KokoroVoice — Kokoro TTS as a macOS Spoken Content voice

Registers Kokoro-82M neural voices as real system voices via an Audio Unit speech synthesis extension (AVSpeechSynthesisProviderAudioUnit). Once enabled, they work with Spoken Content (Speak Selection), VoiceOver, and any app using AVSpeechSynthesizer. Everything runs locally on Apple Silicon via MLX.

Voices included: Kokoro Heart (af_heart, US female — top-rated) and Kokoro Michael (am_michael, US male). Add more in Extension/VoiceManifest.swift — one entry per voice, the model is shared.

Download (no build required)

Grab the notarized installer from Releases → open KokoroVoice.dmg → drag KokoroVoice into Applications → launch it once → then System Settings → Accessibility → Spoken Content → System voice → Kokoro Heart / Kokoro Michael. Requires Apple Silicon + macOS 15+.

The rest of this README is for building from source.

Requirements

• Apple Silicon Mac, macOS 15.0+
• Xcode 16+
• XcodeGen: brew install xcodegen
• Free Apple ID is enough for local signing

Setup

cd KokoroVoice
chmod +x download_models.sh
./download_models.sh      # fetches model (~600MB) + voices (~15MB) into Resources/
xcodegen                  # generates KokoroVoice.xcodeproj
open KokoroVoice.xcodeproj

In Xcode:

1. Select the KokoroVoice target → Signing & Capabilities → set your Team. Repeat for the KokoroVoiceExtension target.
2. Build & run the KokoroVoice scheme (this registers the extension).
3. In the app, hit Refresh — you should see "2 Kokoro voice(s) registered". Click Speak to test. The first synthesis takes several seconds (model load); after that it's faster than real-time.
4. Enable system-wide: System Settings → Accessibility → Spoken Content → System voice → pick Kokoro Heart or Kokoro Michael. Turn on Speak selection to get the read-aloud hotkey (default ⌥⎋).

How it works

App (KokoroVoice)                  ← host; registers extension, test UI
└── PlugIns/KokoroVoiceExtension.appex
    ├── Info.plist                 ← AudioComponents type "ausp" (speech synthesizer)
    ├── AudioUnitFactory           ← NSExtensionPrincipalClass
    ├── KokoroAudioUnit            ← AVSpeechSynthesisProviderAudioUnit subclass
    │     speechVoices             ← advertises voices from VoiceManifest
    │     synthesizeSpeechRequest  ← SSML → plain text → KokoroSwift → PCM buffer
    │     internalRenderBlock      ← streams buffer to system, signals completion
    └── Resources: kokoro-v1_0.safetensors + voices.npz

Dependencies (SPM, resolved automatically): KokoroSwift (MLX Swift port of Kokoro) and MLXUtilsLibrary (reads voices.npz).

Adding more voices

1. Open Extension/VoiceManifest.swift, add e.g.:

   KokoroVoiceDefinition(npzKey: "af_bella.npy", displayName: "Kokoro Bella",
                         identifier: identifierPrefix + "af_bella", language: "en-US"),

2. Check the key exists in the npz (28 voices ship in it): python3 -c "import zipfile; print(zipfile.ZipFile('Resources/voices.npz').namelist())"
3. Rebuild. British voices (b* prefix) automatically use the en-GB G2P path.

Troubleshooting

• Voices don't appear in the app / System Settings — the extension registers on app launch; quit and relaunch once. Still nothing: move KokoroVoice.app to /Applications, launch it, then log out/in. Check registration with: pluginkit -m | grep kokoro
• First "Speak" is silent for a while — normal: the 600MB model loads on the first request inside the extension. Subsequent requests are fast.
• Synthesis fails / no audio at all — check Console.app for messages containing KokoroVoice:. Most likely the model files didn't make it into the extension bundle: verify both files sit in Resources/ before running xcodegen, and that they appear under "Copy Bundle Resources" of the extension target.
• SPM resolution conflict on MLXUtilsLibrary — pin it to the exact revision in KokoroSwift's Package.resolved instead of branch: main in project.yml.
• Memory — expect ~1GB RAM in the extension process while speaking (fp32 model + MLX runtime). Fine on your 24-36GB machine.

Known limitations (v1)

• English only (Misaki G2P). Other Kokoro languages need eSpeak-NG G2P.
• Speech rate/pitch sliders in Spoken Content are ignored (SSML prosody is stripped). Could be mapped to KokoroSwift's speed parameter later.
• No word highlighting yet — KokoroSwift 1.0.8 exposes per-token timestamps, so AVSpeechSynthesisProviderOutputBlock marker events are a natural next step.
• The whole utterance is synthesized before playback starts (no streaming), so very long selections have a noticeable lead-in.

License & credits

MIT — see LICENSE. Built on:

• Kokoro-82M — the neural TTS model (Apache-2.0)
• KokoroSwift — MLX Swift port of Kokoro (MIT). This project pins a one-line fork that adds the missing MLXFast package dependency.
• MLX Swift and MLXUtilsLibrary (MIT)
• Render pattern from bocoup/apple-custom-speech-synthesizer