Vendor Says 95%? Verify Who Said What First

AgentLeverage Team

AgentLeverage Team

2/11/2026

#transcription#speaker-diarization#meeting-summarization#accuracy#ai-tools
Vendor Says 95%? Verify Who Said What First

A vendor percent is measured on a quiet room and one native speaker. Your file is a laptop mic, several people talking over each other, and a last name the model has never seen.

That number does not describe your call. Verify who said what on your audio. You still listen back.

What vendor accuracy percents actually measure

Every transcription vendor leads with a percent. The number is almost always from clean-room audio: one speaker, a studio mic, no overlap.

Your recording is usually none of those. A percent on a marketing page does not transfer to a conference-room laptop, a phone export, or a five-person standup.

Word-error arithmetic on that percent is theater. It assumes the number applies to your file. It does not.

The honest check is to run your own recording and listen back. Wrong labels, dropped turns, and mangled names show up there, not on the vendor page.

We do not advertise a 95% or WER number. Observed use here is people labeling speakers. We do not claim to beat a vendor percent.

If you are still choosing software against a studio demo: choose transcription software for your audio, not the vendor clip

What label-speakers-first returns

Speaker Separation takes a recording and returns a labeled transcript.

Upload the file. The run labels each voice (Speaker 1, Speaker 2). You rename those labels. You can download isolated stems (one audio file per voice) and export txt or srt.

Guest runs are free (2/day). Signed-in orgs pay 2 credits per minute of audio, rounded up, with a 1-minute floor.

That is the job: who said what on a recording you already have. It is not a certified transcript. It is not a meeting bot that joins Zoom. It is not a meaning extractor that skips word errors.

You still review. Overlaps and wrong labels still need a listen.

If the file is any multi-speaker recording and you still need who said what: how to split audio by speaker, then rename the labels

Where errors still show up

A labeled transcript is a first pass. These conditions still break turns:

  • People talking over each other
  • Industry jargon, product names, and last names
  • Accents and regional speech the model rarely heard
  • Phone audio, room echo, or a cheap laptop mic
  • Two voices that sound alike

Isolated stems help you hear one voice at a time. They do not certify the text. You still listen back.

We do not publish a drop-off percent for overlap or cheap mics. Those conditions fail in different ways on different files. The review is the check.

If the job is a deposition recording and you still wait on the certified copy: label speakers for a working transcript, not a court record

Meeting Summarizer is a recap, not a recording tool

Meeting Summarizer does not take a meeting recording. Meeting Summarizer does not transcribe audio. It does not replace a word-error check.

Once you have a transcript you have already checked, it can turn pasted notes or that transcript into a recap, decisions, owners, and action items when those were said.

Guest runs are free (5/day). Signed-in orgs pay 2 credits per summary. It accepts TXT, DOCX, VTT, and SRT.

Skip it if you only need the labeled transcript. Transcribe and recap are two tools. A recap of an unchecked transcript just hides the errors.

How to verify on a real recording

Pick a real recording. Not the vendor clip.

  1. Upload it to Speaker Separation. Guest runs are free (2/day).
  2. Rename Speaker 1 and Speaker 2.
  3. Listen back. Check overlaps, names, and jargon. Download stems or export txt/srt if you want them.
  4. Optional: paste the transcript you checked into Meeting Summarizer for a recap.
  5. You still review before you share.

No account for a guest run. Sign in if you want the job kept. Start with 30 free credits. Pack credits never expire. Unused included credits do not carry forward. Paid checkout charges immediately. There is no 7-day trial.

Try Speaker Separation →

Frequently asked questions

Is this 95%?

No. Vendors advertise a percent on clean-room audio. We do not advertise a 95% or WER number. Test your file. You still listen back.

Is this a certified transcript?

No. It is a working labeled transcript. You still review. Do not file it as an official record.

Does Meeting Summarizer transcribe audio?

No. Meeting Summarizer does not transcribe audio. Paste notes or a transcript you already checked (TXT, DOCX, VTT, SRT). It does not take a recording.

Can I try it without an account?

Yes. Guest runs are free (2/day). Sign in if you want the job kept until you delete it.

What does a signed-in run cost?

2 credits per minute of audio, rounded up, with a 1-minute floor. Start with 30 free credits. Pack credits never expire. Unused included credits do not carry forward. Paid checkout charges immediately. There is no 7-day trial.


Need who said what on a recording that is not the vendor demo? Upload one. Guest runs are free (2/day). You still listen back.

Try it yourself

Speaker Separation

Upload a recording. This speaker separation labels multi-speaker audio. Rename Speaker 1 and download isolated stems plus a timestamped transcript.

Start with 30 free credits — no card required.

Get weekly AI tips

Practical AI productivity tips every week. No fluff.

Vendor Says 95%? Verify Who Said What First | AgentLeverage