MAI-Voice-2.1 vs Sume TTS: narration cost by character count

Microsoft AI lists MAI-Voice-2.1 and MAI-Voice-2.1-Flash per million characters of generated speech. Sume TTS 1.0 bills per thousand characters, rounded up to the cent on each job, and one job takes up to 20,000 characters. The table compares both at every length these posts priced.

Worked example: 15,000 characters of narration: MAI-Voice-2.1 vs Sume TTS

Microsoft AI lists MAI-Voice-2.1 at $22 per 1M characters of generated speech, so 15,000 characters cost $0.330 direct. Sume's TTS 1.0 is billed at $0.0475 per 1,000 characters, rounded up to the cent on each job, and a job takes 1 to 20,000 characters: 15,000 characters is 1 job and $0.720.

Microsoft's figure is from its MAI-Voice-2.1 page, read 2026-10-10; Microsoft lists 23 languages for both voices. Sume's rate is the TTS 1.0 line on the API pricing page. The two products differ in voices and engine, so this is a price table and not a quality ranking.

What does 15,000 characters cost?

Per-character cost is the same in both routes; the only structural difference is Sume's per-job cent rounding and 20,000-character job size.

How do I split a script into Sume jobs?

Count the characters first and split at sentence ends so no job exceeds 20,000. This snippet only does the arithmetic; it calls no API.

import math
import pathlib

text = pathlib.Path("script.txt").read_text()
n = len(text)
jobs = math.ceil(n / 20000)
sizes = [min(20000, n - i * 20000) for i in range(jobs)]
cents = sum(-(-size * 475 // 100000) for size in sizes)
print(f"{n} characters, {jobs} job(s), {cents} cents")

What this comparison leaves out

Treat the numbers as list prices.

• Microsoft's page does not say how it counts characters for markup or pauses.
• Voice choice, language coverage and latency differ; Flash is a latency-focused variant.
• The snippet counts Python characters; confirm the count matches how the API counts your text.

Every narration length in one table

This page replaces 24 posts that asked the same question with different numbers. Each row is one of them, with the answer it gave when it was published.