local-voice

name	local-voice
description	Operate the Local Voice plugin for local STT, VOICEVOX TTS, camera capture, and VRChat OSC speaking cues.
user-invocable	false

Local Voice

Use this skill when the user asks for the Local Voice plugin directly, asks for local TTS or STT outside the Desktop Companion bridge, or wants the legacy chat-command voice surface checked.

Prefer the Desktop Companion control_companion tool for avatar speech and local consent state when the request is about the desktop avatar. Prefer the Hypura voice skill when the request is explicitly about the Hypura mic-to-agent loop. Use Local Voice when the request names this plugin, needs its chat commands, or needs its registered VOICEVOX speech provider.

Defaults

• TTS provider: voicevox.
• Default VOICEVOX endpoint: http://localhost:50021.
• Default VOICEVOX speaker id: 8, Kasukabe Tsumugi normal style.
• STT provider: whisper unless config overrides it.
• VRChat OSC speaking cue: parameter Speaking toggles true during playback and false after playback finishes or fails.

Commands

Use these chat commands only when command dispatch is the intended path:

/voice-assistant
/voice-assistant start
/voice-assistant stop
/voice-assistant speak <text>
/tts <text>
/voice_tts <text>
/camera capture

/voice_tts is the Telegram native command alias for /tts.

Safe operating flow

1. Confirm VOICEVOX Engine is running on the configured endpoint before speech.
2. For speech output, use /tts <text> or /voice-assistant speak <text>.
3. For live voice input, use /voice-assistant start only after the user has requested microphone capture.
4. Use /voice-assistant stop when the voice turn or troubleshooting session is finished.
5. Use /camera capture only when camera capture is requested or already covered by the user's local-consent context.

Configuration notes

The plugin reads plugins.entries.local-voice.config. Important fields:

• ttsProvider: keep voicevox for the local Kasukabe Tsumugi path.
• vvEndpoint: local VOICEVOX HTTP endpoint.
• vvSpeakerId: VOICEVOX speaker id. Use 8 for Kasukabe Tsumugi unless the user asks for another voice.
• sttProvider: whisper or openai-realtime.
• vrchatOscEnabled and vrchatOscPort: local OSC speaking cue behavior.
• cameraDeviceId, cameraWidth, cameraHeight, and cameraMinIntervalMs: camera capture source and throttle.

The registered speech provider also honors baseUrl/speakerId and the legacy vvEndpoint/vvSpeakerId keys so desktop runtime config and provider-style config stay aligned.

Proof path

• For VOICEVOX identity, query /speakers and confirm speaker id 8 is Kasukabe Tsumugi normal style.
• For synthesis, call /audio_query and /synthesis with speaker=8.
• For plugin regression proof, run the Local Voice manifest skill test plus the VOICEVOX provider test.
• For user-facing speech proof, prefer a short phrase and confirm the command returns success instead of silently falling back to browser/system TTS.

Troubleshooting

• Mechanical voice usually means the request did not use the VOICEVOX path or another companion path fell back to browser/system TTS. Recheck ttsProvider, vvEndpoint, and vvSpeakerId.
• VOICEVOX request failures usually mean the engine is not running, the endpoint is wrong, or speaker id 8 is unavailable in that VOICEVOX install.
• If playback works but VRChat does not react, check OSC enablement and the vrchatOscPort setting.
• If microphone capture does not produce transcript text, stop the assistant, check the configured STT provider and audio input device, then start it again only after the user still wants live capture.