How to Separate Speakers on a Panel Recording (2026)

How to separate speakers on a panel recording is the pass you do after the draft exists. The transcript already says Speaker 1, Speaker 2, and whoever else the model invented. Your job is to turn those numbers into the moderator and the named panelists, then fix the mess panels actually contain: overlap, laughter, and someone talking over the host.

This is not the general labeling guide. Two-person files and ordinary meetings are covered in how to add speaker labels to transcripts and how to name speakers in a transcript. Those assume you can hear a clean handoff. A panel often does not give you one. If you do not have a draft yet, record and upload first: how to transcribe a panel discussion.

Start from the numbers on screen

Open the transcript in File Transcribe. Do not re-upload hoping for a cleaner split. The second pass is listening.

You need the roster beside you: moderator, then each panelist. If you were not in the room, the introduction on the tape is the roster. Play it before you type a single name.

Click any segment to hear that moment. A panel label is wrong often enough that renaming from the text alone will bake the error into every later turn.

Step 1: Build a scratch map from the intro

Pause after the welcome. Write this on paper, not in the transcript yet:

• Which speaker number says "I'll be moderating"
• Which number answers the first handoff ("Priya, start us off")
• Which number is the same human again later, because the model split them

A hot lapel on the moderator commonly produces two symptoms. The moderator owns Speaker 1 for the whole hour, or the moderator is Speaker 1 and Speaker 4 because a pause made the model think a new person had arrived. Panelists with quieter mics get merged into whoever talked just before them.

The map can say "Speaker 3 = also Jordan." That is a useful note. It is not a failure.

Check the map against a second handoff before you believe it. One intro is not enough if the room laughed through the names.

Step 2: Rename once, then reassign the strays

When the map survives two handoffs, rename.

In the editor, open the speaker pill on any turn and choose Rename speakers. Type the moderator and each panelist next to the number you mapped. Accept. Every block with that label updates. You should not be retyping "Priya" on forty lines.

If two numbers are the same person, give both the same name, or click the stray blocks and point those pills at the correct person. Renaming the label does not merge audio. It only changes the name printed on that number.

Still wrong on one line? Change that block's speaker from its pill. Leave the global name alone. A single bad turn is normal. A bad global rename is how Sam's quotes become Lena's.

The general version of "rename Speaker 1" is the naming guide. On a panel, do not follow it until the scratch map exists. Speaker 1 is often the moderator, not the first famous person on the flyer.

Step 3: Pull the moderator back out of other people's turns

Moderators speak in short bursts. "Priya?" "Thirty seconds." "Let's take one from the room." "We have to land this."

The model likes long turns. It glues those bursts onto the previous panelist, or it drops them. The recap then looks like the panelist asked themselves the question.

Scan for prompts:

• A name said like a handoff
• A time warning
• A sentence that sets up the next topic and does not answer it

Play it. If the voice is the host, reassign that block to the moderator even when the line is four words. Short lines are the structure of the panel. Without them, the transcript is a pile of speeches with no order.

Step 4: Overlap and laughter

Two panelists answering the same prompt is the usual pile-up. You will see one paragraph that changes opinion halfway through, or two labels that swap every few seconds, or a line that is just vowels.

Listen twice, once for each voice.

• Keep the clearer phrase on the person who said it.
• If you can hear both points and they matter, split them into two turns and name each one. Do not invent words you did not catch. Write that a stretch was unclear if the quote would be fake.
• Laughter, applause, and "ha ha" as fake words can go. Delete the garbage line, or leave a plain note if the laugh was the reaction a writer needs. Do not give the laugh its own panelist name.
• A muttered "right" that stole a whole turn can go back to the person who was actually speaking, or it can be cut if it is not a quote.

Agreement noise is not a speaker you must honor. Spend the time on overlaps you will quote, and skim the rest.

Step 5: Someone talks over the moderator

This is the panel-specific failure. The host tries to close a topic. A panelist keeps going because they have the louder mic, or because they do not want to stop. The draft picks one of them and writes a paragraph that starts as a cutoff and ends as a speech.

Play the boundary. Give the moderator the cutoff line back, even if it is "Sam, we need to move." Then leave the panelist's continuation on the panelist. The reader should see that the floor was contested. If you fold it into one speaker, the writeup will say the host agreed with a point the host was trying to end.

Do the same when the moderator interrupts to correct a fact. That correction is not part of the panelist's sentence.

Step 6: Audience, and the voice you cannot place

A question from the room shows up as a new number, as noise, or as a random panelist. If you can tell it is not on the roster, add a speaker slot and name it Audience. One label for the room is enough unless you know the asker's name and they consented to be named.

If you still cannot tell two panelists apart after the intro and two handoffs, stop renaming and mark the doubtful stretches. A wrong name on a public panel is worse than "panelist, unclear." Go back to the video if you have one. Faces solve what the mix will not.

Step 7: Export after the names are stable

TXT or DOCX for the recap. SRT or VTT if the panel video needs captions (Pro or Plus). If you rename after you export, export again. The file on disk will not update itself.

Spot-check the export, not the whole hour: the intro, one handoff, one overlap you repaired, and the line you will publish. If those four are right, the pass worked.

Upload path, if you are still on the raw file: interview recordings for a sit-down, or the panel steps if it was a stage or a Zoom row. Zoom audio notes: Zoom meetings.

FAQ

Why is the moderator split into two speakers?

A pause, a laugh, or a much hotter mic than the panel. Map both numbers to the moderator after you hear the intro, then rename both or reassign the stray blocks.

Should I delete all the laughter?

Delete laugh lines that the model typed as nonsense. Keep a short note only when the reaction is the point of the quote. Do not create a speaker named Laughter.

What if I cannot tell the panelists apart?

Do not guess for the sake of a tidy doc. Use the intro and the handoffs. If a stretch stays ambiguous, label it unclear. Check the video before you attach a public name.

Can I fix one turn without renaming everyone?

Yes. Change the speaker pill on that block. Use Rename speakers only when the number itself is the wrong person everywhere.

Fix the host first

The moderator's short lines are the skeleton. Put those back, then name the panelists from the handoffs, then touch overlap you will actually quote. Upload is on File Transcribe if the draft is not open yet.

Related: Transcribe a panel discussion · Add speaker labels · Name speakers · Zoom meetings · Pricing

More guides

• Detect topics and keywords with AI
• AI sentiment and intent in transcriptions
• How AI transcriptions save time
• Test transcription accuracy
• Transcription guides

Further reading

• Google Speech-to-Text docs
• Zoom cloud recording help