Azure fast transcription: 500 MB, under 5 h, vs Sume's 900 s detach

Azure's fast transcription API accepts audio files under 500 MB and under 5 hours. Sume's audio detach writes at most 900 seconds of audio per job, from a source video of up to 1800 seconds, so detaching is a way to prepare a track for Azure only when the part you need fits inside those caps. For anything longer, use range to take it in pieces.

Azure's limits are from its fast transcription page, read 2026-10-03. Sume's are from audio detach.

The two ceilings

The limits are of different kinds: Azure's are about the file you upload, and Sume's are about the audio a job may produce.

Taking a long recording in pieces

The 5 hour limit is about 18000 seconds, while Sume's source cap is 1800 seconds. That means Sume's detach will not read a 2 hour video at all, because the worker refuses a source longer than 1800 s with source_duration_exceeded. If your video is 30 minutes or less, you can take it in ranges of up to 900 s each.

Each range is its own $0.01 job. Two ranges cover a 30 minute video: { start: 0, end: 900 } and { start: 900, end: 1800 }. Name your files so the order is clear when you send them on.

The audio shape to ask for

For speech use channels: "mono" and sample_rate: 16000, which Sume's docs call the STT shape, in the default wav format. Azure's page says nothing about which of these you must send, so check Azure's supported formats before choosing mp3.

Because wav here is pcm_s16le, one minute of 16 kHz mono audio is 16000 samples a second at 2 bytes each, about 1.92 MB, by arithmetic. A 900 second range is about 29 MB, well under 500 MB.

• Check probe.has_audio first with video inspect and frames: false.
• A source with no audio fails detach_source_has_no_audio.
• The video file is untouched.

When Sume can transcribe for you

Video inspect has an optional transcript through Sume STT 1.0 at $0.01 per audio minute. If you only need text, that route skips the hand-off. If you need Azure features or a specific region, use detach and send the file there.

Naming and ordering the pieces

When you detach a long video in ranges, keep a list of each range's start and end next to the returned audio_url. After transcription, add each range's start to that piece's word times, so the timings line up with the original video.

Without that offset, every piece restarts at zero and the transcript is hard to align.

Which service holds the long file

The two limits tell you where a long recording should live. A 3 hour interview is within Azure's fast transcription limits as audio, but Sume's detach refuses a source video longer than 1800 seconds, so it cannot prepare that file.

So for long recordings, extract the audio before it reaches Sume, using a tool outside it, and send that straight to Azure.