Transcribing
Which provider, and how long a recording each takes
Section titled “Which provider, and how long a recording each takes”VoxMark sends 16 kHz mono WAV — about 1.9 MB a minute — so a provider’s file-size limit is really a length limit.
| Provider | File-size limit | Longest recording |
|---|---|---|
| Groq (Whisper) | 25 MB free · 100 MB developer tier | ~13 minutes free · ~50 developer tier |
| OpenAI (Whisper) | 25 MB | ~13 minutes |
| AssemblyAI | Around 2 GB per upload | Many hours — no practical limit |
Past about a quarter of an hour you need AssemblyAI, which is also the only provider that identifies speakers. If a render is too big for your provider, VoxMark tells you its size and asks you to switch, rather than letting the upload fail.
On Groq’s paid developer tier VoxMark uses the full 100 MB, with a warning past 25 MB that the free tier stops there. Groq bills a minimum of ten seconds per request and rate-limits its free tier to 20 requests a minute and two hours of audio an hour.
Options
Section titled “Options”Model (Groq only) — Whisper Large v3 Turbo, the default, is faster and places markers more accurately. Whisper Large v3 is slightly better word for word. A wrong word is a double-click to fix in the Navigate tab; a marker in the wrong place isn’t, which is why Turbo is the default.
Language — Auto-detect works for most recordings. Picking the language helps with short or mixed clips.
Detail — Segments gives one marker per phrase; Words gives one per word. Max chars splits long phrases into shorter markers (0 = no limit).
Clear existing comp markers first — removes every marker on the comp before the new ones land. Leave it off to add to what’s there.
Transcript only (no markers) — transcribes and saves the transcript without touching your timeline. Add markers later with Create markers from transcript, from the Transcribe tab or the ⊕ button in the Navigate tab.
Speakers
Section titled “Speakers”With AssemblyAI, speakers are identified by default: each voice gets its own name and marker colour. A recording with a single speaker is left unlabelled.
If you know how many people are talking, enter it in Speakers — an exact number (3) or a
range (2-4). Left on auto, AssemblyAI can merge two similar voices into one; a wrong exact
number makes it worse, so give a range if you’re unsure.
Groq and OpenAI can’t tell speakers apart, but you can assign voices by hand with any provider — see The Navigate tab.
How accurate is the timing?
Section titled “How accurate is the timing?”Whisper tends to report a caption’s start early wherever there’s a pause, sometimes by half a second. VoxMark corrects every caption against the word timings the transcription returns and, where those are wrong too, against the rendered audio itself — measuring where speech actually begins. Where the audio can’t tell, the transcription’s own timing is kept; VoxMark never guesses.
Choosing what gets transcribed
Section titled “Choosing what gets transcribed”Scan Comp lists every layer or track with audio — VO stems, nested comps, several tracks. Untick anything you don’t want in the mix. In After Effects VoxMark renders the comp’s work area; shorten it to transcribe a section.