First-party evidence, not an independent benchmark. HighRoad publishes SongForge. The result below came from a signed App Store-configured candidate and is included so creators can judge the claim, not because it proves a general performance ranking.
Measured result
| Measure | Observed value |
|---|---|
| Wall-clock generation | 70 seconds |
| Delivered music | 133.2 seconds (2 min 13.2 s) |
| Observed throughput | 1.90 seconds of music per wall-clock second |
| Output | Stereo MP3, 48 kHz, 128 kb/s |
| Engine and model | Local ACE-Step 1.5 Turbo Q8_0 |
| Candidate | SongForge 1.0.0 (2), signed macOS build |
What was measured
The test used an instrumental dance-house / nu-disco prompt at 122 BPM. It timed the Generate action through completion after the app and bundled local engine were already running and healthy. The delivered file was checked as a stereo 48 kHz MP3. The test intentionally used a generic musical direction rather than an artist-imitation prompt.
This is local inference using the app’s compatible model set and Apple-silicon acceleration. Apple documents Metal as its graphics and compute API; that technical context does not validate this specific SongForge result. Read Apple’s Metal overview.
What this does not prove
- It is not a cold-start measurement.
- It is not a minimum-specification benchmark or a performance guarantee.
- It does not measure every model variant, song length, vocal language, export path or Mac configuration.
- It does not establish that one product is faster or higher quality than a cloud service.
Earlier short validation tracks (27–44 seconds) completed in 12–25 seconds on the same test environment. They are supporting observations, not a pooled benchmark. For a purchasing decision, test the exact model and song length you expect to use.
Why publish a single-run result?
“Fast” is not useful enough for a local creative tool. A local workflow trades a model download and machine requirements for private, repeatable generation without a hosted per-generation queue. Publishing one complete, bounded measurement makes the trade-off concrete while leaving the variables visible.
Product and release status
SongForge is in release validation. This measurement validates a local audio-generation run at level 4 (signed Release artifact on a physical target). StoreKit loading, purchase, restore, revocation and TestFlight distribution remain untested for the candidate, so it is not evidence of public-release readiness. See the product page for the current status.
Sources and correction policy
Result facts were reviewed against the current SongForge validation evidence on 18 July 2026. This note is updated when a material change affects the method or the claim. Report an error to support@highroadsoftware.com.