Text-to-speech

How it works

Text-to-speech (TTS) models turn written text into spoken audio. They are usually priced per character of input text or per minute of audio produced, and some token-based models bill audio output as tokens.

Models differ in voice quality, available voices, languages, control over style and pace, streaming latency and support for custom or cloned voices. Voice cloning usually comes with consent requirements in the provider's usage policies.

For live voice apps, time to first audio matters as much as quality. Test pronunciation of product names, numbers and abbreviations, which are common failure points.

Worked example

Illustrative: at $15 per 1M characters, a 1,500-word article of about 9,000 characters costs about $0.14 to narrate, and 1,000 such articles a month cost about $135.

Common mistakes

• 1Judging quality on short sample sentences instead of long, realistic scripts.
• 2Not testing how numbers, dates, acronyms and brand names are read aloud.
• 3Using voice cloning without checking the consent and disclosure rules in the usage policy.