SPB Git

Repository Public License

RELEASE-REPORT.md 11.3 KB

# ORIX Audio 1.0.0 (1) — Release report

Date: 2026-11-20 · Bundle ID ai.spboucher.orix.audio · Team 3YM54G49SN

# Summary

ORIX Audio is implemented, builds for devices, and passes its automated suites in the simulator and on the Mac. It is not release-ready for the App Store: no physical iPhone or iPad was available, so the release gate "background recording is physically tested" is unmet. With the owner's explicit authorization on 2026-11-20, the repository was published and build 1.0.0 (1) was uploaded to TestFlight for internal testing — which is now the way to run that device test.

# Implemented features

Area State
Recording engine (tap → wait-free rings → writer thread + analysis thread), pause / resume, markers, six presets Implemented, tested in the simulator through a file-fed input
Crash-safe capture master (append-only CAF), 2 s checkpoints, disk-space watch, verified finalization, recovery UI Implemented and tested, including a real kill-during-recording in the simulator (33 s recovered)
Background / locked-screen recording (UIBackgroundModes = audio), interruptions, route changes, media-services reset Implemented per Apple's APIs — not tested on hardware
Live Activity and Dynamic Island with pause / resume and marker buttons Implemented, compiles and ships in the archive — never seen running
Frequency-coloured waveform (4 styles), spectrum analyzer, live and file spectrogram (tiled, 6 palettes) Implemented and tested
Analysis laboratory: time domain, spectral, BS.1770-4 loudness, LRA, true peak, speech-activity estimates, findings Implemented and tested against reference signals
On-device transcription (SpeechTranscriber, on-device-only legacy fallback), editing, search, manual speakers Engine verified on macOS; UI verified in the simulator with engine output loaded
Meeting intelligence (Foundation Models): grounded summary, key points, actions, decisions, chapters, Q&A Engine verified on macOS; cannot run in the iOS 26.5 simulator
Audio and video import (Files, Photos), metadata, bad-file handling Implemented and tested (Photos picker itself not exercised)
Non-destructive editor with undo / redo, enhancement chain, merge, save as new Implemented and tested
Playback (speed, loop, markers, Now Playing), library (folders, favourites, trash, search incl. transcripts) Implemented and tested
ORIX Actions (workflow builder, batch runner, triggers) Implemented and tested
Export: WAV, AIFF, CAF, ALAC, AAC, TXT, MD, SRT, VTT, CSV, JSON, PDF, waveform and spectrogram PNG Implemented and tested. MP3 is not offered (iOS has no MP3 encoder)
English and French-Canadian localization (629 strings + purpose strings) Complete; French verified by a UI test and screenshots

# Test results

Suite Command Result
Core (DSP, analysis, capture, editing, export) Scripts/test.sh core 83 tests, 14 suites — all passed (macOS 27, Apple M5 Max)
App (capture pipeline, finalization, recovery, import, library, actions, exports) Scripts/test.sh app 21 tests, 4 suites — all passed (iOS 26.5 simulator, iPhone 17 Pro)
UI (first launch, record flow, library, panels, editor, actions, settings, French, accessibility audit) Scripts/test.sh ui 9 tests — all passed (same simulator)

Bugs found by testing and fixed during development: AAC encoding failed for bit rates the encoder does not support at a given sample rate (the capture master was kept, as designed); the live waveform stopped redrawing; generated notes sometimes came back in English for a French transcript; starter workflows were missing in a fresh library.

# DSP accuracy

Full detail in docs/reports/dsp-validation.md. Highlights, all from the test suite:

  • Bin-centred sine amplitude reads within ±0.001 for six windows and four FFT sizes; flat-top off-bin error < 0.02 dB.
  • Dominant frequency within 0.5 Hz at FFT 8192 (61.7 Hz to 9.9 kHz); PSD integrates to the noise variance within 3 %.
  • Spectrogram burst timing within two hops; chirp ridge within two bins.
  • BS.1770-4: K-weighting matches the standard's 48 kHz coefficients to 1e-8; 997 Hz full-scale sine reads −3.01 LKFS at 44.1, 48 and 96 kHz; synthesized EBU Tech 3341 cases 1–5 within ±0.1 LU; EBU Tech 3342 cases 1–2 within ±1 LU; inter-sample peak found within 0.3 dB. Verified against synthesized signals, not the official EBU files.

# Transcription results

Scripts/bench-transcription.sh, SpeechTranscriber on macOS 27 (not iPhone), synthetic speech from macOS voices:

Locale Voice Words WER CER Speed
en-US Samantha 163 1.84 % (3 substitutions) 0.94 % 82× real time
en-US Daniel 163 1.84 % (3 substitutions) 0.94 % 84× real time
fr-CA Amélie 182 0.55 % (1 substitution) 0.09 % 92× real time
fr-CA Thomas 182 0.00 % 0.00 % 116× real time

Clean synthetic speech is an upper bound; real rooms will be worse. No accuracy figure exists for iPhone hardware. Foundation Models notes on the same transcripts (macOS 27): every generated point traced to a transcript line (7/7 in English, 8/8 in French); out-of-transcript questions were refused. See Benchmarks/results/.

# Background recording test evidence

None on hardware. xcrun devicectl list devices showed simulators only. What exists instead:

  • The capture pipeline runs under test with pause / resume, and the bytes written equal the frames accepted.
  • Killing the app during a recording left a valid master that the next launch recovered (observed, simulator).
  • The archive contains UIBackgroundModes = [audio] and the Live Activity extension.

To close this gate, on an iPhone: start a recording, lock the phone, leave it in a pocket for 30 minutes, place a call to it midway, unlock, stop; then check the duration, the interruption marker and the Lock Screen activity.

# Device compatibility

iOS / iPadOS 26.0 or later, iPhone and iPad. Tested in the iOS 26.5 simulator (iPhone 17 Pro, iPad Pro 13-inch M5). Transcription needs a device and language the system supports on-device; summaries need Apple Intelligence. Both degrade gracefully and never block recording or analysis.

# Performance measurements

Mac, Apple M5 Max, single thread, release build (Benchmarks/results/dsp-throughput.md): FFT 4096 in 2.5 µs; a live spectrum frame in about 16 µs; 60 s waveform summary in about 12 ms; full analysis of a 10-minute stereo file in about 2.6 s; AAC export of 10 minutes in about 2 s. No iPhone measurements; no battery or thermal data.

# Known limitations

See README section 32. The main ones: no hardware validation; no speaker diarization; mono-mix waveform; no MP3 export; Foundation Models unverified on iOS; the legacy speech permission prompt uses Apple's stock wording; Metal was evaluated and not used; 57 notes remain in the automated accessibility audit (docs/reports/accessibility-audit.txt), mostly contrast readings under the translucent tab bar and single-line numeric readouts that do not grow with the largest text sizes.

# Repository

Local repository ~/Desktop/Cluster/apple/ORIXAudio, branch main
Remote https://git.spboucher.ai/orix-audio (clone: gitsrv:srv/git/orix-audio.git)
Visibility Public. git.spboucher.ai has no private mode; the owner chose to publish there (2026-11-20) rather than keep the repository off the site.
Commit built and uploaded 69218acef60891cae423bd53b84218f6be7da32b — verified identical locally and on the remote. This report was updated in a later documentation-only commit (git log -1).
README on the site Rendered and checked in Safari: 36 images, 35 confirmed loaded (the last is the lazily loaded footer logo, served correctly); badges and diagrams are local SVG files. The site limits each visitor to 200 requests per minute, so several rapid reloads of this image-heavy page can temporarily show broken images.
Secret scan 176 tracked files scanned for keys, tokens, passwords and personal data: none found. Build products, archives, IPAs and credentials are ignored.

# Build

Version 1.0.0 (1)
Toolchain Xcode 27.2, Swift 6.4.2, iOS 27.2 SDK, deployment target 26.0
Archive Scripts/release.sh archive — succeeded, signed for team 3YM54G49SN, debug hooks confirmed absent from the binary
Export Scripts/release.sh export — succeeded, dist/ORIX Audio.ipa, 4.9 MB
Validation Scripts/release.sh validate — VERIFY SUCCEEDED with no errors (rebuilt from commit 69218ac before validating)
Upload Scripts/release.sh upload — UPLOAD SUCCEEDED, delivery 8d038265-130a-4b8f-8adf-29e0d7490552

# Apple Developer status

Verified in Safari and through the API on 2026-11-20: signed in, team 3YM54G49SN, Apple Developer Program (individual), renewal 5 September 2027, both App IDs registered. Details in docs/release/apple-developer-verification.md.

# App Store Connect status

App record created on 2026-11-20 through the owner's signed-in session, with the owner's authorization: ID 6837544792, name "ORIX Audio" (the name was available), SKU orix-audio-ios-001, primary language English (Canada). The listing itself is still empty: metadata (English and French-Canadian) is ready in docs/release/app-store.md but has not been entered, and no App Store screenshots were uploaded. The account shows an EU trader-status notice that only the account holder can resolve.

# TestFlight status

Stage State (read from the App Store Connect API, 2026-11-20)
Uploaded Yes — build 1.0.0 (1)
Processing Finished — VALID, minimum OS 26.0, expires 2027-02-18
Export compliance usesNonExemptEncryption = false, no question pending
Available to internal testers Yes — IN_BETA_TESTING, in the existing internal group "Test" (all builds), one tester (the account holder, invitation state INVITED)
Available to external testers No — READY_FOR_BETA_SUBMISSION; not submitted for Beta App Review
App Store Not submitted

TestFlight "What to Test" notes were not entered. The build has not been installed on any device yet.

# Outstanding approvals

Granted on 2026-11-20: publish the repository on git.spboucher.ai; create the App Store Connect record; validate and upload to TestFlight.

Still requiring separate authorization: external TestFlight testing (Beta App Review), entering the store listing, any App Store submission.

# Remaining work

  • Device validation: background / locked recording, Live Activity, real microphones (mono and stereo), Bluetooth and wired route changes, phone-call interruption, a multi-hour recording, battery impact.
  • Confirm on-device transcription and Foundation Models on an iPhone; confirm whether SpeechTranscriber needs the Speech Recognition permission on iOS (the app requests it only for the legacy recogniser).
  • Work through the remaining accessibility audit notes; a VoiceOver pass by a person.
  • Per-channel waveform; speaker diarization if Apple ships an on-device API.
  • App Store screenshots at the required sizes; App Review submission (separate authorization).
Files main
Menu

Theme