Overtone

Local, on-device text to speech for macOS, in the menu bar.

Paste text, press ⌘↩, hear it. Nothing leaves the machine: the speech model, the inference server and the voices all live inside the app bundle. No account, no API key, no network.

• Fast. Generation runs at roughly 0.1× real time on Apple Silicon; audio starts in under half a second and playback never waits for the stream.
• Natural. The text is normalized and chunked on sentence boundaries, so it reads like prose rather than a list of words. Why that matters.
• Scriptable. A loopback HTTP API on port 7789, plus a skill file that teaches a coding agent to answer out loud.
• Ten voices, twelve languages, adjustable speed, quality and volume.

Install

Download Overtone-1.0.0.dmg from the latest release, open it and drag Overtone to Applications.

The app is signed ad-hoc, not notarized, so Gatekeeper will refuse the first launch. Either right-click the app and choose Open, or clear the quarantine flag:

xattr -dr com.apple.quarantine /Applications/Overtone.app

macOS 14 or later, Apple Silicon.

Using it

The menu bar item opens a panel with the text to read, transport controls and a progress bar. The ? button opens a help page with every shortcut and the scripting API; esc returns.

Shortcut	Where	Action
⌥⌘P	anywhere	speak the selected text, or the clipboard if nothing is selected
⌥⌘.	anywhere	stop
⌘↩	in the panel	speak the text
⌘.	in the panel	stop

Voice, language, quality, speed and volume are saved as you change them. Only Static MLProgram restarts the speech server; everything else applies to the next utterance.

Reading the selection copies it first, which macOS only allows an app to do by driving the keyboard — so ⌥⌘P asks for Accessibility the first time. Decline it and the shortcut still works, just from the clipboard.

Shortcuts.app exposes Speak with Overtone, Speak selection with Overtone, Speak clipboard with Overtone and Stop Overtone. The overtone:// URL scheme takes text, and optionally voice and language:

open "overtone://speak?text=Привет&voice=F2&language=ru"

Control API

Bound to the loopback interface only.

curl -X POST http://127.0.0.1:7789/speak \
  -H 'content-type: application/json' \
  -d '{"text":"Привет! Это локальная озвучка.","voice":"M1","language":"ru"}'

voice, language, speed, total_step and max_chunk_length are optional and fall back to the current settings.

Endpoint	Purpose
POST /speak	synthesize and play, returns immediately
POST /stop	stop generation and playback
POST /pause	toggle pause, answers with the new state
GET /status	idle/speaking/paused, server state, elapsed and buffered seconds
GET /skill	the agent contract, as markdown
GET /health	liveness

curl http://127.0.0.1:7789/status
{"buffered":7.92,"elapsed":2.14,"error":null,"server":"ready","status":"speaking"}

The low-level streaming endpoint stays on http://127.0.0.1:7788/v1/audio/speech.

Letting a coding agent speak

skills/overtone-voice/SKILL.md is written for coding assistants: when to speak, how to call the API, and how to phrase text that sounds good out loud. The sparkles button in the panel copies it to the clipboard, and the app serves the same file at GET /skill, so an agent can fetch its own instructions. Install it as a personal Claude Code skill:

ln -s "$PWD/skills/overtone-voice" ~/.claude/skills/overtone-voice

The app bundles that file as a resource, so the document, the button, the endpoint and the installed skill can never drift apart.

Build from source

The repository carries the app; it does not carry the 400 MB of model weights and the compiled speech server. Take those from a release build:

brew install xcodegen
cp -R /Applications/Overtone.app/Contents/Resources/Runtime ./Runtime

xcodegen generate
xcodebuild -project Overtone.xcodeproj -scheme Overtone -configuration Release \
  -derivedDataPath .build CODE_SIGNING_ALLOWED=NO build
codesign --force --deep --sign - .build/Build/Products/Release/Overtone.app

Runtime/ needs supertonic-server (built from the rust-server crate of the Supertonic repository), onnx/ with the model files, and voice_styles/ with the voice JSONs.

Unit tests cover the pure logic — text normalization, silence trimming and the control-server HTTP parser:

xcodebuild -project Overtone.xcodeproj -scheme Overtone -configuration Debug \
  -derivedDataPath .build CODE_SIGNING_ALLOWED=NO test

To package a DMG the way releases are built:

./scripts/make-dmg.sh

Chunking and prosody

The speech server splits input into chunks and synthesizes each one as a separate utterance, so every chunk boundary costs the trailing silence of one utterance plus an inter-chunk pause. Small chunks therefore break a sentence into bursts of a few words separated by roughly a second of dead air.

max_chunk_length is an upper bound, not a target: the splitter already breaks on sentence boundaries. The default of 300 characters keeps one sentence per chunk for ordinary text, and the app sends silence_ms=0 so only the model's own sentence pauses remain. Measured on a 372-character Russian paragraph (M1, 8 steps):

max_chunk_length	chunks	audio	wall	pauses at boundaries
32	26	39.5 s	7.9 s	25 × ~0.75 s, mid-sentence
150	6	29.1 s	3.1 s	5 × ~1.15 s
300	4	28.2 s	2.6 s	3 × ~1.1 s, sentence ends only

Raising the limit past ~300 changes nothing but delays the first chunk, since time to first audio depends on the first chunk alone.

The player also drops the half second of silence every utterance starts with, keeping 20 ms of preroll — that is where the rest of the startup latency went.

Input is normalized before synthesis: markdown emphasis and links, list bullets and headings are removed, URLs are read as their host, and a hard-wrapped paragraph is rejoined into one sentence while a heading or list item is terminated as its own. Without that, wrapped text inherits the same chopped-up prosody as over-small chunks.

How the pieces fit

File	Responsibility
AppState	settings, speech lifecycle, wiring
BackendController	spawns, supervises and reclaims the speech server process
StreamingSpeechRequest	one streaming HTTP request to the server
PCMStreamPlayer	playback, lead-silence trimming, pause, device-change recovery
SilenceTrim / SpeechText	pure logic, unit tested
ControlServer / ControlAPI	loopback HTTP framing / routing onto the app
GlobalHotKeys / Selection	Carbon hotkey registration / copying the selection
HelpView / AgentSkill	the in-panel help page / the bundled SKILL.md

If the app is force-killed, its speech server keeps the port; the next launch adopts that healthy server instead of loading the model again. Restarting the server from the UI terminates it first, but only after confirming the process really is this bundle's binary. A second copy of the app notices the first one on the control port and quits rather than fighting over the menu bar and the hotkeys.

Credits and license

Speech comes from Supertonic by Supertone Inc. — sample code under the MIT License, model weights under the OpenRAIL-M License. The release DMG redistributes both; MODEL_LICENSE travels inside the app bundle, and the OpenRAIL-M use-based restrictions apply to the model and to its output.

This app is MIT licensed — see LICENSE.