Get accurate transcripts

Every cut, caption and subtitle in Cutscript is placed by the time of a word, so the transcript is the one thing worth getting right before you edit. It is also a paid cloud step: your audio goes to the provider you chose. This page covers the choices that change the result. The buttons themselves are in Transcription.

Choose the model once

Open Settings › AI › Transcription. The default, Recommended, takes the first provider you have connected, in this order: ElevenLabs, fal.ai, then OpenRouter. These are the models it picks, and what they offer:

The list has more models than these, with live prices. Sort it by Price, or search it, before you pick. Choose a model that labels speakers if there is more than one voice in the recording. Choose one that tags sounds if you want laughter and music marked in the script. A model that cannot give a time for every word cannot drive editing by word, and the list says so. A ChatGPT or Claude plan cannot transcribe at all.

Say the language and the speakers

Under Options in the same page:

• Language: Auto-detect works for most recordings. If you know the language, naming it can help.
• Identify speakers: labels who is speaking. Turn it off for a single voice.
• Expected speakers: a hint from Auto-detect to 8 speakers. ElevenLabs only.
• Mark sounds: marks laughter, music and other sounds. Cutscript never treats these marks as filler words.

Where a model labels speakers one part at a time, the labels can change between five-minute parts. For an interview, use a model that labels the whole file at once.

Decide before you edit

Re-transcribe… replaces the transcript, and it resets your script edits: deleted, ignored and hidden words, corrections, speaker names and shortened pauses. So settle the model and the language first. Skim the first minute of the result. If the language is wrong or the words are poor, re-transcribe at once, while you have not edited anything.

Cutscript does not start a second transcription of a project that already has a transcript. Only Re-transcribe… replaces one. Each new transcript is another charge from your provider.

Give it clean speech

The audio sent to the provider is the dialogue track, extracted as mono, 16 kHz audio. For a Cutscript recording that is the microphone track. For an imported file it is the file's own audio. Clear speech from a close microphone gives the best words. Music or other voices behind you make the transcriber guess. See Get clean sound.

Check that fillers are in the text

Remove filler words can only mark a word that is in the transcript. Cutscript asks Scribe v2 for a verbatim transcript, which keeps "um" and "uh". Other models may leave some of them out. If the tool finds fewer fillers than you hear, the transcript never had them. Skim the script once before you rely on the count. See Remove filler words and pauses.

Fix wrong words the cheap way

Correcting a word changes its text and nothing else. The audio, the timing and your cuts stay as they are, and captions, subtitles and transcript exports use the corrected text.

1. Select the wrong word and press C, or right-click and choose Correct….
2. Type the right text and press Correct.
3. For a name that is wrong everywhere, choose Correct all instances… and use Correct all to fix every place it appears.

Fix names, numbers and product terms before you burn captions into a video. Also fix a filler the model wrote as another word, since the filler search matches the corrected text.

To label speakers properly, click a speaker's label in the script and rename it. Subtitle and transcript exports can then show the names.

Related

• Transcription
• Edit by editing text
• Spend less on AI