Transcription accuracy vs. transcription quality: why the gap matters

Perceived transcription quality measures how accurate a transcript feels to the person reading it—factoring in speaker labels, formatting, punctuation, and whether names and numbers came through correctly. Word error rate measures how many words a model got wrong. The two diverge constantly, and users only ever experience the second one.

Here's the concrete version. Transcript A misses eleven filler words across a thirty-minute call. Transcript B gets every filler word right and mislabels who said the closing commitment. Identical word error rate. One is fine; the other reverses who agreed to what, and no aggregate score will ever tell you which one you shipped.

The gap between accuracy and perception

This is what WER cannot see. It treats every word as equally important, and the person reading the transcript emphatically does not. Nobody has ever noticed a missing "um." Everybody notices their own name spelled wrong.

The economics of that gap are more concrete than they sound. Joshua Grossberg, CTO of Kapwing, puts it in the unit that actually gets budgeted—review time:

"If you have an hour of content, the difference between 99% accuracy and 97% accuracy, it's a lot of time for that person to review. So you could cut down their workflow from taking half an hour, taking 20 minutes, taking 15 minutes — it's huge, right?"

Joshua Grossberg, CTO, Kapwing

Two percentage points. Fifteen minutes per hour of content. That's the conversion rate between a number on a benchmark sheet and a number on a payroll, and it's why "97% vs 99%" is a much bigger deal than the arithmetic suggests.

It also cuts the other way, which is the part vendors don't advertise. A model can improve its WER and get worse to use, if the improvement lands on words nobody was reading and the regression lands on names.

End users are unambiguous about which failures they notice. In AssemblyAI's 2026 Voice Agent Report—a survey of 455 builders fielded across Q4 2025 and Q1 2026—55% named "having to repeat themselves" as the top end-user frustration, and 45% report frequently misheard words. Fully 95% said they had been frustrated by a voice agent at some point. Not one of those complaints is phrased as a word error rate, and every one of them is a transcript problem.

Which makes the target builders set for themselves interesting. In the same survey, 47.5% said they are aiming for better than 95% accuracy—a threshold expressed in words, against a set of complaints expressed in names, repeats, and misheard numbers. That mismatch is the whole subject of this post: the bar people write into their requirements is not the bar their users are judging them on.

Audio tags: when accuracy backfires

Here's a case where being more correct made things worse.

Tag every non-speech event in an audio file—laughter, coughing, a door closing, background music, a pause—and you've produced a more complete and more literally accurate record of what happened. You've also produced a transcript that reads like a stage direction, and users hate it.

The reason is that a transcript is not a recording. It's a representation, and every representation makes editorial choices about what to include. Readers have strong, mostly unstated expectations about those choices, formed by decades of subtitles, meeting minutes, and court records. Violate them and the transcript reads as wrong even when every character is right.

Which is why the same audio tag can be essential or intolerable depending entirely on the job:

• Accessibility captioning: [laughter] and [music playing] are required. Omitting them is the error.
• Meeting notes: the same tags are noise. Nobody wants their standup minutes annotated with coughs.
• Qualitative research: pauses and overlaps are data, sometimes the most important data in the file.
• A voice agent transcript: tags are actively harmful—they get fed to an LLM that now has to reason about a door closing.

Same model, same accuracy, four different right answers. That's not something you can benchmark your way out of, and it's the clearest illustration that "quality" is a property of the fit between output and use, not of the output alone.

Formatting works the same way. Punctuation, casing, and inverse text normalization—turning "twenty five dollars" into "$25"—change no words and transform how a transcript reads. We've covered the mechanics in automatic punctuation, casing, and ITN, but the point for this argument is that all of it is invisible to WER, because punctuation is normally stripped before scoring. An entire dimension of perceived quality is measured at exactly zero.

Speaker diarization: the trust multiplier

Speaker labeling errors damage perceived quality more than word errors do, because a mislabeled turn changes who said something rather than just what was said—and AssemblyAI optimizes diarization for cpWER (30.17 average) rather than DER for exactly that reason.

Consider what a diarization error actually does to a reader. A word error makes them think the model is imperfect. A speaker error makes them think the model is confused about the fundamental structure of the conversation, and that suspicion generalizes: if it doesn't know who was talking, why would I trust anything else in here? One bad label poisons the whole document.

The standard metric for speaker diarization is DER—diarization error rate—which scores speaker segmentation in isolation from the words. It answers "did the model correctly identify when the speaker changed?" without asking whether the resulting transcript is usable.

cpWER—concatenated minimum-permutation word error rate—scores the transcript and the attribution together. It asks the question a reader would ask: is the right text attached to the right person?

Universal-3.5 Pro is optimized for cpWER rather than DER, which is a small choice with a large implication. It means the model was tuned on what users experience instead of on the thing that was easiest to measure. Which is, more or less, this entire post's argument applied to a model training objective.

Measured across meetings, telephony, far-field, and conversational audio—deliberately including the conditions where diarization is hard, rather than the two-person studio recordings where every model looks competent.

The hard problem of streaming corrections

Streaming diarization has a problem batch diarization doesn't: it has to commit to a speaker label before it has heard the conversation.

Speaker identity in audio is a clustering problem, and clustering works better with more data. Thirty seconds in, the model has heard two voices briefly and has to guess. Twenty minutes in, it has enough signal to be confident—and to know that the guess at second thirty was wrong.

So you get to choose your failure. Never revise, and early labels stay wrong forever. Revise continuously, and the transcript flickers while the reader is trying to read it, which is worse. Watching speaker labels reshuffle in real time is one of the most viscerally untrustworthy experiences a transcription product can deliver.

Universal-3.5 Pro Realtime takes a third path: label live, then re-cluster and send a single correction within about half a second of the stream ending, for up to 10 speakers. One visible revision at a natural boundary, rather than continuous churn. The streaming diarization deep dive covers how the re-clustering works.

Note that this is a perceived-quality decision, not an accuracy decision. Continuous revision would produce a marginally more accurate transcript at every intermediate moment. It would also be unusable. The design choice is to be slightly less right in a way nobody minds, instead of slightly more right in a way everybody hates.

What "quality" means in production

Once you stop treating all words as equal, quality resolves into four separable dimensions:

1. Word-level accuracy—did it get the words right? This is WER, and it's the least interesting of the four once you're above roughly 95%.
2. Entity accuracy—did it get the names, numbers, addresses, and product terms right? These are the words users check.
3. Structural accuracy—speaker attribution, turn boundaries, paragraph breaks. Does the transcript reflect the shape of the conversation?
4. Temporal accuracy—timestamps and alignment. Invisible until you build a player or a caption file on top, then suddenly the only thing that matters.

Entity accuracy is where the largest measured gaps between providers sit. On the Pipecat open speech-to-text benchmark, built from real voice-agent conversations, the spread on entities is several times the spread on words. Lower is better throughout, and note these are streaming measurements on voice-agent audio.

Read the top two rows together. On words, the spread is roughly two-fold. On entities, it's more than three-fold—and entities are the only part anyone proofreads. A transcript that scores 9.76% WER and 39.70% on entities will feel dramatically worse to a human reader than the raw word score implies, because the errors are landing precisely where attention is.

The same asymmetry shows up in multilingual audio. Code-switching—moving between languages mid-sentence—is common in real conversation and rare in benchmarks, and it's the fastest way to make a transcript feel broken to a bilingual reader:

Coverage, since it's the obvious next question: Universal-3.5 Pro covers 18 languages with native code-switching built into the model. Universal-2 remains the path for the long tail of 99+ languages.

Where perceived quality bites hardest is anywhere a human reads the transcript to make a decision. In a contact center, a supervisor reviewing a call for a compliance dispute needs the attribution to be right more than they need every word. For a voice agent, no human reads it at all—the entity is the whole payload, and structural niceties matter less than getting the account number right. Different jobs, different definitions of quality, same model underneath.

Why this is hard to benchmark

Perceived quality resists measurement for a structural reason: it depends on the reader's job, and a benchmark has to fix a single scoring function.

That doesn't mean you're stuck with WER. It means you pick the metric that matches your job. Here's the map:

Two of those deserve their own reading. Semantic WER is the attempt to score meaning preservation rather than string match, and why WER alone is insufficient makes the full structural case against the default metric.

There's also a measurement problem underneath the measurement problem: the reference transcript. WER is scored against a human transcript, and human transcripts disagree with each other about filler words, number formatting, and where a sentence ends. Some meaningful share of any published WER figure is transcriptionist convention rather than model error, which is a strange foundation for an industry-standard number.

What transcription quality standards actually require

Outside the API world, "transcription quality" has meant something specific for decades, and it's worth borrowing the vocabulary—because the people searching for this term are often asking an operational question, not a model question.

Verbatim vs clean read. Verbatim transcription captures everything: false starts, repetitions, filler words, stutters. Clean read removes them and produces readable prose. Neither is more accurate—they're answers to different questions. Legal and qualitative research usually require verbatim; business documentation almost always wants clean read. The mistake is delivering one when the reader expected the other, which reads as an accuracy failure and isn't one.

Speaker attribution. Standards typically require every turn attributed, with consistent labels throughout the document and a stated convention for unidentified speakers. "Speaker 1" and "Speaker 2" satisfy this; labels that drift mid-document do not.

Timestamping. Convention ranges from every speaker change, to fixed intervals, to word-level. Word-level alignment is what caption files and audio-synced players need, and it's the dimension nobody checks until it breaks.

Inaudible and uncertain markers. A good transcript flags what it couldn't hear rather than guessing. This is where automated systems have historically been worst and where confidence scores earn their keep—a model that silently invents a plausible word is doing more damage than one that marks the gap.

Map those four onto the four dimensions from the last section and they line up almost exactly. Which is reassuring: the human transcription industry arrived at the same taxonomy decades earlier, by watching what clients complained about.

Perceived quality is the next competitive frontier

The interesting shift isn't that WER is a flawed metric. Everybody working on speech has known that for years, and the field has been publishing critiques of it for longer than most of the current models have existed.

The shift is that WER has stopped being discriminating. When the leading models were at 12% and 18%, the headline number told you something. At 5.6% English mean WER, the remaining differences between serious providers are mostly noise on the aggregate and enormous on the specifics—three-fold on entity errors, meaningful on speaker attribution, invisible-but-decisive on formatting.

So the competition has quietly moved. Not to a better single number, but to the parts of a transcript that a benchmark was never designed to score: whether the name is right, whether the attribution holds, whether the thing reads like a document a person wrote.

Which suggests an uncomfortable test for anyone evaluating providers. Take the transcript your current vendor produced for your hardest real file, and the transcript from a candidate. Don't score them. Read them, and notice which one you'd hand to a customer.

If your scoring script and your gut disagree, trust your gut and go fix the script—because your users don't run the script. The speech-to-text and streaming speech-to-text product pages document which of these levers are configurable.

Frequently asked questions

What is perceived transcription quality and how does it differ from WER?

Perceived transcription quality measures how accurate a transcript feels to the person reading it—factoring in speaker labels, formatting, punctuation, and whether names and numbers came through correctly. Word error rate measures only the share of words substituted, deleted, or inserted against a reference. A transcript can score well on WER and still read as broken if it attributes the wrong line to the wrong speaker.

What is a good word error rate?

There is no universal threshold, because the right target depends on what the transcript feeds. For summaries and meeting notes, 10–15% WER is usually survivable since meaning survives scattered word errors. For a transcript a human will read, or one an agent will act on, entity accuracy matters far more—one wrong digit in an account number fails the task no matter how strong the WER.

How do you measure transcription quality beyond word error rate?

Use metrics that weight the words users actually notice. Missed Entity Rate tracks names, medications, addresses, and numbers; Semantic WER discounts errors that do not change meaning; and cpWER scores the transcript and speaker attribution together. AssemblyAI optimizes Universal-3.5 Pro's diarization for cpWER rather than DER for exactly this reason, averaging 30.17 against Deepgram Nova-3 English at 37.92.

Why do users complain about transcription quality even when WER is low?

Because WER treats every word as equally important and users do not. Misspelling a filler word costs the same in WER as misspelling the customer's name, but only one of those gets reported as a bug. Errors concentrated on names, numbers, and speaker turns dominate perception regardless of what the aggregate score says.

How do you calculate the word error rate?

WER is the sum of substitutions, deletions, and insertions divided by the total number of words in the reference transcript, expressed as a percentage. A score of 0% means a perfect match. The formula is simple, which is a large part of why it persists as the industry default despite weighting every word identically.

How does streaming transcription handle speaker label accuracy in real time?

Streaming diarization has to commit to a speaker label before it has heard the full conversation, which is why early labels are often wrong. Universal-3.5 Pro Realtime labels speakers live, then re-clusters and sends a single correction within about half a second of the stream ending, for up to 10 speakers—one visible revision rather than labels flickering throughout the call.

‍