A good AI voiceover is now indistinguishable from a hired read for most ads. The differences that still matter are cost per finished minute, cloning rights and what the free tier is allowed to do.
Our pick
ElevenLabs for quality and voice cloning — the benchmark most people compare against, billed per character, so long narration needs watching. Kokoro (open weights) for budget narration you can run free on a laptop. Whisper for the other direction: transcription, not synthesis, at fractions of a cent per minute.
What one job costslower is better
The 1 cheapest of 1 models with a published price. Bars are logarithmic so the cheap end stays readable; every row carries the exact figure.
ElevenLabs v3
ElevenLabs · Speech & VoiceThe quality benchmark for synthetic voice and cloning. Billed per character, so long-form narration costs real money.
Quality —
Kokoro 82M
Hexgrad · Speech & VoiceTiny enough to run on a laptop and startlingly good for its size. Fewer voices and languages than the commercial options.
Quality —
Remove a filter, or submit the model you are looking for.
Prices checked hourly against the LiteLLM dataset · last updated 27 minutes ago. API rates, not seat subscriptions. Watching 320 models for new releases.
What this task actually costs
A minute of finished narration is roughly 750 characters. At $0.18 per 1,000 characters that is about $0.14 a minute on the API — trivial for one video, noticeable across a 40-episode podcast. Subscription tiers bundle this into a monthly credit pool; we show which is cheaper at your volume.
The catches, out loud
- Free tiers usually watermark the audio or forbid monetised use — that kills it for client work.
- Cloning a real person’s voice needs their consent. Full stop.
- Prosody still gives machines away on emotional or very short reads; listen before you ship.
See also: all speech models · compare side by side.