InstaGigDev · Voice Technology Lab
Speech Lab
Speak naturally and watch your words stream into chat. After a pause, your next turn starts a separate bubble below the agent's reply while audio continues through the E2Ev5-protected SignalR channel. See the latest Qwen3 4B turn status pinned above the chat. Choose an OpenRouter conversation model and hear the agent through local Qwen3-TTS, or compare with ElevenLabs when configured.
Speech capture and live conversation
Qwen3-TTS settings
Waiting for the server to report settings.
Advanced synthesis
Audio always streams over the encrypted connection. Complete-text prompt mode controls how Qwen receives the reply text; it does not turn audio streaming off.
Server configuration and model capabilities · read only
Reported by the server. GPU, batching, and deployment settings apply to the shared worker and cannot be changed for a session.
Live conversation
Chat
Your live Parakeet transcript and the InstaGig agent's replies will appear here.
Microphone stream
Start once. Speak continuously.
The microphone continuously forwards one 40 ms PCM16 packet in an encrypted E2Ev5 SignalR command, including during silence. Parakeet pushes each available hypothesis from continuous 320 ms inference frames; speaking again after a pause starts a new chat bubble. Messages appear in conversation order, with new turns at the bottom. Agent voice returns as encrypted, ordered PCM chunks over that same SignalR channel and stops immediately when you speak. Choose the voice before listening; it stays fixed for that session. Our local models are Parakeet for transcription, Qwen3 4B for turn detection, and Qwen3-TTS for speech. Conversation text is processed through the selected OpenRouter model. ElevenLabs processes agent speech text externally when selected. No application REST API or separate audio WebSocket is used.