frames sampled over 2.12 seconds
Benchmarks · measured August 26, 2026
Numbers with methodology attached.
These are engineering probes, not a claim that every library behaves identically. Hardware, drive speed, file size, codec, background analysis, and library topology all affect real-world performance.
7.82 smedian synthetic discovery + hashing pass across 10,000 small nested files
17.0 msp95 frame interval across a 127-frame scroll-idle renderer probe
0%frames above 33 ms in that renderer probe; 2.12-second observation window
2 KBraw 512 × float32 semantic vector payload per analyzed audio hash
10,000-file discovery probe.
Three cold database runs on an Apple M4 Mac mini with 16 GB RAM. Fixture: 10,000 synthetic 4 KB files across five directory levels, total 39.06 MB.
Run 1
5.93 s
Run 2 · median
7.82 s
Run 3
13.07 s
Average: 8.94 seconds; reported hash throughput: 1,243 files/second. The spread is shown rather than hidden.
Linear planning estimate
Scale the median discovery pass
A linear projection is 78.2 seconds.
This is arithmetic, not a measured result at the selected scale. Real audio is much larger than the synthetic fixture and drive behavior is not linear at every scale.
Important limitation
The synthetic generator uses audio extensions but not valid encoded audio. The result measures directory traversal, stat work, and hashing attempts—not full metadata ingestion or DSP/embedding analysis. We will not present it as “10,000 samples fully analyzed in 7.82 seconds.”
Renderer frame-pacing probe.
A short instrumented session captured a scroll-to-idle transition in the production renderer on the same M4/16 GB machine.
millisecond median frame interval
millisecond p95 frame interval
long tasks and frames above 33 ms
Stop reason: scroll_idle. Input latency was not exercised, no analysis jobs were running, and the window is too short to represent a full production session. It is evidence for one interaction, not a blanket performance guarantee.
Storage math you can audit.
Sample stores one 512-dimension float32 vector for each unique analyzed audio hash.
| Unique analyzed files | Raw vectors | What is excluded |
|---|---|---|
| 10,000 | 19.5 MiB | SQLite rows, ANN index overhead, DSP metadata, waveforms, model files, filesystem allocation, and duplicate paths. |
| 50,000 | 97.7 MiB | |
| 100,000 | 195.3 MiB | |
| 250,000 | 488.3 MiB |
Formula: unique hashes × 512 dimensions × 4 bytes. Duplicate file paths can point to the same content hash and therefore need not duplicate the raw semantic vector.
Architecture choices aimed at large libraries.
Performance is a system property, not one headline query time.
Stat-first rescans
Unchanged files can be reassociated from a path/size/mtime cache rather than rehashed on every root scan.
Hash-based identity
Analysis follows file content across known paths. This supports path healing and avoids duplicating analysis for identical content.
Approximate vector index
HNSW nearest-neighbor retrieval narrows candidates without exhaustively comparing every 512-dimension vector per search.
Separated workers
Scanning, DSP, embedding, vector maintenance, and renderer work are deliberately separated so background work need not own the interface thread.
Confidence gates
BPM/key filters exclude ineligible values instead of surfacing false precision. One-shots do not inherit loop assumptions.
Bounded work
Searches, catch-up batches, watch events, and UI prefetches use explicit limits to reduce unbounded latency as the catalog grows.
What we will measure next.
A credible benchmark page should expose the gaps.
Valid-audio cold analysis
10K and 50K mixed WAV/AIFF/FLAC fixtures, reporting DSP and embedding phases separately.
Search latency distribution
p50/p95/p99 across text, filtered text, similarity, and blend queries at multiple catalog sizes.
Session endurance
Memory, frame pacing, input latency, and search responsiveness during background analysis and external-drive changes.
One-time purchase · $129 launch price
Buy the workflow—not an unqualified benchmark headline.
Sample searches and analyzes locally on Apple Silicon. Your audio stays on your Mac. Use it offline after activation, drag results into any DAW, and request a refund within 14 days if it does not improve your workflow.