Cheap TTS API: what a thousand characters actually costs
Speech is the cheapest generative modality per unit of output and the easiest to overpay for. Most TTS pricing pages quote a plan, not a unit, which makes real comparison impossible. This page is built the other way round: unit price first, live from our catalogue.
Why per-character is the only useful unit
- Audio length is a consequence, not an input. You pay for the text you send, so a 30-second clip and a 3-minute clip of the same word count cost the same.
- Voice quality is not linear in price. The cheapest voice is often indistinguishable for narration; premium voices matter for ads and character work.
- Character counting has traps. Some providers count SSML tags and punctuation; ours counts the text you pass in.
Worked example: a 10-minute narration is roughly 9,000 characters. At $0.0105 per thousand that is about $0.09. At the $0.015 list price it is $0.135 — the difference matters at podcast or audiobook volume, and is irrelevant for a handful of UI prompts.
Live speech prices from our catalogue
USD per thousand characters, sorted cheapest first. Read from the catalogue at page load.
| Model | Vendor | Per 1K chars | List price | vs list |
|---|---|---|---|---|
| — | — | — | — | — |
Calling it
OpenAI-compatible audio endpoint, same key as chat:
POST https://tapixai.com/v1/audio/speech
{"model":"gemini-3.8-flash-tts","input":"Your narration text here","voice":"altair"}
For transcription the endpoint is /v1/audio/transcriptions; both are billed per thousand characters of text processed.
Where cheap TTS goes wrong
The two failures that cost real money: regenerating the same paragraph because the voice mispronounced a name (fix the pronunciation, not the whole file), and chunking text badly so you lose prosody at every seam. Batch by paragraph and cache the audio you have already generated — a prepaid per-character model makes caching strictly profitable, because you stop paying twice for the same sentence.
Try it free with $1 credit See full pricing