SPB Git

Repository Public License

performance.md 3.5 KB

# Performance Report

Date: 2026-11-20 · Tool: orix-bench (release build of the core package) · Machine: Mac, Apple M5 Max, macOS 27 · Threads: one

These are not iPhone measurements. They show what the algorithms cost on a fast desktop-class chip. An iPhone will be slower by a factor that has not been measured, and nothing here speaks to battery, heat or memory pressure on a phone.

Raw output: Benchmarks/results/dsp-throughput.md.

# Signal processing

Operation Result
FFT 1024 — window, transform, scaling 0.8 µs
FFT 4096 2.5 µs
FFT 16 384 12.8 µs
Live spectrum frame — 4096-point FFT, smoothing, peak hold, seven bands 15.1 µs
Waveform summary with band colours, 60 s of 48 kHz audio 14 ms (4433× real time)
Spectrogram, FFT 2048, hop 512, 60 s (5622 columns) 15 ms (3904× real time)
Rasterise a 512 × 512 spectrogram tile through a palette table 0.65 ms
BS.1770 loudness with true peak, 60 s mono 95 ms (634× real time)

Timings vary by a few percent between runs; an earlier run of the same binary gave 0.61 ms for the tile and 12 ms for the waveform summary.

# Whole files

Ten minutes of 48 kHz stereo pink noise, 24-bit source.

Operation Time Speed Output size
Full analysis pass (levels, loudness, spectrum, speech activity, waveform) 2.74 s 219× —
Export to WAV, PCM 16-bit 0.10 s 5739× 115.2 MB
Export to Apple Lossless 1.28 s 470× 104.6 MB
Export to AAC 2.11 s 285× 8.7 MB
Noise reduction + voice enhance + export to WAV 3.67 s 163× —

(Pink noise is close to incompressible, which is why Apple Lossless saves little here; speech and music compress much better.)

# Transcription and notes

Operation Result
Transcribe about 60 s of speech (SpeechTranscriber, model installed) 0.59–0.70 s, 82–116× real time
Generate notes from a one-minute transcript (Foundation Models) about 7–8 s
First-time speech model download several minutes, once per language

# What these numbers support

  • Canvas and Core Graphics instead of Metal. A spectrum frame costs microseconds and a spectrogram tile is rasterised once in well under a millisecond, against a 16.7 ms frame. See visualization.
  • Analysing every recording automatically. A ten-minute file takes under three seconds here.
  • Encoding on stop rather than while recording. Encoding is hundreds of times faster than real time, so finishing a long session is a short wait.

# Simulator observations

Not benchmarks, but worth recording:

  • In the iOS 26.5 simulator (Debug build) the analysis thread produced every expected waveform bucket and spectrogram column while capturing 48 kHz stereo and 22.05 kHz mono; this is asserted in the app tests.
  • No samples were dropped in any test run.

# Not measured

Item Why
CPU while recording on an iPhone no device
Battery drain per hour, screen on and locked no device
Memory over a multi-hour recording no device; by design only the waveform summary grows with duration (about 13.5 MB per hour at 48 kHz)
Frame rate of the live views on older iPhones no device
Long-file transcription throughput only one-minute samples were run
Thermal behaviour no device

No figure is given for any of these, and none should be inferred from the Mac results.

Files main
Menu

Theme