Acoustician

Development biography · earlier archive

Development profile compiled from project source and retained records in October 2026. Dates and test results describe those records; they do not imply a new build, live acceptance test or public release.

Detailed project biography

Acoustician brings together two activities that are often separated in music software: working creatively with a recording and examining whether a processing result is actually trustworthy. A producer may want to isolate a kick, identify a break, explore a harmonic layer or recover the character of an instrument. A researcher may want to compare algorithms, measure spill, or understand why a quiet cymbal disappears. Acoustician gives these activities a shared workspace while keeping an estimate distinct from the original recording.

The standalone application grew from the Acoustician module inside AssistantApp. Its studio compiles the shared Acoustician and StemSplitter source rather than maintaining a second separation engine. The interface gives that engine a dedicated home: a real short-time Fourier-transform spectrogram, an amber-on-dark acquisition display, a timeline with markers, a source panel and a categorized module rack. Floating parameter panels expose different operations without requiring the user to treat every audio problem as the same kind of separation.

The central problem is that a drum recording is more complicated than a list of instruments. Kick and snare can arrive together. A cymbal continues ringing while another piece is struck. Ghost notes may sit below louder attacks, and a microphone assigned to one drum also hears its neighbours. Electronic drums introduce another opportunity: some hits repeat closely enough to support reference matching, but sampled or noise-based sounds still vary. The project's documented direction is to distinguish these situations and assess individual pieces rather than equating a broad “drums” stem with a recovered kit.

A typical session starts by importing audio and examining its spectrogram. The user auditions the recording, moves through the timeline, marks passages and chooses an operation suited to the material. The source panel makes the resulting layers available for selection, soloing, muting and playback. Generic DSP quality modes trade processing breadth against isolation and resource limits. Neural actions use an available local backend, while an unavailable model produces a visible status instead of silently claiming that another method performed the work.

For a complete drum bus, the optional LarsNet integration offers five broad stereo groups: kick, snare, toms, hi-hat and cymbals. Acoustician's adapter processes the complete original stereo acquisition and returns estimates at its original rate and duration, with a separate remainder and an inference receipt. This is an integration of the published third-party model, not an Entertrainment-authored neural architecture. It also does not imply separate recovery of every tom, open and closed hat, ride, crash, rim or clap.

When suitable isolated instrument stems are available from the same recording, reference-kit fitting offers another route. The default spectral method learns several shapes per instrument and fits them jointly. An optional event method also models approximately 300 milliseconds of attack and decay, allowing several instruments to activate together. This gives the user a concrete comparison between spectral similarity and time-varying hit structure. The result preserves original complex stereo through masks and retains unassigned material in the remainder; neither convergence nor reconstruction conservation proves that the labels are correct.

Kit Lab exposes the investigation itself. Its stages cover onset detection, frequency representation, decomposition and bleed analysis, followed by measured results. The user can compare magnitude-flux and phase-aware onset approaches, explore constant-Q or variable-Q frequency axes, and inspect the effects of different decomposition choices. Elsewhere in the rack, tonal/transient/noise decomposition, query-by-example extraction, drum resynthesis, cross-refinement, fingerprint identification and break analysis provide further ways to examine or transform material. Their presence in source is distinguished from a claim that every operation has completed independent runtime acceptance.

Acoustician also makes a deliberate distinction between separation and restoration. Source-mask drum processing is accompanied by a remainder; adding modelled timbre is an explicit synthetic-restoration choice. Exports identify the stems and processing metadata, and can include the original decoded stereo alongside the estimates. The source file remains authoritative. The shared segmentation model records the acquisition, analysis priors and derived overlays using audio-native provenance, including JSON-LD representation.

The research workflow is an important part of the project's identity. On 11 October 2026, the event-reference method improved the mean stereo SI-SDR on a controlled 909 excerpt from 3.002 to 5.666 dB, while Brazilian percussion comparisons remained mixed and sometimes poor. A later conditional-residual experiment reduced false hat and ride output, but damaged some quiet notes and tails; it failed acceptance and was not installed. Hash-checked manifests, per-piece scores, omission challenges and same-position listening comparisons make those limitations reviewable. These are recording-specific adaptation results, not proof of universal accuracy.

For musicians and audio developers, the distinctive proposition is therefore a studio where listening, experimentation and evidence can inform one another. The roadmap aims at stronger attack/body/decay representations, multiple articulations, conservative absence handling and validation on distinct kits, players and rooms. Acoustician is already a substantial local application and research tool; perfect individual-piece recovery from arbitrary acoustic or electronic mixes remains an open objective.

Features and what a visitor can see or hear

Technical foundations

Swift/SwiftUI/AppKit provide the native studio, AVFoundation handles audio, and Accelerate supports spectral and fitting computation. The linked standalone package uses Swift tools 6.2 in Swift 5 language mode and a macOS 26 deployment floor. The shared STFT uses a 2048-point Hann window and 512-sample hop for the reference/reconstruction paths.

Local packages are SemanticCompute, Kalman (including the listed harmonics/DKF/prosody products) and FIRFilters. October evidence records SC 1.28.0 with 261 families; that is a dated verification, not a fresh 17 October server-capability claim. Neural helpers use local Python/PyTorch environments. The experimental custom Core ML query path has a corrected versioned stereo/raw-complex contract, but no quality-trained custom model is supplied by the documented training workflow.

The immutable-acquisition/provenance model distinguishes original audio, estimates, remainder and explicitly synthetic derivatives. Separation metadata and model receipts record the operation/settings rather than treating a UI confidence value as measured correctness.

Maturity and dated evidence

Roadmap and limitations

• Separate attack, body and decay evidence; learn multiple articulations and dynamic levels.
• Develop presence/absence/uncertain states using omitted-piece negatives and quiet positives, without cutting tails or ghost notes.
• Qualify machine-specific waveform matching for offsets, gain, polarity, collisions and hat chokes.
• Improve fine kit taxonomy beyond LarsNet's five groups, including individual toms and cymbals.
• Expand evaluation to independent kits, sessions, players and rooms; acoustic, electronic and Brazilian sources need distinct evidence.
• Evaluate phase-capable reconstruction and listening quality once instrument identity improves.

Existing limits include overlap, bleed, weak/noisy pieces, dependent reference examples, analysis/model resource bounds and optional external-model installation. A stem sum matching the input is a conservation check, not proof of correct instrument assignment. The event fitter stays optional because several Brazilian comparisons regress.

FAQs

Does Acoustician recover every drum perfectly? No. Broad neural groups, discovered timbre clusters and reference-conditioned pieces are different outputs. Quiet hits, overlap and tails remain active research problems.

Can I supply my own kit stems? Yes, the reference workflow accepts isolated instrument stems from the same kit/session. Overheads, room channels, full mixes and duplicate bounces should not be treated as isolated instruments.

Is the result generated audio or separated source audio? Reference fitting uses source masks plus a remainder. Synthetic restoration is a separate explicit choice and creates a derivative. The original source remains authoritative.

Does it work locally? Most analysis and installed-model processing is local. Models/dependencies need provisioning, and optional ShazamKit recognition needs network access and entitlements.

Can I download a released version? A local installed macOS app is verified. No public Acoustician download was verified for this intake. The linked build requires macOS 26; source-level iOS/AUv3 work is not a confirmed released product.