Nightly voice-agent test calls: 200 utterances, a month of TTS

Testing a voice agent every night with 200 caller utterances of 100 characters each is 20,000 characters a night, or 600,000 over 30 nights. That costs $9.00 on MAI-Voice-2.1-Flash, $13.20 on MAI-Voice-2.1 and $28.50 on Sume TTS at list. A test lane that runs on every build still costs less than a monthly SaaS seat.

The bill

The test script is fixed, so the question is whether to synthesize it again every night at all.

Render once, replay nightly

The 200 utterances (about 7 seconds each at 100 characters, an estimate) never change between runs unless you edit the script. Render them once on Sume, store the job id and each segment's audio_url, and replay the files every night. A one-off render of the script is 20,000 characters, which is $0.95 on Sume at list and, at about 1,400 seconds of speech, needs two jobs of 100 utterances to stay under the 1,200-second limit, and 30 nights of replay cost nothing in TTS.

That makes the real monthly TTS bill whatever the script edits are, not 30 full runs. If the script changes 10 percent a week, you re-render roughly 2,000 characters, or about 10 cents.

When fresh voices help

Fresh synthesis is worth paying for in two cases: you want to test with different voices, or you want varied wording. For those, pick several voices and rotate them per night, one Sume job each. Twenty thousand characters is exactly the request cap and the speech runs past 1,200 seconds, so one night is two jobs of 100 utterances.

Use idempotency_key per night, such as the date, so a retried run of the test harness never bills the same night twice. The job envelope gives you status and result URLs, so a harness can poll them with a plain loop.

What the table leaves out

Microsoft's figures are the launch prices from 2026-10-01 (Microsoft AI, read 2026-10-05): $22 per 1M characters for MAI-Voice-2.1 and $15 per 1M characters for MAI-Voice-2.1-Flash, which makes up to 45 seconds of audio per call. Sume's figure is its list rate per transcript character, and on Sume spaces and punctuation count toward usage, so count characters of the final text.

The cheaper per-character price does not make the Microsoft route a drop-in swap. A Sume TTS request is an async job whose result carries a hosted audio_url, duration, optional word timings and sentence slices, and that hosted file is what Timeline audio and renders accept. Count the extra work of storing and joining files outside Sume before you call the gap a saving.

Check language coverage first. Microsoft lists 23 languages and 26 locales for MAI-Voice-2.1, and Sume takes a language field that you must set for every non-English transcript. A cheaper model that does not speak your target language is not an option, so settle that before you compare totals.

Format and the streaming boundary

A caller test needs the agent to hear audio, so the format matters. Sume returns MP3 by default and a wav container on request. Pick the one your telephony test tool accepts and keep it for the whole script.

Streaming is a different problem. A live voice agent needs the speech the moment a sentence is ready, which is what Microsoft's streaming launch addresses. Sume TTS is a job with a result URL, which fits prerecorded caller turns and not live replies from the agent under test. Keep that boundary clear when you plan the harness.

Also record which job produced which night's audio. When a test fails, you want to replay the exact caller audio that broke the agent, and a stored job id with its segment index lets you find the same file again instead of regenerating something that may sound slightly different.