Deepgram's 2 GB limit says extract audio first: use Sume audio detach

Deepgram's pre-recorded audio page gives a maximum file size of 2 GB and tells you to extract the audio stream first when the file is a large video. Sume's audio detach is that extraction step for a video already on media.sume.com: it returns a new wav or mp3 file and leaves the video untouched.

Deepgram's guidance is from its pre-recorded audio page, read 2026-10-03. Sume's is from audio detach.

What Deepgram says

The page is short on formats, and points to a separate list.

Why extract first

A video file carries a video stream that a speech API does not use. Sending only the audio makes the upload far smaller and keeps you well below any file ceiling. Deepgram's own advice is to extract it yourself rather than send the whole video.

On Sume, the video is already hosted, so the extraction runs next to it and you pass on only a URL to the audio.

The Sume request

The source must already be a media.sume.com video, such as a Sume-generated clip, or a TikTok or Instagram video mirrored with POST /v1/media-imports ($0.15). Then call POST /v1/audio-detach with that URL. Defaults are wav and the source channel layout. For speech, channels: "mono" and sample_rate: 16000 is the documented STT shape. Each job costs $0.01.

Output is capped at 900 s, and a source over 1800 s is refused with source_duration_exceeded. For a source between 900 and 1800 s, use range and take it in pieces.

• format mp3 is 128 kbps, wav is the default.
• Check probe.has_audio before paying.
• The result gives an audio_url to pass on.

Or skip the hand-off

If you only need a transcript, Sume's video inspect can transcribe the clip itself with transcribe: true, at $0.01 per audio minute, and returns text, words and optional sentence segments. Use detach when you want to send the audio to Deepgram or another service for features Sume does not offer.

A short checklist

Before you detach, confirm the video has audio with video inspect and frames: false. After detach, check duration_seconds against source_duration_seconds so you know whether the whole track came through.

Send Deepgram the returned audio_url only if it is reachable by Deepgram, or download the file and upload it, since the page you read lists both wav and mp3 as example files.

Size you can expect

A 2 GB ceiling is far above what Sume's detach can produce, since output is capped at 900 seconds. As arithmetic, a 900 second 16 kHz mono wav is about 28.8 MB, and a 900 second mp3 at 128 kbps is about 14.4 MB. Neither is near the Deepgram limit.

These sizes come from the bit depth and rate, not from either vendor's page, so read the real file size after the job.

Run the detach once per video and reuse the returned file if you send it to more than one provider, so you pay the $0.01 only once.