How to caption a music video in Premiere Pro — word by word, on the beat

6 min read · by the CapBeat team

Type "auto captions music video" into any forum and the answer is always the same: "recognition doesn't work on songs — type the lyrics by hand." That advice is outdated. Here's the workflow that gets word-by-word lyric captions onto a Premiere timeline in about fifteen minutes — including why the usual way fails.

Why your transcription comes back as gibberish

Speech recognition models (including Whisper, the best openly available one) are trained on spoken language. Put drums, bass and synths under the voice and the model literally hears percussion where consonants should be. It starts guessing — you get fragments, wrong words, gaps.

The fix isn't a better recognition model. It's giving the model what it was trained on: a clean voice. That's what vocal separation does — an AI model (Demucs, open-sourced by Meta) splits the finished mix into vocals and instrumental. Transcribe the isolated vocals, and suddenly every word lands. On our reference clip that's 14 recognized words without separation versus the complete 47-word lyric with it.

What you need

• Adobe Premiere Pro 2023 or newer on Windows 10/11
• The CapBeat plugin — the only Premiere plugin with vocal separation built in. Music mode is part of Pro (€79 one-time); the free tier lets you test the full speech workflow first.
• No cloud account, no upload — everything below runs on your machine.

Step 1: Install and open the panel

Download the installer from the download page and run it (Windows shows its standard warning for new installers — "More info" → "Run anyway"). Start Premiere and open Window → Extensions → CapBeat. On first start the plugin fetches its AI models once (~1.5 GB); after that it works offline.

Step 2: Open the sequence, set the range

Open your music video sequence. If only part of it is sung, set in and out points and tick "Only between in and out points" — separation takes roughly as long as the material, so a tight range saves real time.

Step 3: Hit the music button

Click "♪ Transcribe music – detect lyrics". CapBeat now exports the sequence audio, separates the vocals from the instrumental on your CPU, and transcribes the isolated voice. For a three-minute song expect roughly three to five minutes total — Premiere stays fully usable while it runs.

Model tip: "Medium" is the sweet spot for lyrics. "Large-v3" buys a little more accuracy on dense or fast vocals at double the time.

Step 4: Fix the words that need fixing

The transcript appears word by word, and here's the part that saves your evening: words the recognition was unsure about are highlighted. You don't proofread everything — you check the marked spots. Click a word, type the correction, press Enter. Ad-libs and stacked harmonies are the usual suspects.

Step 5: Style, preview, render

Pick a caption style, your font (every font installed on your system is available), colors and words per line — the preview is pixel-identical to the final render. Then hit render: the captions arrive as a transparent clip on the next free video track, word-timed to the vocal. Your edit underneath stays untouched.

Common problems, honest answers

The vocals are still not recognized well

Extreme metal vocals, heavy autotune-stacking and very dense mixes remain hard even after separation. Test with the free tier on your actual material before buying — that's what it's for.

Can I do this without a plugin?

Partially. You can separate vocals in external tools, transcribe the stem, then rebuild timing by hand in Premiere — it works, and it costs you an afternoon per song. The plugin exists because that afternoon got old.

Does this work for karaoke-style videos?

Yes — word-by-word timing with a highlighted active word is exactly the karaoke pattern. Combine it with the drop-in animations for intros and titles.

Want the bigger picture first? Read our honest comparison of all caption plugins for Premiere Pro.