CASE STUDY / REPOSITORY-VERIFIED

beatrice

A desktop instrument that turns beatbox audio into an editable arrangement.

project
beatrice
what
A desktop instrument that turns beatbox audio into an editable arrangement.
stack
Rust · Tauri · React · WebAudio · WASM · SQLite
year / status
2026 / active
Treated capture of Beatrice showing a detected beatbox performance and its arrangement.

Beatbox in. A synth arrangement out.

TREATED REAL DEMO
01

problem

Beatrice takes a recorded or live beatbox performance and carries its rhythm through sound classification, tempo estimation, quantization, harmonic arrangement, playback, and export. The repository keeps offline, compose-quality analysis separate from a causal streaming detector because those paths have different timing constraints.

02

how it works

Beatrice architecture: React input reaches a shared Rust DSP crate through Tauri for offline analysis or WASM and an AudioWorklet for streaming analysis, then continues through groove, harmony, arrangement, playback, and export.
03

proof

  1. 01

    On AVP v4, the Gaussian model with MAP calibration reached 81.6% participant-wise accuracy across 9,357 held-out utterances from 28 participants.

    docs/avp-results-2026-07-15.md — Overall; Protocol
  2. 02

    A three-run final-build acoustic loopback measured 163.7–165.5 ms P95. The predeclared 60 ms live-synth gate failed, so Jam Mode shipped as visual capture instead.

    docs/latency.md — Phase 3 Task 6 re-measure
  3. 03

    The same beatrice-dsp crate supplies offline native analysis and the WASM StreamingDetector used inside the browser AudioWorklet.

    crates/beatrice-dsp/src/lib.rs; src/worklet/detector.worklet.ts; Cargo.toml
  4. 04

    The streaming regression gate compares seven deterministic fixtures and fails below 95% same-class onset matches within ±20 ms of the offline pipeline.

    crates/beatrice-dsp/tests/streaming_tolerance.rs
04

source

github.com/bobbyrathoree/beatrice