bug.** Legacy: streaming PV output at ~35–50% input RMS.
Re-verified on the offline wide path (`stretch()` at 1.30×, `normalize`
off): broadband content drops to 0.64–0.68× input RMS (kick pattern, EDM
mix) while a pure sine keeps 0.99×, confirming the coherence mechanism.
The engine-graph path (1.10×, SOLA + varispeed) shows no deficit
(0.93–1.00×). Mitigation today is the opt-in `params.normalize` RMS match;
there is no spectral gain-compensation stage, by design.
- **Envelope preservation does not measurably affect time-stretch quality;
it matters for pitch shift.** The current architecture embodies this:
spectral-envelope correction runs only on the pitch-shift/formant path
(`src/lib.rs`, gated on `pitch_factor != 1.0`); the offline wide PV
constructs with envelope preservation off.
### Partially reproduced (2026-08)
- **Phase-locking mode ranking (legacy: ROI ≥ Identity > Selective).** On the
current offline wide path at 1.30× scored against RubberBand references,
the result is mixed: RegionOfInfluence improves spectral-flux similarity
(kick 0.69 → 0.79, EDM mix 0.79 → 0.83) and nudges spectral/perceptual
similarity up, but regresses transient onset match (kick 1.00 → 0.89,
EDM 0.95 → 0.89). No clear winner; `Identity` remains the wide-path
choice. Revisit only with a listening pass.
### Not reproduced in isolation (2026-08)
- **"Transient phase resets hurt harmonic quality" (legacy: −52 points).**
Isolated test on the current PV (streaming entry point, Identity locking,
2048/512, steady harmonic chord, periodic `reset_phase_state()`): output
re-seeds phase but shows no level dips or spectral degradation
(similarity 0.9995 vs the no-reset run). The legacy harm was evidently an
interaction with the deleted engine's scheduling/overlap machinery, not a
universal PV law. Treat as historical, engine-specific guidance. The reset
machinery is live: the wide keylock stage calls
`reset_phase_state_bands` on onset events
(`src/engine/stages/wide_keylock.rs`).
### Historical (died with the streaming PV engine — kept as context)
- Adaptive phase-locking **mode switching** mid-stream caused phase
discontinuities and hurt quality. (No adaptive switching exists today.)
- PV preserved high frequencies better than WSOLA; WSOLA lost highs on
transient material. (WSOLA is deleted; the SOLA high-band corrector is the
current time-domain analog, chosen after listening passes rejected a
small-FFT PV corrector as phasey — see `src/engine/stages/keylock.rs`.)
- Architectural limits of the streaming PV that motivated the engine rewrite:
EDM 1.5× centroid shift (broadband 300–3000 Hz loss indistinguishable
per-chunk from percussive loss), chunk-boundary effects on harmonic
content, and a structural quality gap between streaming PV and the batch
hybrid segmentation it was scored against.
---
## Sources
### Patents / Patent Records
- [US8805696](https://patents.google.com/patent/US8805696)
- [US8837750B2](https://patents.google.com/patent/US8837750B2/en)
- [US9230557B2](https://patents.google.com/patent/US9230557B2/en)
- [US9240196B2](https://patents.google.com/patent/US9240196B2/en)
- [EP1918911A1](https://patents.google.com/patent/EP1918911A1/en)
- [US20050010397A1](https://patents.google.com/patent/US20050010397A1/en)
- [US10283130B2](https://patents.google.com/patent/US10283130B2/en)
- [US9236062B2](https://patents.google.com/patent/US9236062B2/en) — Fraunhofer transient cut/stretch/reinsert; active to 2029-09-29 per Google Patents (ROADMAP Parity Track IP note)
- [US6549884B1](https://patents.google.com/patent/US6549884B1/en) — Laroche–Dolson PV pitch shifting by peak-region shift; expired 2019
- [US8489404B2](https://patents.google.com/patent/US8489404B2/en) — transient frames left unscaled; lapsed 2021
### Papers / Technical Sources
- [Laroche & Dolson 1999 (IEEE record)](https://ieeexplore.ieee.org/document/759041)
- [Dolson 1986 tutorial text](http://www.panix.com/~jens/pvoc-dolson.par)
- [Röbel 2003 DAFx PDF](https://www.dafx.de/paper-archive/2003/pdfs/dafx32.pdf)
- [Driedger & Müller 2016 review (MDPI)](https://www.mdpi.com/2076-3417/6/2/57)
- [TSM Toolbox DAFx 2014](https://www.dafx.de/paper-archive/2014/dafx14_jonathan_driedger_tsm_toolbox_matlab_imple.pdf)
- [WSOLA original IEEE record](https://ieeexplore.ieee.org/document/319366)
- [Enhanced WSOLA w/ transients IEEE record](https://ieeexplore.ieee.org/document/4381234)
- [PVSOLA DAFx 2011 PDF](https://www.dafx.de/paper-archive/2011/Papers/57_e.pdf)
- [Průša & Holighaus 2022 — Phase Vocoder Done Right (arXiv)](https://arxiv.org/abs/2202.07382)
- [Průša & Søndergaard 2016 — Real-Time Spectrogram Inversion Using PGHI (DAFx)](https://dafx.de/paper-archive/details/KyhYU44R13wwkkqNs3bgPw)
### Open-Source Implementations
- [Rubber Band](https://github.com/breakfastquay/rubberband)
- [SoundTouch](https://codeberg.org/soundtouch/soundtouch)
- [Paulstretch](https://github.com/paulnasca/paulstretch_cpp)
- [librosa](https://github.com/librosa/librosa)
### Product Documentation (latency budgets, §9.1)
- [Rubber Band integration guide (start delay table)](https://breakfastquay.com/rubberband/integration.html)
- [Elastique Pro V3 SDK documentation](https://licensing.zplane.de/uploads/SDK/ELASTIQUE-PRO/V3/manual/elastique_pro_v3_sdk_documentation.pdf)
- [Elastique Pro V3 DirectAPI documentation](https://licensing.zplane.de/uploads/SDK/ELASTIQUE-PRO/V3/manual/elastique_pro_v3_sdk_directapi_documentation.pdf)
- [Elastique Efficient V3 SDK documentation](https://licensing.zplane.de/uploads/SDK/ELASTIQUE-EFF/V3/manual/elastique_efficient_v3_sdk_documentation.pdf)
- [zplane Elastique Pitch (plugin latency figure)](https://products.zplane.de/products/elastiquepitch)
- [zplane technology overview](https://licensing.zplane.de/technology)
- [DJ TechTools — This is Pioneer DJ's CDJ-3000](https://djtechtools.com/2020/09/09/this-is-pioneer-djs-cdj-3000/)
- [DJ TechTools — Latency, a brief guide](https://djtechtools.com/2011/05/03/latency-a-brief-guide/)
- [Rekordbox latency adjust thread (buffer guidance)](https://forums.pioneerdj.com/hc/en-us/community/posts/4406806265881-Rekordbox-Latency-adjust)
- [Ableton audio setup guide (buffer sizes)](https://www.audeobox.com/learn/ableton/ableton-audio-setup-guide/)