Vocello 2.3.0

Vocello 2.3.0

Released 2026-07-31. Vocello 2.3.0 ships a new generation of model packages that are
smaller on disk and lighter in memory at verified equal quality, makes warm generation
measurably faster, and hardens model delivery on both platforms. Everything still runs
locally on Apple Silicon; no audio leaves the device unless you export it.

Upgrading from 2.2? Drag-install the new DMG as usual. Your installed models, history,
and saved voices carry over. After updating, Settings → Model downloads shows
Update available on installed models: one click refreshes each package to the new
2026.07.26.1 artifacts (smaller download, shared components are reused, nothing else to do).
The old artifacts keep working until you update them.

Download v2.3.0 · All releases

Requirements

• macOS 26.0 or later
• Apple Silicon Mac
• 8 GB RAM minimum (16 GB+ recommended for Quality variants and long projects)
• ~6 GB free disk for the three recommended Speed (4-bit) model variants (was ~7 GB);
  ~15 GB if you also install Quality (8-bit)
• Internet for one-time model downloads from Hugging Face
• iPhone: the public TestFlight beta is live (join link);
  build 22 carries this release.

Headline themes

• Smaller, lighter models. All six model packages were re-published (artifact version
  2026.07.26.1) with the talker's text-embedding table quantized to 8-bit. The three Mac
  Speed packages now total about 6 GB (down from about 7 GB), and peak generation memory on
  the 8 GB floor drops by roughly 280 MB. Output quality was verified before promotion with
  a fixed-seed four-arm comparison battery and the standing automated audio-QC gates; the
  full measurement record is in benchmarks/OPTIMIZATION.md §N.
• Faster warm generation. The per-frame code-predictor pass is now compiled, worth
  8–11% warm throughput across modes and lengths, and stream-chunk materialization is
  pipelined off the token hot path with byte-identical output. Net of the new artifacts,
  warm generation runs up to about 10% faster than 2.2.2 on the canonical Mac mini M2
  (8 GB), and every mode remains faster than realtime there.
• Model delivery hardening. Installed models detect stale artifacts and offer an
  authenticated in-place update on both platforms. iPhone model downloads are pinned to
  Wi-Fi (previously, Wi-Fi Assist could silently reroute multi-gigabyte downloads over
  cellular at sub-MB/s rates). On the Mac, switching models no longer reloads the shared
  682 MB speech tokenizer when the bytes are identical, cutting about a quarter second
  from the first chunk after a switch.
• Clone and long-form polish. Clone reference encoding appends a half second of
  trailing silence for a steadier clone onset. On iPhone, selecting a saved clone voice
  primes the clone model in the background so the first take starts sooner. Long-form
  assembly records a warn-first advisory when a segment boundary jumps unexpectedly, and
  custom delivery text that requests an explicit duration now gets an advisory explaining
  that delivery text cannot control take length.

Fixes

• iPhone: updating a model's artifacts no longer strands the download ledger on a
  superseded request.
• iPhone: the clone prime no longer spikes resident memory (previously up to ~5 GB on a
  supported device) and a completed model load is no longer discarded when its task is
  cancelled late.
• macOS: repeated model switches reuse the byte-identical speech tokenizer instead of
  re-reading it from disk.

Under the hood

Every benchmark take now carries a typed quality identity (audio QC, prosody, ASR, and
long-form continuity gate verdicts) in the published history schema, and the quality
registry runs on the shipping generation path. This is QA infrastructure rather than user
behavior, but it is the evidence backbone for the claims above.

Evidence

Recorded per the standing release-candidate procedure (docs/reference/macos-release-qa.md):

• macOS smoke lane: macos-xcui-smoke-20260731-190550-a314890e — PASS, all seven
  journeys (navigation/readiness, Custom generation + History, cancellation, recording,
  library surfaces, long-form project, line batch). Long-form segment end-footprint growth
  first→last −6.8%: flat.
• Engine regression bench: label release-QA-2.3.0
  (macos-engine-20260731-192156-208e2873, schema v3) — clean-source run on the release
  commit, audio QC clean with zero warnings, warm throughput within noise of the
  2026.07.26.1 promotion-battery cells (custom warm about 1.2× realtime medians on the
  canonical Mac mini M2 8 GB).
• Artifact promotion evidence: the 2026.07.26.1 packages passed the fixed-seed
  four-arm promotion battery (13 takes per arm, all QC-clean, RTF parity within noise on
  medium/long cells) and the fail-closed catalog contract
  (model_catalog_contract.py validate --require-complete) before upload; measurements in
  benchmarks/OPTIMIZATION.md §N.
• Deferred, tracked: the fresh characterization-fixture rebind and device memory
  re-qualification on the new artifacts are mid-flight (six of nine control lanes banked,
  all PASS) and continue after this release; state in docs/development-progress.md.
  Deterministic release gates are the packaging authority and ran in the managed release
  process on the tag per the schema-v2 release-evidence contract.

TestFlight — What to Test (build 22)

Build 22 carries all of 2.3.0 plus two fixes that specifically matter on iPhone. Paste
this section into the build's Test Details in App Store Connect when distributing.

• Model updates. If you installed models with an earlier build, Settings shows
  "Update available" on them. Run one update: the download should complete, and
  generation should work normally afterward with the same voices.
• Wi-Fi-only downloads. Model downloads now run only on Wi-Fi. On cellular they wait
  instead of silently crawling at low speed. Try starting an update on cellular and
  confirm it clearly waits for Wi-Fi rather than hanging ambiguously.
• Faster first clone take. Select a saved clone voice, then generate: the first take
  should start noticeably sooner than on build 21, and the app should not be terminated
  by the system while preparing the voice (a build 21 memory issue this build fixes).
• The usual pass. One short take in each mode (Custom, Design, Clone), one long
  script (past 900 characters) as a project, and one mid-take cancel: cancel should
  reset cleanly with nothing added to History.

Report anything odd through TestFlight feedback, including the moment it happened so it
can be matched with diagnostics.

Known limitations

• macOS 15: stay on v1.2.3.
  No 2.x backport planned.
• Long-form scale: flat-memory behavior is validated up to twelve-segment projects
  (about ten minutes of joined audio); larger audiobook-scale evidence is ongoing work.
• Single takes and normalization: the long-form planner's spoken-text normalization does
  not yet apply to short single takes.
• Voice Cloning: only clone voices you own or have permission to use.

Install

1. Download Vocello-macos26.dmg from the v2.3.0 release.
2. Open the DMG and drag Vocello.app to /Applications.
3. Open Vocello → Settings → Model downloads → install the recommended Speed packages
   (~6 GB total).
4. Generate from Custom Voice, Voice Design, or Voice Cloning.

The DMG is Developer ID signed and notarized with a stapled ticket. release-metadata.txt on
the release records build provenance.