Voice AI research notes

Architecture, pricing and turn-taking for developers building voice agents. Every rate links to the provider's own page, with the date we checked it.

OpenAI and voice architecture

• GPT-Live API vs. Realtime API: one voice frontend, one voice model
  GPT-Live API vs. OpenAI Realtime: how gpt-live-1 delegates to a backend, what 10 minutes costs on each, plus transports, modalities and session limits.
• OpenAI Realtime API alternatives: decide which layer you are replacing
  OpenAI Realtime API alternatives compared by layer: Gemini 3.8 Live, GPT-Live, Deepgram, Pipecat, LiveKit, open models and Vapi or Retell, with tools and rates.
• Full-duplex voice AI: listening while speaking, and what turn-based agents do instead
  Full-duplex voice AI listens while it speaks. How that differs from turn-based agents with VAD and barge-in, with correction, backchannel and tool examples.
• Choosing a low-cost voice API with function calling
  Voice API with function calling compared: who runs the tool, whether speech continues while it runs, what an interruption cancels, and the published rate.

Frameworks and platforms

• Pipecat vs. LiveKit: compare the framework, the transport and the host
  Pipecat vs LiveKit for voice agents: LiveKit's media server, LiveKit Agents and Pipecat compared on languages, tools, interruptions, hosting meters and cost.
• Vapi alternatives for developers: buy the platform or own the stack
  Vapi alternatives sorted by layer: Retell, LiveKit, Pipecat, Deepgram, GPT-Live and Gemini Live. Current Vapi pricing, plus Vapi vs LiveKit and Vapi vs Retell.
• Self-hosted voice AI vs. managed APIs: what a GPU conversation-hour costs
  Self-hosted voice AI with Nemotron VoiceChat 11B or Moshi vs. a managed voice API: GPU memory, licences, tool limits and a break-even formula you can fill in.
• Nemotron VoiceChat tool calling: what the model supports and what hosts expose
  Nemotron VoiceChat tool calling: how the 11B model's tool channel works, offline demo vs. live container, documented limits, and why .wave's beta has no tools.

Pricing

• AI voice agent pricing: billing formulas, worked examples and a calculator
  AI voice agent pricing compared by billing unit: session minutes, tokens, audio direction, platform fees and GPU hours. Formulas, worked examples, calculator.
• LiveKit pricing for voice agents: the agent-session minute is one line of the bill
  LiveKit pricing for a voice agent: plan fees, included minutes, agent sessions, inference, telephony and observability, itemised for a web app and a phone app.
• Retell AI pricing: what 10,000 voice minutes cost
  Retell AI pricing worked through three named configurations at 10,000 minutes, with telephony, transfers, silence, add-ons and concurrency fees shown apart.
• Gemini Live API pricing: why a long conversation costs more per minute
  Gemini Live API pricing for Gemini 3.8 Live: per-token audio rates, why retained context is billed every turn, and what transcription and grounding add.
• Session time and audio time have different prices.
  Published Gemini Live, Grok Voice, and .wave Nemotron prices, with hourly audio estimates and billing assumptions.
• What does an hour of conversation cost?
  Compare GPT-Live and .wave voice API pricing, tool-calling limits, and the GPU utilization needed for lower-cost full-duplex voice infrastructure.

Turn-taking, tools and reliability

• Semantic VAD vs. server VAD: stop cutting off unfinished sentences
  Semantic VAD vs. server VAD in the OpenAI Realtime API: silence thresholds, eagerness timeouts, a timing table for a mid-sentence pause, and LiveKit's modes.
• Silero VAD and Smart Turn solve different parts of turn-taking
  Silero VAD finds speech; Pipecat Smart Turn decides if the turn is over. A Python sketch of the two-stage detector with 16 kHz resampling and per-turn resets.
• Why your Realtime voice agent keeps talking after an interruption
  OpenAI Realtime API interruptions: cancel the response, flush queued audio, truncate unheard speech, and handle pending tool calls over WebRTC and WebSocket.
• OpenAI Realtime API function calling: from a spoken request to a finished action
  OpenAI Realtime API function calling, event by event: tools in session.update, call_id, function_call_output, response.create, timeouts and idempotency.
• Gemini Live API session resumption: reconnect without starting over
  Gemini Live API session resumption explained: 10-minute connections, SessionResumptionUpdate handles, GoAway timeLeft, compression, and recovering tool calls.
• Voice agent evaluation: check the final state, not the transcript
  A voice agent evaluation method: a scenario template for corrected requests, state checks, p50 and p95 timings with sample counts, and cost per successful task.
• A good voice isn't always a good agent.
  How to read full-duplex voice AI benchmarks. Understand tool-call success, spoken reply quality, confidence intervals, and the limits of cross-system comparisons.
• Build an AI voice tutor that hears the student's correction
  Build an AI voice tutor for a learning app: a narrow tutoring loop, lesson and progress tools, state kept outside the voice session, and a 30-minute cost.