A voice agent feels like one product, but you are billed for three. Speech-to-text turns the caller's audio into a transcript, priced per minute. An LLM answers, priced per token. Text-to-speech turns the reply back into audio, priced per character. Three different meters on three different bases, which is why the total is so hard to feel until it arrives. This calculator runs all three on one conversation and splits the result.
The LLM is usually the cheapest of the three
The counterintuitive result, which the breakdown makes obvious, is that the language model is often the smallest line on the bill. Take the defaults — a five-minute call, a cheap fast LLM, a mid-tier voice — and the transcription and synthesis together dwarf the tokens between them. Speech is simply expensive to move in and out of text: a per-minute transcription rate and a per-character synthesis rate both scale with how much anyone talks, and people talk a lot in five minutes. Swapping to a pricier LLM barely moves the total; swapping the transcription tier moves it a lot. The three share stats tell you which lever you are actually holding before you tune anything.
What drives each of the three lines
Transcription is priced on the audio you send it, so the "caller talk share" field matters: you only transcribe the human side, and a call that is 55% caller costs 55% of the wall-clock minutes to transcribe. Synthesis is priced on the characters the assistant speaks, which this calculator derives from the LLM's output tokens at roughly 4.3 characters per token — so shorter spoken replies cut the text-to-speech line directly and linearly. The LLM line follows the usual token arithmetic, with one voice-specific quirk: the input grows every turn because the running transcript is re-sent, the same accumulation that makes agents expensive. If your conversations run long, price that growth in the AI agent cost calculator.
The levers, in order of effect
Because synthesis and transcription dominate, that is where to optimise first. Shorten spoken replies — voice users prefer terse answers anyway, and every character you do not speak is a character you do not pay to synthesise. Choose the transcription tier deliberately: the spread between the cheapest and dearest speech-to-text models in the index is several-fold for accuracy differences that many use cases will not notice. Only after those two should you reach for a cheaper LLM, and by then the LLM is usually already the small number. The speech AI cost calculator prices transcription and synthesis on their own if you want to compare voices in isolation.
What this deliberately leaves out
This is a per-turn-batched model, not a real-time-streaming one. It does not price the streaming/realtime premiums some vendors charge for low-latency bidirectional audio, nor interruption handling, silence, or the audio-token pricing of the newer end-to-end speech models that skip the transcript entirely. It also assumes one round of speech each way per the tokens you enter. Treat the output as the floor for a classic three-box stack; a realtime API with its own per-minute audio pricing is a different model and usually a higher one.
Prices used by this calculator were last verified on 31 August 2026 from the vendors' own pricing pages. See the full price index for every figure and its source, the change log for what has moved recently, and the methodology for how the index is maintained and where its limits are.