One minute and one hour of AI narration: MAI-Voice vs Sume TTS cost

One minute of narration costs about 2.0 cents on MAI-Voice-2.1, 1.4 cents on MAI-Voice-2.1-Flash and 5 cents on Sume TTS 1.0; one hour costs $1.19, $0.81 and $2.58. These use an assumed 900 characters per spoken minute (150 words at six characters including the space), Microsoft's listed $22 and $15 per 1M characters, and Sume's $0.0475 per 1,000 characters (read 2026-10-08).

MAI-Voice is cheaper per character, on both models. The Sume figures are higher partly because each job rounds up to the cent.

The arithmetic

900 characters at $22 per 1M is $0.0198; at $15 per 1M it is $0.0135. On Sume, 900 characters is 4.275 cents, billed as 5 cents for a one-minute job.

An hour is 54,000 characters. Sume caps a request at 20,000 characters and at 1,200 seconds of audio, and 1,200 seconds at this speaking rate is 18,000 characters, so I use three jobs of 18,000 characters: 85.5 cents each, billed as 86 cents, $2.58 in all.

What the price does not cover

Microsoft's Flash is documented with a 45-second audio limit per request on the launch post, so an hour on Flash is many calls that you must stitch. MAI-Voice-2.1 is described for long-form work. Both are public previews without an SLA, per the Learn page. Voice choice, languages and cloning gates differ too; a cheaper rate on a voice you cannot use is not a saving.

Sume's jobs are asynchronous and return mirrored audio artifacts you fetch from the job result; there is no streaming.

How to use the numbers

Compute your own characters per minute from a real script: count characters including spaces, divide by the finished audio length. Languages with different character densities, such as Korean or Chinese, change the figure a lot. Then re-run the table with your own number. For a one-off ad read, the difference is cents; for 10,000 hours it is not, and then volume terms with the vendor matter more than any list price.

Where the gap closes

Sume's per-character rate is higher than the MAI list prices, but Sume's price includes a voice workflow that already returns word timestamps on request, sentence segmentation and mirrored audio artifacts you can hand to the timeline and caption routes. If your narration goes straight into a rendered video, the extra cost per minute is cents, while the integration work saved can be hours.

If you only need raw audio at maximum volume and you accept a preview service without an SLA, MAI-Voice is cheaper on paper, and its Flash model is cheaper still. The honest summary is that price per character favors Microsoft's list and workflow favors whichever route you already integrated.