Text-to-Speech Cost Calculator

Characters into dollars, with SSML markup counted or not counted, an audio cache for your fixed phrases, and retakes billed in full.

Requests per monthUtterances delivered to a listener.

requests / mo

Characters of spoken text per requestThe words, before any markup.

characters

SSML markup overheadCharacters your tags add, as a share of the spoken text. Breaks, emphasis, phoneme hints and prosody add up fast.

Is markup billed?Providers differ, and this is the field most estimates get wrong. Check your own docs.

Served from a cache of pre-rendered audioFixed phrases — greetings, confirmations, error messages — synthesised once and replayed.

RetakesExtra syntheses per 100 requests, billed in full.

Synthesis rate$ per 1M characters. Price lists are often quoted per 1K characters — multiply that figure by 1,000 before typing it here.

per 1M chars

Monthly text-to-speech spend

$423.36

42,000 syntheses for 50,000 delivered utterances — $0.0085 each.

Spoken characters per request: 600
SSML markup characters: 72 — billed
Billable characters per synthesis: 672
Syntheses per month: 42,000
Characters billed: 28,224,000
Rate per 1,000 characters: $0.01
Cost of the markup: $45.36
Avoided by the audio cache: $105.84
Monthly total: $423.36
Cost per delivered utterance: $0.0085
Annual: $5,080

What this assumes: billing is per character of input, not per second of audio produced and not per token — if your provider bills by generated audio duration this arithmetic does not apply; a "character" is counted as one unit here, but some providers count bytes, which makes Cyrillic, Greek and CJK text cost two to three times what its character count suggests; cached audio is free to serve, which ignores the storage and CDN behind it; retakes are billed at full price because a partial synthesis is still a synthesis; voice cloning, custom voice hosting and per-voice premiums are not modelled.

Text-to-speech is billed on what you send, not on what the listener hears, and that one fact produces every surprise on the invoice. SSML is the clearest case: a paragraph wrapped in break tags, emphasis and phoneme hints can be a third longer than the words inside it, and if your provider bills the raw input then a third of your synthesis budget is being spent on angle brackets. Whether that is true of your provider is a documented fact you can check in a minute, and it is worth checking before you optimise anything else.

The second lever is that most voice products say the same things over and over. Greetings, confirmations, error messages, the twelve responses your IVR gives ninety per cent of the time — these are fixed strings, and a fixed string should be synthesised once and stored as audio. The cache field above is usually the largest single saving available, and unlike prompt caching it never expires, because the text does not change.

What this leaves out: latency, which for a conversational product is the actual constraint and which pushes you toward streaming synthesis that you cannot cache; storage and delivery of the audio, which for long-form content dwarfs the synthesis charge; and licensing, if the voice you want is not one of the free ones. If you are comparing TTS against recording a human, note that the human is a one-time cost per script and TTS is a per-request one — the break-even is a volume question, and it moves the moment your copy changes.

Text-to-Speech Cost Calculator

The reasoning behind this

Related tools