Transcription Accuracy Explained: What 99% Really Means

Transcription accuracy is the metric every vendor advertises and almost no vendor explains. The short version: "accuracy" means Word Error Rate flipped to a percentage, the benchmark it came from matters as much as the number itself, and the file you upload will rarely match the conditions that produced the marketing claim.

What Word Error Rate Actually Measures

The formula is:

WER = (Substitutions + Deletions + Insertions) / Total reference words

Three error types each count as one point:

• Substitution. "He said yes" transcribed as "she said yes."
• Deletion. "Send the report" transcribed as "send report."
• Insertion. "We met" transcribed as "we have met."

A 1,000-word reference transcript with 50 errors gives a WER of 5%, which becomes "95% accuracy" on the product page. The math is fine. What the number hides is that WER treats every word identically: dropping "not" in "do not approve the transfer" costs one point, and so does dropping "uh." The damage to your workflow is not equal.

WER also does not penalize wrong context. A model that consistently transcribes "two" as "to" will show almost no WER penalty on most audio, but it will break any script that parses numbers.

What Accuracy Bands Feel Like in Practice

Numbers in isolation are easy to misread. For a one-hour interview (roughly 9,000 words):

For most publishing workflows, 95% is the floor and 97%+ is what you should expect from a good AI tool on clean audio. If a tool delivers below 90% on your files, the problem is almost always the audio input, not the model. The companion post on how to improve transcription accuracy covers the specific fixes; the post on dealing with background noise goes deeper on audio prep.

Why Benchmark Numbers Mislead

Vendors choose their benchmarks. Here is what the common ones actually test:

LibriSpeech. Audiobooks read by volunteers in quiet rooms. Models routinely score 97-99% here. Whisper Large-v3 hits approximately 2.7% WER on the clean test set. This is the best-case scenario for any recording.

TED-LIUM. Conference talks with accents and some cross-speaker variation. Harder, but still rehearsed speech. GPT-4o-transcribe reaches roughly 2.5% WER here; AssemblyAI Universal-3 Pro scores around 6.8% WER.

CallHome / Earnings-22 / AMI. Phone conversations, earnings calls, multi-speaker meetings. Whisper Large-v3 scores 26.4% WER on CallHome and 15.9% WER on AMI meeting audio. These are the numbers that reflect what real business audio looks like.

The number on the landing page is almost always from LibriSpeech or a similar clean dataset. The number you get on a Zoom call is closer to the CallHome or AMI number, sometimes worse.

My take: when a vendor quotes "99% accuracy," the first question to ask is "on what dataset?" If they can't answer, treat the number as decorative.

Independent third-party benchmarks routinely produce higher error rates than internal vendor ones. Deepgram's internal benchmarks place Nova-3 at roughly 5-7% WER on their curated test sets; third-party benchmarks on mixed real-world audio put the same model around 18% WER. Neither is wrong. They are measuring different things.

For a detailed breakdown of how individual API providers compare, see the best speech-to-text APIs 2026 post.

What Actually Moves Your WER

The model choice is one variable. These factors move accuracy more on most files:

Audio quality

A clean 16 kHz mono recording with the speaker close to a USB microphone can hit under 3% WER. The same speaker on a speakerphone in a noisy room hits 15-25% WER on the same model. Verified real-world data confirms the range: clean studio audio lands at 3-5% WER for top models; far-field room audio lands at 18-28% WER.

The specific culprits: background noise above -40 dB relative to speech, heavy MP3 compression (64 kbps loses the fricative consonants models depend on), room reverb from hard walls, and distance from the microphone.

Accent and dialect

A model trained mostly on American English will score 96% on American English and closer to 85% on heavily accented speech. Whisper Large-v3 on accented English sits in the 8-15% WER range. The gap has narrowed with multilingual models, but it has not closed.

For non-English audio, using a provider that natively supports your language matters more than picking the highest-rated English model.

Domain vocabulary

General-purpose models trip on medical, legal, scientific, and niche technical terms. "Encephalitis" might come back as "in cephalitis." Brand names and proper nouns are especially fragile because they appear rarely in training data.

Three practical fixes: vocabulary boosting (most APIs let you pass a term list), custom models for high-volume domain work, or a targeted find-and-replace pass on known recurring terms.

Number of speakers and overlap

Single-speaker monologue is the easiest condition. Two speakers in a structured interview is manageable. Three or more people, especially with crosstalk, compounds errors quickly. Diarization errors and recognition errors stack on top of each other.

For multi-speaker meetings, expect 10-15% WER even with a top model unless the audio is well-separated. The speaker diarization explained post covers how diarization interacts with accuracy.

File length

Most modern pipelines chunk audio into 30-second windows with overlap to handle long files. Whisper in particular can hallucinate at chunk boundaries when chunking is naive, producing fluent-sounding text that has nothing to do with the audio. Research published in 2024-2025 confirmed that a small subset of Whisper's decoder attention heads drive most of these hallucinations on non-speech segments. The practical fix is using voice activity detection before chunking to strip silence.

Confidence Scores: What They Tell You and What They Don't

Most modern speech APIs return a confidence score per word or segment. A 0.92 confidence means the model assigns 92% probability to that word being correct.

Confidence is useful for:

• Flagging spans to review. An editor can jump directly to low-confidence words instead of reading the whole transcript.
• Automated quality control. Reject segments below a threshold and re-run with a different model or human review.
• Estimating edit time. A 90-minute transcript with 30 low-confidence words gets reviewed faster than one where errors are spread evenly.

Confidence is not useful for:

• Proving correctness. A model can assign high confidence to a hallucinated word. Whisper is the most-cited example, producing fluent-sounding output on silent or noisy input with high confidence scores.
• Cross-model comparison. A 0.90 from Deepgram and a 0.90 from Whisper use different internal scales. They are not the same number.

Use confidence to find where to look, not as a substitute for looking.

Human Transcription as the Reference Point

Trained human transcriptionists score 98-99% on clean audio and 95-97% on difficult audio. The AI-to-human gap on clean audio is now small enough that the difference rarely justifies the cost for most use cases.

The gap on hard audio is still meaningful. Courts, qualitative researchers, and medical documentation workflows still use human transcription for material where errors carry real consequences.

Worth noting: two humans transcribing the same file will not produce identical transcripts. Inter-annotator agreement on challenging audio sits around 95%, which sets a practical ceiling on what any system, AI or human, can achieve on that material.

Measuring Your Own WER

If you want to know what accuracy you are actually getting on your audio:

1. Take a 5-minute sample of your typical audio, something representative of your real workload.
2. Transcribe it carefully by hand (this is your reference transcript).
3. Run the same clip through your AI tool (this is your hypothesis transcript).
4. Compute WER using the Python jiwer library or the NIST sclite toolkit.

Five minutes of your actual audio gives you a more actionable number than any vendor benchmark. The comparison to aim for is not the marketing page, it is your own reference against your own audio.

If you want a fast check before committing to a paid plan, the audio-to-text tool at ConvertAudioToText lets you upload a sample and inspect the output directly. If your specific audio type produces poor results, you will see it before you pay.

When WER Is Not the Right Question

Three use-case rules:

• Publishing verbatim (podcast show notes, press quotes): aim for 97%+ and plan a 15-20 minute editing pass. A single wrong number or name in a published transcript is the kind of error that gets noticed.
• Search indexing or note-taking: 90-92% is often fine. Errors in filler words and minor variants do not break a search query.
• Quoting specific passages in writing: always verify the specific quote against the audio, regardless of overall accuracy. The overall WER can be excellent while one sentence is wrong.

The right accuracy threshold is set by what the transcript will be used for, not by what sounds impressive in a product comparison.

FAQ

What is Word Error Rate (WER)?

WER is the standard metric for measuring transcription accuracy. It is calculated as (Substitutions + Deletions + Insertions) divided by the total number of words in the reference transcript. A WER of 5% is reported as 95% accuracy. It counts three error types equally: a substituted word, a dropped word, and an extra word each cost one point regardless of how much they change the meaning.

Why do vendors claim 99% accuracy but my files come out worse?

Because benchmark choice matters enormously. Most vendors quote accuracy from LibriSpeech, a clean audiobook dataset where even mid-tier models score 97%+. Real-world audio, especially meetings, phone calls, or noisy environments, produces significantly higher error rates. Gladia researchers have documented the same model reporting 5% WER on a benchmark and 25%+ WER in production on equivalent real-world audio. The benchmark number is not a lie, but it is not your number.

What does 95% accuracy actually feel like on a one-hour recording?

A one-hour interview runs roughly 9,000 words. At 95% accuracy, that is 450 errors, or about 7 mistakes per minute. The text is readable, but proper nouns and technical terms often break, and you will need a 30-45 minute editing pass before publishing. At 99% accuracy (90 errors), it reads close to a clean human transcript and needs only light spot-checking.

Can I trust confidence scores to find errors?

Confidence scores are useful for flagging which words to check, not as proof of correctness. A model can assign high confidence to a hallucinated word, especially on noisy or silent segments. Whisper is particularly known for generating fluent-sounding text on non-speech audio. Also, a 0.9 confidence from one model is not the same as 0.9 from another: the scales are not calibrated across providers.

How do I measure my actual WER?

Take a 5-minute sample of your typical audio, transcribe it manually, run it through your AI tool, then compare the two with the Python jiwer library or the NIST sclite toolkit. Both compute WER from a reference and a hypothesis transcript. Five minutes of your actual audio gives you a more useful number than any vendor benchmark.

Sources

• Word Error Rate is Broken: How to Actually Evaluate Speech-to-Text in 2026 - AssemblyAI
• How Accurate is Speech-to-Text in 2026? - AssemblyAI
• What Is WER in Speech-to-Text? - Vatis Tech
• What Is Word Error Rate (WER): How It's Calculated - Gladia
• How Accurate Is Whisper in 2026? (WER Data by Language) - VexaScribe
• Best Open Source Speech-to-Text Model in 2026 (with Benchmarks) - Northflank
• jiwer - Python WER Calculator - PyPI
• AI vs Human Transcription Accuracy - Rev
• Calm-Whisper: Reduce Whisper Hallucination - Interspeech 2025