Speech API pricing

Speech APIs split into two jobs priced two different ways. Speech-to-text (transcription) is billed per minute of audio processed. Text-to-speech is billed per character — or, increasingly, per second of generated audio, which comes out to roughly the same thing once you account for speaking rate.

Transcription: the per-minute rate isn't the whole cost

The advertised per-minute price usually assumes batch processing of a clean recording. Real costs move on three things the pricing page rarely leads with. Streaming — real-time transcription for a live call or a voice agent — costs more per minute than batch, sometimes double. Diarization, punctuation, and language detection are often add-ons that each bump the rate. And model tier matters: the cheap "base" model and the accurate "enhanced" model can differ by 3x, and for anything a human will read, the enhanced model is not optional.

For a rough monthly estimate, take your total minutes of audio, multiply by the streaming-or-batch rate for the model you'd actually ship, and add a margin for the features you'll turn on.

Text-to-speech: characters, voices, and the clone tax

TTS is billed per character of input text. A useful shortcut: one minute of speech is about 900 characters, so a provider quoting "$15 per million characters" is charging roughly $0.013 per minute of audio. Watch for two multipliers. Premium or "realistic" voices are often a separate, higher tier than the standard set. Voice cloning — training a custom voice on your own recordings — usually carries both a one-off fee and a higher per-character rate on anything that voice generates.

Comparing providers

Decide first whether you need batch or streaming, and whether accuracy or price is the binding constraint — for compliance transcripts it's accuracy, for a bulk archive it's price. Then price your actual monthly volume against two or three providers using the model tier you'd deploy, not the cheapest one on the page. The individual comparisons here do that math for the common matchups. The full rate dataset is on the [data page](/data).