Transcribe 100,000 two-second voice command recordings: the bill

Transcribing 100,000 two-second voice command recordings with Sume STT 1.0 bills 200,000 audio seconds, which is 3,333.3 minutes, so $33.33 at $0.01 per audio minute prorated by the second. One clip costs about $0.000333. The catch is the hold: if you omit duration_seconds, each job in flight reserves a full minute ($0.01) until it settles. Rate and proration are from Video inspect, read 2026-10-06. This is for recorded clips, such as a log of commands you want to review, not live voice control.

The arithmetic at three scales

For comparison, PYMNTS reports an introductory $0.54 per audio hour for Microsoft's MAI-Transcribe-2-Streaming. Converted linearly, that is $0.0003 per two seconds. The sources I read do not say how that service rounds short sessions, so treat the right-hand column as a straight conversion, not a quote for 2-second clips.

What holds your money while jobs run

A hold is not a charge, but it ties up balance and can trigger a 402 on a small wallet. With a Pro plan's accepted capacity of 24 jobs, 24 holds of $0.01 each total $0.24 if you leave duration_seconds out, against about $0.008 when you send 2. Capacity numbers are from Generation admission, read 2026-10-06.

Compute your own total

Plug in your clip count and length. The result matches the table above.

RATE_PER_MINUTE = 0.01


def stt_bill(clips, seconds_each):
    seconds = clips * seconds_each
    return round(seconds / 60 * RATE_PER_MINUTE, 2)


def hold_without_hint(jobs_in_flight):
    return round(jobs_in_flight * RATE_PER_MINUTE, 2)


print(stt_bill(100_000, 2))
print(stt_bill(1_000_000, 2))
print(hold_without_hint(24))

Is it worth batching them?

Each clip is still one job, with its own submit and result. If you control the recording side, concatenating many short clips into one file up to 600 seconds cuts the number of jobs by orders of magnitude; the price is the same, and you split the transcript by word times afterwards. If you need one job per clip, use the retry-aware submit patterns and the capacity table in the 10,000-clip post.