Extract audio from a video for transcription: 16 kHz mono wav

To transcribe a video, detach its audio first as 16,000 Hz mono wav, then pass the new audio URL to speech-to-text. Sume's audio detach docs call that shape the STT shape. Detach costs $0.01 per job and leaves the video untouched. Transcription then costs $0.01 per audio minute.

Steps

• Import the video with POST /v1/media-imports if it is not already on your workspace's media.sume.com host. Audio detach does not fetch the open internet.
• Call POST /v1/audio-detach with an Idempotency-Key header and video_url, format: "wav", channels: "mono" and sample_rate: 16000.
• Poll GET /v1/jobs/:id/status, then read GET /v1/jobs/:id/result for audio_url and duration_seconds.
• Call POST /v1/stt-1.0/transcribe with that audio_url and the duration, rounded up to a whole second.

The limits

A long video

The duration_seconds hint on the transcribe request tops out at 600 s, so keep each request to 10 minutes of audio or less. A detach output is capped at 900 s, so for a longer video use the range field, {start, end} in seconds, with one detach per window; the source can be up to 1,800 s. If the whole track is 900 s or less, you can instead detach once and split with Timeline audio, which takes up to 20 ranges in a single $0.01 job. Add each window's start offset back to the word times when you stitch transcripts together.

Errors to expect

• detach_source_has_no_audio: the video has no audio track. Check first with video inspect.
• audio_detach_range_empty: the range end is not after its start, or the range is over 900 s.
• unsupported_media_source: the URL is not on the Sume media host.

Limits

16 kHz mono is smaller and fast to process, and it discards stereo and high-frequency content; keep the original for anything you will listen to. Costs add up per window: a 30-minute video is three 10-minute windows, so three detaches ($0.03) plus 30 audio minutes of transcription ($0.30), or $0.33 in all. See also subtitle cues from sentence segments.