11 SEPTEMBER 2026 / DEVELOPMENT BETA

Tone Assist 0.4.1

An Apple Silicon guitar rig for macOS 14 and later, available as a standalone app, Audio Unit and VST3. The current installer uses ad-hoc development signatures and is not notarized. Public installers remain unavailable pending release qualification.

New in 0.4.1

A library that remembers. Gear and Presets retain your search, filters and List/Grid view, with locally saved favorites. Search across all six artist packs, narrow sounds by instrument and role, and show only presets whose capture is available offline. Gear instrument filters use supplied metadata; unknowns remain visible as Unspecified. Preset cards wrap properly at smaller window sizes.

Visible tone builds. The assistant shows preparation, planning and validation. Cancel a pending request or replace it with another; late responses cannot take over the new job. Complete proposals are retained natively and checked against the current rig revision before applying. Errors and conflicts preserve the sound you were playing.

Experimental audio recovery. The opt-in prepared-render path now guards processor failures and exposes a recovery state. This improves that development path; it does not qualify the complete audio callback or finish the lock-free audio migration. Sustained live-play and host testing remain required.

Introduced in 0.4.0

First sound, offline. Both installer formats include eight original Buffle virtual preamps, four synthesized cabinet responses and 24 complete rigs. They require no account or language model. These are original circuit designs, not measurements of branded amplifiers. The session room can also load these voices directly for live browser audition.

Three views, one rig. Studio, Pedalboard and Compact preserve the working chain. Hardware and Parameters tabs share their controls; a mini-chain keeps the selected device in context. Density, text size and the optional companion are remembered.

Experimental graph routing. Connect processor outputs and inputs to build nested splits and merges, with connection level, pan and polarity. The current limit is 16 processors and 64 connections. Drafts are validated before Apply; cycles, disconnected nodes and stale changes are rejected. Previous two-path sessions retain their original routing until you choose to convert them.

More specific planning. The local assistant can select cabinets and individual effect settings, respects explicit effect exclusions and narrows captures by musical role. Personal TONE3000 selection and permitted feeds remain available; unrestricted automatic online search is disabled. This free, open-source application works with its original library independently of the online service.

Measured boundaries. Pitch latency is now reported and parallel paths are aligned, but the current pitch algorithm still adds around 64 ms; a low-latency live mode remains unfinished. The embedded model is being evaluated on an M4 Max with 64 GB. Recommend 16 GB or more for this configuration until smaller-machine tests pass; M1/8 GB performance is not yet qualified.

Support and recovery. Diagnostics offer a bounded, redacted preview before copying. Release helpers verify their containing bundle and ignore development executable overrides. Live browser failures return to the original native render with a retry path.

Continuing from 0.3.0

Offline planning. The Mac app, AU and VST3 bundle a local AI runtime. Assistant setup downloads and verifies the Qwen3.5 4B model; the offline installer includes it. Setup supports progress, pause, resume and a test request. Built-in recipes remain available. TONE3000 captures are obtained separately.

Legacy preset review. Inspect POD HD and Line 6 L6T/X3 files, review source gear and settings, select replacement captures and effects, and arrange the reconstructed rig across two paths. Export an inspection report or save the resulting rig. Source parameters are not automatically translated into unrelated NAM controls. Axe-Fx Standard/Ultra support is inventory-only.

The session room. Each guitar and bass stem exposes the per-take chain used for its original native render. Load your own compatible 48 kHz NAM to audition the original DI through a serial browser chain. Supported controls, bypass and order affect live audio. A/B shares a transport; live output requires manual level matching. Captures remain local and are not uploaded.

Hardening. Model inputs are checked before native or WebAssembly allocation. Live demo preparation cancels cleanly when playback, seeking or preset selection supersedes it. Inspector controls retain keyboard focus, and touch targets remain usable on narrow screens.

Audio recovery. A --no-audio standalone launch opens the editor without automatically opening the saved interface. Startup recovery stays armed until audio setup finishes. A CoreAudio delay observed during interface restoration remains part of live-play acceptance.

The complete rig workflow

Describe a sound, inspect available TONE3000 captures, build and edit the chain, then save the rig. Hardware views show real processing controls and captured variants. A chromatic tuner, eighteen built-in effects, cabinet IRs, undo/redo and two complete A/B rig memories are available. A/B rig memories do not automatically match loudness.

Existing split routing retains its original source/destination behavior. The new graph editor provides explicit connections for more complex rigs. The assistant currently constructs serial chains and preserves supported existing routing during refinements; fully graph-aware generation remains in progress.

A fixed NAM capture does not recover hardware knob positions that were never recorded. The editor provides input drive, output trim, wet/dry mix and separate EQ, plus actual captured variants when available. Pitch effects currently add about 64 ms of wet-signal delay.

Quick Capture

The standalone workflow provides curated hardware/interface profiles, manual entry, wiring preparation, verified official NAM stimulus, guarded recording. Managed trainer setup is paused before downloads: current NAM 0.13 requires PyTorch Lightning 2.6.1 or earlier, while checkpoint-loading security fixes require Lightning 2.6.6 or later. Unsafe or incompatible installed trainers are rejected before import. Training needs a supported dependency set; physical capture remains unqualified end to end. This does not affect NAM playback or offline AI. If the official stimulus download is unavailable, select its WAV manually.

Measured QA uses independent recordings, alignment and level matching. Correlation and residual error are measurements of that comparison, not a percentage of perceptual identity. Physical cabling, analog levels and hardware mixer settings need user calibration.

What the browser demonstrates

The listening room uses thirteen authored DI riffs, including separate crunch, ambient clean, octave-fuzz breakdown and lead examples. The preset library selects performances by role. The session room uses the three exported songs, with guitars and bass rendered through Tone Assist. Existing render provenance retains the engine version that actually produced each file.

Original tones are rendered natively. Live WebAssembly NAM supports a bounded WaveNet/A2 subset and serial gate, compression, drive, EQ, chorus, delay and an approximate room reverb. This is not the complete native plugin embedded in a webpage. MIDI-guitar stems are audio prints.

Release boundaries

This is a development beta, not a finished public release. Developer ID signing, notarization, clean-machine installation, sustained live-play checks and physical hardware capture acceptance remain release gates. Complete DAW session save/reopen needs host-specific verification.

The GuitarGeek library, exact automatic legacy-tone conversion, scenes, dry-input audition looping, automatic live loudness matching and coordinated updates remain unfinished. Source photographs and metadata are creator-supplied; other device surfaces are representative. Artist names describe inspiration, without endorsement or a claim of exact album recreation. The installer includes only original Buffle model assets; public artist recipe packs do not redistribute third-party NAM captures.

Independent, and still taking shape.

Built with Neural Amp Modeler and the TONE3000 library. This is an independent Buffle Audio project. Hardware names belong to their respective owners.