AI Pipeline

🚧 Coming soon. Architecture deep-dive lands with the first backend release.

WaveKat Voice streams audio through four stages with overlap at every boundary, targeting p95 end-to-end latency under 400ms:

caller audio ──▶ VAD ──▶ ASR ──▶ LLM ──▶ TTS ──▶ caller audio
                                 │
                                 └─▶ tool calls (booking, lookup, transfer)

Planned topics:

• The Frame model (a Rust port of Pipecat’s streaming Frame concept)
• VAD backend choices (WebRTC, Silero, TEN-VAD, FireRedVAD)
• Turn detection vs VAD — why they’re different
• ASR backends: local (Whisper.cpp, SenseVoice) vs cloud (Deepgram, OpenAI Realtime)
• LLM provider abstraction (OpenAI-compatible, llama.cpp, Ollama, Anthropic)
• TTS backends and voice cloning
• Barge-in and interruption handling
• Latency budget breakdown