Gemini SynthID audio check: detach the track from a Sume video

Gemini's SynthID checker takes audio under 1 hour. Detach a Sume video's audio as wav or mp3 and see which format stays under 100 MB.

5 min readSume
All posts

Can I check just the audio of a video for SynthID?

Yes, as a separate check. Google's help page lists audio files under 1 hour and 100 MB per file, with about 10 audio checks per 24 hours (up to 3 hours total). To get an audio file out of a Sume video, use audio detach, which writes the track as a new wav or mp3 and leaves the video untouched.

A miss proves little. The page says SynthID currently works for content from Google's AI, so a voice from another vendor will not be recognised.

Detach the track

Audio detach takes one clip on media.sume.com and returns an audio artifact. The default is a sample-exact wav (pcm_s16le); format: mp3 gives 128 kbps. range takes start and an optional end, channels is source or mono, and sample_rate is 16000, 44100 or 48000. The docs put the source cap at 1800 seconds and the output cap at 900 seconds, so a longer track needs a range. A job costs $0.01 per the docs; confirm it in the catalog.

If the clip has no audio track the job fails with detach_source_has_no_audio. Check probe.has_audio first with video inspect and frames: false.

curl -X POST https://api.sume.com/v1/audio-detach \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: synthid-audio-001" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_demo/ad.mp4",
    "format": "mp3"
  }'

Which format stays under 100 MB

This is arithmetic from the format definitions in the docs, treating 100 MB as 100,000,000 bytes. Uncompressed 16-bit PCM is sample rate x channels x 2 bytes per second.

Seconds of audio before a file reaches 100 MB (arithmetic, 100 MB = 100,000,000 bytes)
FormatBytes per secondSeconds to reach 100 MB
wav, 48 kHz stereo192,000520
wav, 16 kHz mono32,0003,125
mp3, 128 kbps16,0006,250

Shrink the file without cutting the track

Mono halves the size of an uncompressed file. A 48 kHz mono wav is 96,000 bytes per second, which is 86.4 million bytes for a full 900-second detach, under 100 MB. If you need only a section, pass a range with start and end. Reserve the 16 kHz mono wav for speech-to-text, which is what the docs say it is for; it throws away detail above 8 kHz, so use the source rate for the file you intend to check.

Sume's video inspect transcript audio is also a 16 kHz mono wav, but the docs describe it as speech-to-text input, not a full-fidelity rendition, so detach the track instead of reusing it.

What this tells you about a normal clip

A full 900-second detach as 48 kHz stereo wav is 172.8 million bytes, over the limit, so choose mp3 or a mono, lower-rate wav, or cut a range of up to 520 seconds. An mp3 of 900 seconds is 14.4 million bytes. Google's page does not list which audio formats the app accepts, so confirm in the app that your chosen format uploads before you build a routine around it.

Remember that a separate audio check does not replace a video check. The two have their own daily counts on Google's page.

  • Detach the final mix, not a stem.
  • Keep the job id with the result of the check.
  • Do not state a clip is human-made because the check found nothing.

Sources

Related posts

More in Media tools

All Media tools posts

Written by Sume