Voice Pipeline

Voice Pipeline

CortexPrism includes a full voice pipeline: speech-to-text (STT), text-to-speech (TTS), voice activity detection (VAD), and real-time audio streaming.

Components

Component	File	Description
STT	src/voice/stt.ts	Speech-to-text via OpenAI Whisper
TTS	src/voice/tts.ts	Text-to-speech via OpenAI TTS or ElevenLabs
VAD	src/voice/vad.ts	Energy-based voice activity detection
Audio	src/voice/audio.ts	Audio format conversion (ffmpeg fallback)
Channel	src/voice/channel.ts	Voice channel plugin implementing ChannelPlugin
Pipeline	src/voice/pipeline.ts	Auto-TTS post-output hook
Manager	src/voice/manager.ts	Voice mode state management

Voice Mode

cortex voice enable         # Enable voice mode
cortex voice disable        # Disable voice mode
cortex voice status         # Show current voice config
cortex voice set-voice <id> # Change default voice

Voice Settings

In ~/.cortex/config.json:

{
  "voice": {
    "enabled": false,
    "provider": "openai",
    "defaultVoice": "alloy",
    "autoTTS": false
  }
}

OpenAI TTS voices: alloy, echo, fable, onyx, nova, shimmer

ElevenLabs is supported as an alternative TTS provider. Set voice.provider to elevenlabs and configure your API key in settings.

Auto-TTS

When voice.autoTTS is enabled, a post-output pipeline hook automatically synthesizes agent text responses to audio. Audio is forwarded to the WebSocket client before the done signal.

VAD (Voice Activity Detection)

Energy-based VAD with configurable parameters:

• Frame size: Duration of each audio frame
• Speech threshold: Energy threshold to detect speech
• Silence timeout: Duration of silence before ending capture
• Minimum speech duration: Minimum length to consider valid speech

WebSocket Audio Protocol

Real-time audio streaming over WebSocket:

Client → Server:

{ "type": "audio_chunk", "data": "<base64>" }
{ "type": "audio_end" }

Server → Client:

{ "type": "speak", "text": "...", "voice": "alloy" }
{ "type": "audio", "data": "<base64>" }
{ "type": "voice_state", "enabled": true }

Transcribed speech is dispatched directly into the agent loop as a user message.

Agent Tools

Tool	Description
speak	Text-to-speech via configured TTS provider
listen	Speech-to-text via configured STT provider

REST API

Endpoint	Description
POST /api/voice/transcribe	Speech-to-text transcription
POST /api/voice/synthesize	Text-to-speech synthesis
GET /api/voice/synthesize/:text	TTS via GET request
GET /api/voice/providers	List available TTS providers

Web UI

Voice settings tab includes:

• Provider selection (OpenAI / ElevenLabs)
• Default voice choice
• Language setting
• Auto-TTS toggle
• Microphone button in chat (with recording animation)
• Speaker button on each assistant message for on-demand TTS
• Speaking pulse animation indicator

See Also

• REST API — Voice API endpoints
• Configuration — Voice config options