Speech to text accuracy: word error rate explained, and how to test it

Speech to text accuracy is usually measured as word error rate (WER): the number of words the system got wrong, divided by the number of words actually spoken. A WER of 5% means about one word in twenty was substituted, missed or added. Lower is better. But WER counts every word equally, so a transcript can score well and still get the one word that matters wrong: a name, an amount, or a "not". The only accuracy figure that tells you about your own dictation is one you measure on your own voice, which takes about ten minutes.

We have not run a controlled benchmark of any product. The published numbers below describe their publishers' test datasets, not your voice, accent, microphone or vocabulary.

How is word error rate calculated?

The Open ASR Leaderboard gives the standard formula in its About tab:

WER = (S + I + D) / N

• S, substitutions: a word replaced by a different word.
• I, insertions: a word that was never spoken.
• D, deletions: a spoken word that is missing.
• N: the number of words in the correct (reference) transcript.

A worked example.

Suppose you said: "Do not send the 14 March invoice to Jordan Chen." That is 10 words.

Transcript A has one deletion and one substitution. WER = (1 + 0 + 1) / 10 = 0.2, or 20%.

Now a second transcript: "Do not send 14 March invoice to Jordan Chen, um." It drops "the" (one deletion) and adds "um" (one insertion). WER is also 20%.

Same score, very different damage. Transcript A reverses the instruction and misspells the person. Transcript B is harmless. That is the main limit of WER: it measures how much is wrong, not how much it matters.

Two more details. Because insertions count, WER can go above 100% if a system adds lots of words that were never said, which can happen when a model "hallucinates" over silence. And for languages written without spaces between words, publishers often use character error rate (CER) instead; OpenAI's Whisper README shows CER in italics for those languages.

Why can a low WER still get the important word wrong?

• Every word weighs the same. "The" and "not" each count as one error.
• Names are rare words. A surname the model has seldom seen may come out as a common word that sounds alike. One wrong name in a 200-word note barely moves WER.
• Numbers have many correct forms. "Fourteen", "14" and "14th" may all be acceptable, so benchmarks standardise them before scoring. That means a benchmark often does not test whether numbers come out in the form you want.
• Punctuation and capitals are not scored. The leaderboard says it normalises text before computing WER, which "removes punctuation and casing," converts numbers (for example "zero" to "0") and removes filler words. A transcript with no commas and no capital letters can score perfectly.

OpenAI's Whisper paper explains why benchmarks normalise: "systems that output transcripts that would be judged as correct by humans can still have a large WER due to minor formatting differences." It reports that its own normaliser cut WER by up to 50 percent on some datasets. Normalisation makes comparisons fairer, but it also hides exactly the formatting a dictation user sees.

How do leaderboards test accuracy?

The Open ASR Leaderboard runs each model over public datasets and averages WER "over all audios in the benchmark (in the order of thousands of audios)." Its English table uses sets such as:

It reports speed as RTFx, seconds of audio processed per second of compute, and states that WER and RTFx come from the same run. It also has a private-data tab, which it says exists "to address risks of overfitting to public datasets."

The spread between datasets is the lesson. In the leaderboard's results file for its 25 September 2026 update, Whisper large-v3 shows a WER of 1.56 on LibriSpeech clean and 13.63 on AMI meetings. The same model is many times worse on spontaneous meeting speech than on clean reading. Your dictation sits somewhere on that range, and only a test tells you where.

What about Whisper's accuracy?

OpenAI's README says "Whisper's performance varies widely depending on the language" and shows WER and CER by language on the Common Voice 15 and Fleurs datasets. The large-v3 model card adds that the models "exhibit disparate performance on different accents and dialects," and may produce "texts that are not actually spoken in the audio input (i.e. hallucination)." Whisper large-v3 vs turbo lists the leaderboard figures for both large models.

A 10-minute method to measure accuracy on your own voice

You need a text editor, the microphone you normally use, and the tool or tools you want to compare.

1. Write a fictional script of about 60 words. Include two names, a date, a time, an amount, a percentage and a negation. For example: "Please tell Amara Lindqvist that the 22 July review moves to 3:30 pm. The budget is 48,500 dollars, up 7.5 percent. Do not copy Ruben Okonkwo on the draft. The Q4 target stays at 12 units." Do not use real client or personal details.
2. Count the words. That is N. Count numbers as one word each.
3. Dictate it three times: once in a quiet room, once with normal background noise, and once at your normal speaking speed if you slowed down for the first two.
4. Mark each transcript against the script. Count substitutions, insertions and deletions. Ignore punctuation and capitals for this count.
5. Work out WER for each run: (S + I + D) / N.
6. Separately count critical errors: wrong names, wrong numbers, lost negations. Write these in their own column.
7. Note what the WER missed: punctuation, capitals, number format.

Repeat with each tool or model on the same script. If you dictate in a second language, do a second script in that language. The tool with fewer critical errors usually saves more time than the one with the lower WER.

Where does SuprFlow fit?

SuprFlow 1.0.0 uses a WhisperKit (Whisper) model, with Small and Medium options in multilingual and English-only variants. We publish no accuracy figure for it, because we have not benchmarked it. Use the method above to compare its models, and see Compare SuprFlow speech models for a fair way to do that.

Questions people ask

What is a good word error rate for speech to text?

There is no universal threshold. The same model can score under 2% on clean audiobook audio and over 13% on meeting recordings, as Whisper large-v3 does in the leaderboard's results file. Compare tools on the same audio, and weigh critical errors more than total WER.

Is WER the same as accuracy?

People often quote accuracy as 100% minus WER, but that is only a rough shorthand. Because insertions count, WER can exceed 100%, and it ignores punctuation and how serious each error is.

Why do my numbers come out as words, or the other way round?

Formatting of numbers is a separate step from recognition, and benchmarks normalise it away. Test number-heavy sentences specifically. Testing amounts and percentages in dictation has a script for that.

How do I make speech to text more accurate?

Use a better microphone position, speak at a steady pace, add names to the tool's dictionary if it has one, and choose a model suited to your language. Improve Mac dictation accuracy covers the steps.

Related guides

• Improve Mac dictation accuracy
• Run a fair Mac dictation comparison
• Best speech to text models
• Parakeet vs Whisper
• Compare SuprFlow speech models
• How speech recognition works

Sources

• Hugging Face — Open ASR Leaderboard, including its About tab
• English results dataset
• Radford et al. — Robust Speech Recognition via Large-Scale Weak Supervision
• OpenAI — whisper GitHub repository and README
• OpenAI — whisper-large-v3 model card on Hugging Face