120 ms first partial vs a 600-second file job: pick the STT shape

Choose by what is waiting for the text. MAI-Transcribe-2-Streaming lists about 120 ms to a first partial transcript and about 128 ms end-to-end, which suits live captions and voice agents. Sume STT is a file job: submit up to 600 seconds of audio, poll the job, read the transcript. It does not stream, so it fits recorded audio where the text feeds an edit, not a live conversation.

Five workloads

Prices are Microsoft's page prices (introductory) and Sume's $0.01 per audio minute. The 25-second row shows the effect of Sume's 1-cent minimum: list math would be $0.0042 but the request bills at least 1 cent.

How a Sume STT job returns

Sume jobs are async by default: you receive 202 and poll status, or pass a webhook. A sync mode waits up to 30 seconds on the submit call, then you keep polling (Jobs and results). The catalog entry lists POST /v1/stt-1.0/transcribe and the job status and result endpoints. A 600-second file may not finish inside the 30-second wait, so code the poll path first and treat the sync answer as a bonus.

Use a webhook if you have many chunks. For the one-hour recording above, six chunks run as six jobs that you can submit together, each with its own idempotency key and its own duration_seconds.

When to run both

A product with live captions and a published transcript can run streaming for the live view and a batch pass for the final text. On Microsoft's page prices one hour through both is $0.54 + $0.10 = $0.64. If the final pass runs in Sume instead, one hour is $0.54 + $0.60 = $1.14. The extra $0.50 an hour buys the transcript inside the Sume job and Timeline flow.

Misfit warnings

Two mistakes recur. The first is using a file job for a live feature and covering the delay with a spinner; there is no job setting that turns Sume STT into a stream. The second is using a streaming model to caption finished video and then discovering that the page lists no word timestamps for it. If you are unsure which side you are on, ask whether the audio exists as a file before the text is needed. If it does, use a file job; if it is still being spoken, use a stream.