Video Transcription Service

Uploading a video vs extracting its audio

Uploading the video file works and is what most people do. Extracting the audio first is faster when you control the source.

A one-hour video carries an order of magnitude more data than its audio track. Uploading the full file over a slow connection is where uploads time out and users give up. Strip to audio first and the transfer drops by roughly the compression ratio of the video codec — typically 10x — while transcription quality is unaffected, because the model only ever listens to the audio anyway.

What to expect with picture in the file

Picture does not help accuracy and can hurt it, indirectly. A talking-head interview and a screen recording with a voiceover transcribe equally well.

What actually affects accuracy is the audio track itself: microphone distance, room noise, music beds, and whether speakers overlap. A high-resolution 4K export with a lav mic beats a 480p capture of a laptop microphone every time.

Do not re-export a low-quality file in an attempt to improve it. That cannot add detail that was never captured.

From transcript to subtitles

For a web video, subtitles materially change reach: they are how a meaningful share of viewers consume, and they are what makes the content indexable as spoken words rather than an unwatched video.

Export SRT or VTT with word-level timestamps, then check cue boundaries against the picture. Auto-timing drifts over long files, and badly timed subtitles are worse than none — they pull the viewer's attention to the timing rather than the content.

Break long passages into shorter cues. A cue spanning three sentences is unreadable at playback speed.

When the video is the hard part

Screen recordings with heavy background music, phone-shot video of a presentation, and recordings with a single speaker more than a few feet from the microphone all need a hearing check before you process a long batch.

Upload a short clip first. Ninety seconds tells you whether the result will be usable, and costs a minute instead of discovering it forty minutes into a two-hour render.