Transcribe and subtitle

This lesson gets the words out of a recording: a transcript, a subtitle file, or subtitles burned into the video. You'll transcribe the lighthouse narration with Gemini 3.5 Transcribe, fix the SRT file by hand, and burn subtitles into a video with Scenario Caption Studio. It's also the fastest way to check any speech you generate.

Before you start

Pick the output first, then the model:

Transcription lives under Create › Text ("Transcribe audio, summarize images, or run text-to-text models"), not under Audio.

Step 1: Open the transcription model

1. Open Create › Text.
2. Pick Gemini 3.5 Transcribe in the model menu.
3. Under Audio, add your recording: drag and drop it, click Select from Library, or Upload. Here, the ElevenLabs v4 narration from Narration and speech.

lighthouse-narration-elevenlabs-v4-alice.mp3Audio: lighthouse-narration-elevenlabs-v4-alice.mp3

Step 2: Help it with names and language

• Language hints: type a language code and click Add, here en-US. Leave it empty and the model detects the language among more than 85, even when a speaker switches mid-sentence. A hint helps most on short clips, accents and close dialects.
• Custom vocabulary: add names, product terms and invented words, one at a time with Add, so they're spelled your way. Here, Keeper. Keep the list to distinctive terms; everyday words weaken it.

Step 3: Choose the settings and generate

Under Settings:

• Speaker diarization labels who speaks (speaker_0, speaker_1 and so on). Turn it on for interviews and dialogue.
• Word timestamps splits the transcript into timed, numbered segments. Turn it on if you'll fine-tune subtitles. Here it's on.

Either setting lowers the longest file the model accepts from 60 minutes to 30.

Click Generate. You get two files in your session: the text, and an SRT subtitle file.

Step 4: Fix the SRT by hand

An SRT is plain text: a number, a time range, and the words, then a blank line, repeated. Open it in any text editor. The narration came back as one subtitle:

1
00:00:00,300 --> 00:00:12,500
Every evening at six, the keeper climbs the 112 steps of the lighthouse. The lamp is old, the stairs are older, and the cat, as usual, is already at the top.

Every word matches the script, but the subtitle is far too long to read. To fix it:

• Fix words in place and keep the time lines.
• Split long subtitles at a comma or a breath: one or two short lines each, on screen for no more than about seven seconds. Give each part its own number and time range.
• Keep the numbering in order after a split.

The same narration, split at a breath:

1
00:00:00,300 --> 00:00:05,900
Every evening at six, the keeper climbs
the 112 steps of the lighthouse.

2
00:00:06,300 --> 00:00:12,500
The lamp is old, the stairs are older,
and the cat, as usual, is already at the top.

Decide how numbers are written. The narration said "a hundred and twelve"; the transcript wrote the digits. Either is fine, as long as it's the same everywhere.

Step 5: Burn subtitles into the video

When a platform can't take a subtitle file, burn the subtitles into the video with Auto Subtitles or Scenario Caption Studio. Find them by name in the model menu, and add the video under Input Video.

In Scenario Caption Studio:

• Style Preset: the look of the captions. The film uses Modern Chip, a rounded background behind the text. The others include TikTok Bouncy, Karaoke Fill, Karaoke Underline and Word Pop.
• Translation › Target Language: Auto keeps the spoken language; pick another to translate.
• Text Segments: Add Text Segments puts your own timed lines on screen instead of transcribing, for a title or a hook.
• Subtitles (SRT): use a fixed SRT for the text and timing, or get the SRT file out with the video.
• Segment limits: shorter segments, with fewer words on screen at once, are easier to read on a phone.

Auto Subtitles is the simpler choice: a fixed style you adjust with font, size, colors and border, the original language or a translation to English, and limits on how long a line stays on screen.

Generate, then watch the result on the device it's made for: a phone for vertical video, a TV or laptop for a trailer.

Step 6: Prefer a file where you can

If the platform takes an SRT, deliver the SRT. You can fix it any time, and players and platforms load it. Burned-in subtitles can't be turned off or corrected without generating the video again.

Use transcription as a quality check

Transcribe every generated speech take and every instrumental:

• Speech: compare the transcript with the script word by word. A changed word, such as "night house" for "lighthouse" after a voice change, is easy to read and hard to hear.
• Music: an instrumental should return no speech. Any text means the model sang or spoke something.

Common problems

What's next

That's the course: music, voice and sound on Scenario, and every take checked before it ships. To go back over model choice and prompts, see Audio on Scenario; to automate transcripts for a series, put Gemini 3.5 Transcribe in a workflow.