macOS Local Voice

Local STT and TTS on macOS using native Apple capabilities. Speech-to-text via yap (Apple Speech.framework), text-to-speech via say + ffmpeg. Fully offline, no API keys required. Includes voice quality detection and smart voice selection.

Security Scan

The skill's code, instructions, and required binaries line up with its stated purpose (local macOS STT/TTS) and do not request unrelated credentials or network access.

Purpose & Capability

Name/description (local macOS STT/TTS) match the actual requirements and behavior: required binaries are yap, say, and osascript (all relevant), code uses Apple Speech.framework via yap and AVFoundation via osascript, and ffmpeg is optional for format conversion.

Instruction Scope

SKILL.md directs the agent to run the included scripts and to use local system settings for downloading voices. The instructions do not ask the agent to read unrelated files, export environment secrets, or contact remote endpoints. It does reference the agent 'message' tool for sending generated audio (expected for delivering voice notes).

Install Mechanism

This is instruction-only with included scripts (no install spec). The README suggests installing yap/ffmpeg via Homebrew — a standard, traceable source. No downloads from arbitrary URLs or archive extraction are present.

Credentials

No environment variables or credentials are requested. The scripts use HOME to write output under ~/.openclaw/media/outbound (reasonable for generated media). There are no requests for unrelated secrets or config paths.

Persistence & Privilege

Skill is not forced always-on and does not modify other skills or system-wide settings. It only creates per-user media files in a reasonable path and requires user action to download premium voices.

Assessment

This skill appears coherent and local-only, but review these before installing: 1) Install yap and ffmpeg from trusted package sources (Homebrew/taps shown in README). 2) The scripts invoke local system commands (yap, say, osascript, ffmpeg) — ensure you are comfortable granting the skill the ability to execute those on your Mac. 3) Generated audio is saved under ~/.openclaw/media/outbound and the SKILL suggests using the agent 'message' tool to send it — verify recipients before sending sensitive audio. 4) The locale detection and voice selection are heuristic (CJK-character ratio, calling voices.mjs) — test with your language. If you need further assurance, inspect the included scripts (they are readable JS) and confirm no network calls are added in your environment.

Fully local speech-to-text (STT) and text-to-speech (TTS) on macOS. No API keys, no network, no cloud. All processing happens on-device.

Requirements

• macOS (Apple Silicon recommended, Intel works too)
• yap CLI in PATH — install via brew install finnvoor/tools/yap
• ffmpeg in PATH (optional, needed for ogg/opus output) — brew install ffmpeg
• say and osascript are macOS built-in

Speech-to-Text (STT)

Transcribe an audio file to text using Apple's on-device speech recognition.

node {baseDir}/scripts/stt.mjs <audio_file> [locale]

• audio_file: path to audio (ogg, m4a, mp3, wav, etc.)
• locale: optional, e.g. zh_CN, en_US, ja_JP. If omitted, uses system default.
• Outputs transcribed text to stdout.

Supported STT locales

Use node {baseDir}/scripts/stt.mjs --locales to list all supported locales.

Key locales: en_US, en_GB, zh_CN, zh_TW, zh_HK, ja_JP, ko_KR, fr_FR, de_DE, es_ES, pt_BR, ru_RU, vi_VN, th_TH.

Language detection tips

• If the user's recent messages are in Chinese → use zh_CN
• If in English → use en_US
• If mixed or unclear → try without locale (system default)

Text-to-Speech (TTS)

Convert text to an audio file using macOS native TTS.

node {baseDir}/scripts/tts.mjs "<text>" [voice_name] [output_path]

• text: the text to speak
• voice_name: optional, e.g. Yue (Premium), Tingting, Ava (Premium). If omitted, auto-selects the best available voice based on text language.
• output_path: optional, defaults to a timestamped file in ~/.openclaw/media/outbound/
• Outputs the generated audio file path to stdout.
• If ffmpeg is available, output is ogg/opus (ideal for messaging platforms). Otherwise aiff.

Sending as voice note

After generating the audio file, send it using the message tool:

message action=send media=<path_from_tts.sh> asVoice=true

Voice Management

List available voices, check readiness, or find the best voice for a language:

node {baseDir}/scripts/voices.mjs list [locale]     # List voices, optionally filter by locale
node {baseDir}/scripts/voices.mjs check "<name>"     # Check if a specific voice is downloaded and ready
node {baseDir}/scripts/voices.mjs best <locale>       # Get the highest quality voice for a locale

Quality levels

• 1 = compact (low quality, always available)
• 2 = enhanced (mid quality, may need download)
• 3 = premium (highest quality, needs download from System Settings)

If a voice is not available

Tell the user: "Voice X is not downloaded. Go to System Settings → Accessibility → Spoken Content → System Voice → Manage Voices to download it."

Notes

• The say command silently falls back to a default voice if the requested voice is not available (exit code 0, no error). Always use voices.mjs check before calling tts.mjs with a specific voice name.
• Premium voices (e.g. Yue (Premium), Ava (Premium)) sound significantly better but must be manually downloaded by the user.
• Siri voices are not accessible via the speech synthesis API.