Vocello 3.0.0

Vocello 3.0.0

Released October 2026. Vocello 3.0 brings a redesigned Mac app that runs its engine inside the
app and updates itself, and a new command-line tool, vocello, installed with a signed package and
built for scripts and AI agents. Everything still runs locally on Apple Silicon; no audio leaves
your Mac unless you export it. The iPhone app's 3.0 version, on the same engine, follows in the
coming days through TestFlight and the App Store.

Upgrading from 2.4? Install the new DMG as usual. Your History, saved voices and settings
carry over. 3.0 uses updated model packages, which now include the audio-marking weights, so the
model manager offers an update for each model you installed with 2.x. From 3.0 on, the app
updates itself: choose Vocello > Check for Updates.

Download v3.0.0 · All releases

What's new on the Mac

• The app updates itself. Vocello looks for new versions with Sparkle, on its own schedule (a
  setting you can turn off) or when you choose Vocello > Check for Updates, and installs one when
  you accept it. Every update is signed, and the app checks that signature before installing.
• A redesigned Mac app. The Mac app adopts the iPhone app's visual language: a new window
  shell with four sidebar destinations, the Studio canvas with the phone's setup chips, dock and
  mode colors, and Settings organized in categories. The Custom Voice mode is now called
  Built-in Voice on both platforms.
• The Mac engine runs inside the app. The separate XPC engine service is gone. The Mac app
  hosts the engine in-process, as the iPhone app does, so one store, one lifecycle and one memory
  policy serve both: the engine's own pressure responder trims and unloads, and an idle model
  unloads according to the memory tier.
• Protect accepted audio when an operation fails. Failed generation cleans up its own
  staging files rather than deleting an existing output destination. Saved Voice transactions
  retain recoverable backups when restoration fails, and app/CLI store mutations use a shared
  cross-process lock. Contention reports a busy store instead of attempting competing recovery.
• Safer long-form updates. Segment replacement stages and checks candidate audio and the
  project manifest before acceptance. Recovery records protect interrupted updates, previous
  joined outputs retain their History ownership, and reused segments count once in progress.
  Continuing generation remains session-scoped; this is not automatic background resume.
• Keep output language separate from reference language. Clone reference metadata no longer
  silently selects the generated language. Explicit output-language choices remain authoritative.
  macOS Voice Design snapshots language, delivery, variation and pinned seed before asynchronous
  startup, so later UI edits cannot alter the submitted request.
• More dependable completion and recovery. History-save failures remain distinct from
  successful synthesis, with recovery/export paths. Stale completion callbacks cannot take
  over a newer Mac generation's player. Explicit Play after streaming follows the heard
  position rather than assuming all queued audio was played.

What's new: the vocello command line

• Install it with a signed package. Vocello-macos26-cli.pkg installs vocello in
  /usr/local/lib/vocello and links /usr/local/bin/vocello and its zsh completion. It is signed
  with Developer ID and notarized. sudo sh /usr/local/lib/vocello/uninstall.sh removes it and
  leaves your models, voices and History untouched. If you prefer, copy the Vocello CLI folder
  from Vocello-macos26-cli.dmg instead.
• The Mac app's features, from the terminal. Generate in all three modes, long-form and line by
  line included; enroll, review and delete saved voices; install, verify and remove models; read and
  export History; read and change settings. The CLI uses the app's models, saved voices and
  settings, so both see the same library. CLI takes are saved as files and do not appear in History
  yet; that comes in a 3.0 update.
• Built for scripts. --json and --ndjson give every command a stable machine contract: one
  envelope, typed error codes, documented exit statuses and progress events. vocello capabilities --json describes every command and option, and vocello run --request takes JSON requests.
• An MCP server for AI agents. vocello mcp lets Claude Code, Claude Desktop and other MCP
  clients list voices and models, generate speech with progress and cancellation, enroll and delete
  voices, and export takes. Cloning needs consent in each request, and files stay in Vocello's
  output folder unless you start the server with --allow-dir. vocello mcp --print-config claude-code prints the command that adds it.

What's new in both apps

• Generated audio is marked as AI-generated. Every generated WAV, in the apps and the CLI,
  carries an inaudible AudioSeal watermark with a fixed Vocello payload and a machine-readable
  provenance note (a LIST/INFO chunk naming the generator, version, engine, model, mode and creation
  time). The watermark weights download with each model package. Live preview audio heard before
  the file is written is not marked. Publishing a cloned real voice still calls for your own
  disclosure that it is AI-generated.
• The interface speaks ten languages. The Mac and iPhone interfaces are available in English,
  French, Spanish, German, Italian, Brazilian Portuguese, Simplified Chinese, Japanese, Korean and
  Russian, chosen in Settings independently of the speech language.
• Pin a take's seed. Every finished take records the seed it used. Pin it from History;
  Studio shows the pinned seed and can unpin it, and later generations with the same settings
  reproduce the take. Leave it unpinned for a fresh take each time.
• Delivery choices for curated cloned voices. A cloned voice with a curated emotion reference
  bank offers a delivery choice, each delivery backed by its own verified reference clip rather
  than an engine instruction.
• Clearer delivery expectations. The picker offers Neutral, Calm, Whisper and Sad as its
  measured delivery group; Happy, Angry, Fearful and Surprised remain directional hints, not
  promises of exact emotion. Excited has folded into Happy and Dramatic has been retired from
  the current roster. Existing raw delivery instructions remain verbatim for reproducibility.
  Automatic acoustic analysis is not proof of listener-perceived improvement.
• Built-in Voice and Voice Design takes read the whole script first. As in the reference
  Qwen3-TTS, a take now sees its full text before it speaks, instead of receiving it word by word
  while the audio streams, so a take that opens with a sigh no longer loses its script. In an
  automated screen, with no listening, takes with wrong or missing words fell by about half.
  Long-form projects keep the previous layout. A seed pinned on an earlier build no longer
  reproduces its take exactly.
• Review imported Clone references before saving. iPhone Studio's Reference Clip picker
  includes Import audio file. Imports use permanent voice enrollment with on-device
  transcription, editable text, explicit audio-only confirmation when needed, and Save Voice
  confirmation. A delayed transcription cannot overwrite manual edits. macOS enrollment also
  retains typed transcription outcomes and reference-language metadata.

iPhone

The iPhone app's 3.0 version brings the same engine and the shared changes above. It follows in
the coming days through the public TestFlight beta
(join link) and then the App Store; until then the
beta runs version 2.4.

• Design & Clone Export on iPhone, not yet available. The official iOS app is prepared to
  offer one non-consumable purchase (US base price $19.99, with Apple-managed regional prices)
  that unlocks outward Design and Clone audio export. Generation, listening, History, voice
  enrollment and Built-in exports stay free, and macOS and the CLI remain unrestricted. The
  product's App Store details are entered and local purchase testing passed. The processed-build
  purchase check, Apple's paid-app agreement and App Store review are still pending, so this is
  not an availability claim.

Requirements

• macOS 26.0 or later, Apple Silicon, and at least 8 GB unified memory (16 GB or more recommended
  for the Quality models and long projects).
• An administrator password to install the CLI package.
• Internet access for the one-time model downloads from Hugging Face; model files are not bundled.
  Check the displayed model sizes and available storage before installing. Generation runs locally
  once a model is installed.
• Microphone access for recording, and on-device Speech Recognition for automatic transcription.
  Manual transcript review stays available when recognition cannot finish.
• iPhone: iOS 26.0 or later on iPhone 15 Pro or newer.

Install

Download from the v3.0.0 release.

1. Mac app: open Vocello-macos26.dmg and drag Vocello.app to Applications.
2. Command line: open Vocello-macos26-cli.pkg and follow the installer, then run
   vocello --help. Or open Vocello-macos26-cli.dmg and copy the complete Vocello CLI folder
   to a folder you own, keeping its executable and resource bundles together. The installer only
   adds links in folders that root alone can change: if your /usr/local/bin belongs to your user
   (as some older Homebrew setups leave it), it installs the command but adds no link there, so run
   /usr/local/lib/vocello/vocello or link it into your path yourself.
3. AI agents: with Claude Code, run claude mcp add --scope user vocello -- /usr/local/bin/vocello mcp.
4. Open the model manager (Settings → Models & Files, or vocello models list) and install or
   update the models you use. Back up important exported audio and saved references before
   upgrading.

Evidence and limitations

Every 3.0 file is built by the release workflow from the tagged commit, only after CI has passed on
that exact commit. Both DMGs and the CLI package are signed with Developer ID, notarized and
stapled, attested, and checked again after upload; the CLI package is installed, upgraded and
uninstalled on a clean machine before release. Deterministic tests cover request preservation,
transactional failure recovery, cross-process exclusion, cancellation and the CLI's machine contract.
No new benchmark record accompanies these notes, so they make no speed claim over 2.4.0.

Known limitations: some French and Chinese takes fail the mandatory audio check, and Vocello
rejects them instead of saving them. Rarely, an English long-form take develops a long silent gap and
reaches its length limit without finishing; one retained French take shows the same pattern. The
cause is still being investigated. When a take is rejected, generate it again: Vocello never
silently regenerates, substitutes or accepts a failed take.
Long-form projects using the Consistent take variation can still stall, though far less often
than in earlier builds: a segment of about a minute may pause for several seconds or never finish.
Audio QC rejects that segment instead of saving it, and Resume retries it with a fresh seed. This
was not observed with Expressive, the default.

Take quality in 3.0 is not verified by human listening. Automated checks catch gross failures (silence,
a cut-off or unfinished take, missing words, the wrong language), but a take can still carry subtle
defects: a whispered or swallowed syllable, a short false start, an inserted vowel or over-linked French
liaison, an English-sounding accent in French, or delivery that misses the requested style. Generate
again for a different take. Generation now samples from full-precision (float32) scores, as Qwen's
reference implementation does, on a newer MLX whose sampler draws differently for the same seed, so
a seed from an earlier version does not reproduce the same take. As in the reference, the speech
detail stage keeps its own sampling whichever take variation is chosen, and the end of speech is
sampled like any other token, so takes can differ from earlier versions, Balanced and Consistent
takes most of all.

TestFlight: what to test (build 24)

For testers of the iPhone app's build 24 in TestFlight, once it is distributed:

• Exercise first launch, model install/cancel/relaunch/removal and offline generation on supported
  hardware; preserve existing personal models and voices.
• Generate Built-in, French Voice Design with a pinned seed, and transcript-backed Clone takes.
  Confirm selected language, saved transcript and reference identity are preserved.
• Import a test-owned reference without a text sidecar, review transcription or manual/audio-only
  fallback, save it, generate a Clone take, then remove only that test-owned voice.
• Generate a long-form project and regenerate one segment. Check the joined result, History and
  playback without losing the prior accepted output after cancellation or failure.
• Exercise Play/Pause, scrubbing, export, cancellation and History recovery. Report silence,
  unusually spaced words, language/pitch mismatch, a stuck player or any missing saved output.

Use TestFlight feedback only on an explicitly distributed candidate, or the
support page. Do not include private reference audio or transcripts
in public reports.