Which MAI-Voice-2.1 model for ad voiceover: standard or Flash?

For ad voiceover pick standard MAI-Voice-2.1, not Flash. Microsoft's own model page puts audiobooks, content creation and voice-over with the standard model, and call-center agents, voice assistants and IVR with Flash.

Microsoft's split

Microsoft's pages describe Flash as the low-latency variant: about 45 ms model latency at $15 per million characters. Standard is listed at about 550 ms and $22 per million.

What an ad pipeline needs

An ad is a script, a voice and a finished file. Nobody waits for the first syllable, so latency does not matter. What matters is that the voice can be matched to a brand, takes can be repeated and the file lands in an editing timeline.

The Sume path for finished files

Sume's TTS Router is built around async jobs: POST /v1/tts-router/generate returns a request id, you poll status and read the result, and a job covers up to 1,200 seconds of audio. There is no streaming and no SSML field. Output defaults to mp3 at 44.1 kHz and 128 kbps; output_format also takes wav or raw.

curl -sS -X POST https://api.sume.com/v1/tts-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"sonic-3.6","transcript":"Fresh bread, baked at six.","avatar_handle":"product_host","language":"en","output_format":{"container":"wav","sample_rate":48000,"encoding":"pcm_s16le"}}' \
  | jq '.data | {request_id, status_url, result_url}'

When Flash is the right call

If the same script is read live to a caller, Flash fits and Sume does not: Sume has no streaming. Call-center and IVR work is Microsoft's own listed use for it.