Transcription accuracy for deal teams

The short answer

Junior records the lowest error rate in all four categories we measure, against all three of the generic tools deal teams usually arrive with; on entities, Junior is wrong 6.1% of the time, against 58.2% for Otter.ai, 71.5% for Fireflies and 75.3% for Granola.

Entities are the company names, product lines and people in a call; key points are the statements that carry the meaning. In diligence work those are the only parts of a transcript anyone actually uses, and they are badly served by the single number most buyers compare tools on.

Otter.ai, Fireflies and Granola are the three we get asked about most, and they are good products for the market they were built for, which is general meetings. None of them was built for a diligence call, and the benchmark shows it in the same place every time.

If you evaluate on word error rate alone you will not see any of this; word error rate counts every word the same way, so a tool can miss nearly every name and figure in an hour-long call and still post a perfectly respectable score.

What we measure, and why it is four numbers

Junior reports word error rate, number error rate, entity error rate and key point error rate; only the first of those is standard in the industry, and on its own it tells you very little about deal work.

A one-hour expert call runs to roughly 9,000 words, of which the target's name, its competitors, its product lines and the figures discussed might account for 200. Those 200 are what ends up in the deliverable; the other 8,800 are connective tissue, and nobody has ever gone back to a transcript to check an "and".

So get all 200 wrong and word error rate shifts by around two points, which reads as a rounding difference on a procurement sheet; meanwhile the transcript is useless for the job it was bought to do.

That is why we publish four of them: word error rate tells you whether a transcript reads cleanly, while entity, number and key point error rates tell you whether it is right about anything an analyst will quote, chart, or take into an IC.

The benchmark

Error rates below, lower is better, all measured on public earnings-call audio.

We used earnings calls because the conditions match a real customer call fairly closely - dense in company names, product lines and figures, several speakers talking over each other, and audio quality in the range of what firms actually record.

Junior posts the lowest error rate in every cell of that table; the three generic tools separate from each other a little on word error rate, then land in much the same place on the two measures that decide whether a transcript is usable.

From conversation to deliverable

An entity error rate between 58% and 75% means the transcript is not a record of the call, it is a first draft that somebody has to reconcile from memory before the work can start.

Three things happen to teams running generic note-takers, and we hear all three in nearly every conversation we have with a new firm.

Junior is built to clear that second bar, with clean transcripts structured the way deal teams work and the quote and its source one click apart in Insight Grid, so anything that lands in a client document can be traced back to the moment it was said.

The firm's institutional memory

Entity errors do not cost you once; they make the archive unsearchable, which is the more expensive problem.

Every call with an adviser, an expert, an industry veteran or a management team adds to what the firm knows; that only becomes an asset if someone can find it again 18 months later.

Take a payments business mentioned in passing on a customer interview: it is peripheral to the live deal so the team captures the call and moves on, and then a year and a half later it turns up as a target. Your earlier calls already hold informed views on its product, its customer relationships and where it sits competitively; none of that surfaces if the name was misspelled, or transcribed five different ways across five calls.

At the entity error rates the generic tools post, anywhere from 58.2% to 75.3%, that is the likely outcome rather than bad luck, and an archive nobody can search is worth roughly what no archive is worth.

Alexandria is the part of Junior that solves this; approved BD conversations and expert calls become searchable across the firm by deal, sector or attendee, and because the entities are right in the first place, the search actually returns the call you were thinking of.

Passing a compliance review

75% of our private equity customers ran a firm-wide ban on transcription before they adopted Junior; the objection was never the software, it was where the audio went and who could reach it.

Junior was built to pass that review rather than to argue with it, and the controls that get a compliance team comfortable are standard rather than something you negotiate for.

Junior is SOC 2 Type II certified, and all data sits on Junior's own infrastructure, not on user laptops and not on third-party servers.

How we built it

The gap on entities comes from domain tuning rather than from a bigger base model, which is why it holds across call types instead of moving around with audio quality.

• A proprietary enrichment stack: a custom named-entity recognition engine tuned for what investors and advisers care about, so deals, companies, markets, products and people; there is no vocabulary list for an admin to keep topped up every time a new target enters the pipeline, which is how most tools ask you to solve this.
• Tuned for M&A language: the same stack handles the jargon and the numerical detail that generic models round off, drop, or mishear entirely.
• Expert network pipelines: we connect directly to nearly all the leading expert networks, so that audio lands in Junior automatically instead of being recorded on somebody's phone and uploaded afterwards.
• Insight Grid: pull an insight out of a project and verify where it came from in one click, so nothing reaches a client document without a source somebody can check.
• Alexandria: approved conversations become a searchable firm-wide library, indexed by deal, sector or attendee, so the work done on one project is still reachable on the next.

Test it on your own calls

Any vendor benchmark is worth checking, including this one, and this one takes an afternoon to verify.

• Take one recorded expert call or management presentation from your own pipeline, with real company names, real product lines and real figures in it.
• Run it through whatever you use today and through Junior, using the same audio file for both.
• Ignore how the two transcripts read; count the entities, count the figures, and score only those.
• Then answer one question: could an associate paste this straight into a deliverable, or would they rewrite it first?

Common questions