Voice AI calling: how it works, what it costs, what the law says

Cascade or speech-to-speech?

The table describes a cascade: three separate models. Speech-to-speech models such as OpenAI’s realtime models and Gemini Live hear audio and answer in audio, which removes two hand-offs. The trade is control: a cascade lets you swap any part, log the transcript and pick a cheaper voice.

Where the milliseconds usually go

• Endpointing. Deciding the caller has finished. Twilio notes fixed end-of-turn defaults are often set around 500 ms, which can eat half the budget on its own. Semantic turn detection cuts it.
• Long prompts. Time to first token grows with the context you send each turn. Trim history and cache the system prompt.
• Distance. ITU-T G.114 says one-way delay under 150 ms feels “essentially transparent” (ITU-T G.114). Put your models in the same region as your phone carrier.

The worked example, line by line

Our defaults: the agent speaks for 50% of the call and replies 6 times a minute. Each reply sends 1,500 input tokens and writes 40. Speech runs at 800 characters a minute, the top of Cartesia’s stated 750–800 range. These are assumptions; change them in the calculator.

Google lists gemini-3.8-flash at double the price from 1 Jan 2027, which would lift the LLM line to $0.0153. In India, Plivo charges ₹0.38 a minute for domestic calls in 30-second pulses (Plivo), so a 40-second call bills as 60 seconds.

All-in platform prices

For the business case, use the AI calling ROI calculator. It compares these costs with the cost of a human caller.