A 150 ms TTS claim vs Sume's async TTS: which jobs need which

If you need speech to start within a fraction of a second, as in a live voice agent, Sume TTS is not the tool. Sume TTS 1.0 is an asynchronous job: you submit text, poll the job status and fetch an audio file, and Sume's TTS Router docs list streaming TTS as a non-goal of its first release. Microsoft's MAI-Voice-2.1-Flash, launched 2026-10-01, is a vendor-claimed 150 ms end to end at $15 per million characters (read 2026-10-06). It is not available through Sume.

If you need finished audio for a video, a lesson or an ad, the picture reverses. A file that exists before anyone presses play does not care about 150 ms, and the same balance that pays for the voice also pays for the clip, the captions and the render. Pick by the latency budget first and the price second.

The cost of 1,000 short turns

A short spoken turn is the worst case for Sume's rounding, since each job rounds up to a cent. A 200-character turn is $0.0095 in raw characters and bills $0.01, so 1,000 turns cost $10.00. The same 200,000 characters list at $4.40 on MAI-Voice-2.1 and $3.00 on Flash. For a conversational loop the list price is a worse fit and the latency is a worse fit too, so use a streaming-capable vendor for that job.

The same arithmetic favours Sume the other way when the text is long. A 2,000-character narration bills $0.10 (raw $0.095), a 5 percent premium, and one 20,000-character job bills $0.95.

• A turn of 200 characters bills $0.01 on Sume: 1,000 turns cost $10.00.
• The 150 ms figure is the vendor's claim. Measure it in your own region before you rely on it.
• Sume TTS jobs reserve and capture like other jobs, so a failed job does not cost you.
• Pre-generate whatever you can, and keep the live path for text that cannot exist yet.

Price a turn-by-turn loop

The script below prices a day of short turns on Sume and at the two MAI list prices, so you can see the gap for your own turn length and volume.

def sume_turn_cents(chars: int) -> int:
    return -(-chars * 475 // 100_000)


def day(turns: int, chars: int) -> dict:
    sume = turns * sume_turn_cents(chars) / 100
    mai = turns * chars * 22 / 1e6
    flash = turns * chars * 15 / 1e6
    return {"sume": round(sume, 2), "mai": round(mai, 2),
            "flash": round(flash, 2)}


print(day(1000, 200))
print(day(1000, 2000))

What to do with this

Split the work by latency. Keep live turns on a real-time voice vendor, and send everything that can be a file to Sume TTS, where it sits next to video, captions and the timeline render. If you are unsure which side a feature falls on, ask whether anyone waits on the audio. The rates are on the API pricing page, and the vendor figures are from the October 2026 model tracker.

A practical split for a product with both needs

Most products with a voice feature have both kinds of audio. A support bot answers live, and the same company records a product tour, a set of hold messages and the narration for a launch video. Treat these as two lanes with two vendors if the live lane needs it, and do not force the file lane onto a streaming service because the first lane chose one.

The file lane gets cheaper to manage on Sume because the audio, the video and the render share one wallet and one usage ledger. You can read one balance and see captures for the voice line next to the clip it narrates. The live lane stays with whichever vendor meets your latency target. When MAI-Voice is the vendor you pick, it is billed by Microsoft and it does not touch your Sume balance.