Speech Recognition Accuracy: Word Error Rates

Speech recognition is very accurate on clear English and much weaker almost everywhere else. A leading open model reaches 1.53 percent word error on LibriSpeech test-clean ¹. The same family of technology scores 100.7 percent on Malayalam read speech ², which means more wrong words than there were words spoken. Differences between the top English systems are small. Differences between languages, noise levels and speaking styles are large.

This page draws on 574 published figures from 32 sources. Each is a word error rate (WER): the share of reference words a system got wrong, so lower is better. For background on how the technology works, see our overview of speech recognition.

Top English systems are now separated by fractions of a point

The Artificial Analysis index (AA-WER) ranks commercial services on a blended English test. The leaders are tightly bunched. Fun-Realtime-ASR-preview scores 1.7 percent ³, MAI-Transcribe-2 scores 2.0 percent ⁴, and ElevenLabs Scribe v2 scores 2.2 percent ⁵. Further down, Whisper Large v3 hosted by Together AI sits at 4.5 percent ⁶, and Deepgram Nova-3 at 5.2 percent ⁷.

Showing 8 of 38 rows. See every row in the dataset.

The Open ASR Leaderboard, which averages several English test sets, tells a similar story. Otter's speech-v3.1 averages 3.8844444444444446 percent ¹³. IBM Granite Speech 4.1 2B averages 5.33 percent ¹⁴ and leads the open models in one snapshot. NVIDIA's Parakeet TDT 0.6B v3 averages 6.32 percent ¹⁵, and Whisper large-v3 averages 7.841428571428572 percent ¹⁶.

Not every entry is good. Google's Gemma 4 12B averages 20.86875 percent ¹⁷, which shows that a general-purpose model is not automatically a good recognizer. The older Chirp from Google scores 31.2 percent on AA-WER ¹⁸, against 4.3 percent for its successor Chirp 3 ¹⁹. Model generation matters more than brand.

Clean read speech is a solved problem, and it flatters every system

Audiobook-style speech gives the lowest rates in the table. Canary 1B Flash reaches 1.48 percent on LibriSpeech test-clean ²⁰. Whisper large-v2 scored 2.7 percent on the same set ²¹. Even the tiny Vosk model manages 9.85 percent ²², and its much larger sibling reaches 5.69 percent ²³.

Because the ceiling is so low, this benchmark no longer separates good systems from great ones. It also says little about how you will fare. Read, studio-quality sentences are the most favorable condition there is. One research system scored 3.7 percent in a close-talking studio ²⁴, and the authors themselves call that a best case.

Conversation costs several times more errors than reading

Spontaneous speech is harder. On the Switchboard telephone test, Whisper large-v2 scores 13.8 percent ²⁵, compared with 2.7 percent on read speech ²¹. In the Voxtral paper, GPT-4o mini Transcribe reaches 17.31 percent on Switchboard ²⁶, while Gemini 2.5 Flash reaches 9.57 percent ²⁷. Meetings are similar: Canary 1B Flash scores 13.11 percent on AMI ²⁸.

Humans are not error-free here either. Professional transcribers scored 5.9 percent on Switchboard ²⁹, and the best of the transcribers in IBM's study scored 5.1 percent ³⁰. Research systems matched them: IBM's recognizer reached 5.5 percent ³¹ and Microsoft's reached 5.8 percent ³². On the harder CallHome calls, humans scored 6.8 percent ³³ against 10.3 percent for IBM's machine ³⁴. Those systems were tuned to one benchmark, so read this as "machines can match careful humans on a narrow task," not "machines hear like people."

Showing 8 of 20 rows. See every row in the dataset.

Distance and noise are the biggest practical threats

Distance from the microphone hurts a lot. Whisper large-v2 scores 16.9 percent on AMI meetings recorded with headsets ⁴², but 36.4 percent with a single distant microphone ⁴³.

Added noise shows a steep curve. Canary-Qwen 2.5B scores 2.41 percent with mild white noise ⁴⁴ but 30.60 percent when the noise is louder than the speech ⁴⁵. Parakeet TDT v3 shows the same pattern with MUSAN noise: 7.12 percent at mild levels ⁴⁶ against 19.88 percent at the worst level ⁴⁷. On the CHiME-4 noisy-room test, Scribe scored 8.35 percent ⁴⁸ and Gemini 2.5 Flash scored 14.79 percent ⁴⁹.

Specialist vocabulary adds errors too. Earnings calls, full of company names and figures, give Canary 1B v2 11.79 percent ⁵⁰, even though the same model scores 2.28 percent on the more formal SPGI recordings ⁵¹.

Language matters more than the system

English and the large European languages are well served. Whisper large-v2 scores 3.0 percent in Spanish ⁵² and 5.6 percent in Russian ⁵³. Lower-resource languages fall off sharply: 33.0 percent for Welsh ⁵⁴, 75.6 percent for Lingala ⁵⁵, and 140.3 percent for Amharic ⁵⁶. A rate above 100 percent is possible because inserted words count as errors.

Newer systems narrow the gap for some languages. On FLEURS Hindi, Whisper large-v3 scores 28.87 percent ⁵⁷, while ElevenLabs Scribe scores 5.67 percent ⁵⁸. Fine-tuning for a language helps as well: Whisper fine-tuned on Kurmanji reached 10.5 percent ⁵⁹.

Dialects, children and unusual speech are harder still. Whisper large-v3 used as is on Swiss German scored 23.0 percent ⁶⁰, against 14.0 percent for a fine-tuned model ⁶¹. YouTube captions scored 18 percent on Mexican Spanish ⁶² and 24 percent on Argentinian Spanish ⁶³. Whispered speech gave Whisper 18.8 percent ⁶⁴, and a Whisper model tuned on child speech reached 80.4 percent for isiXhosa ⁶⁵.

Studies disagree, and the disagreement is informative

The same model can look good or bad depending on the test. Whisper large-v3 scores 15.44 percent on FLEURS Arabic ⁶⁶ but 50.58 percent on Common Voice Arabic ⁶⁷. Common Voice has varied speakers and microphones. FLEURS is read Wikipedia text in quiet rooms.

Rankings flip as well. In a study of university lectures, Google scored 20.1 percent ⁶⁸ while Whisper large-v2 scored 2.9 percent ⁶⁹. A preprint put ElevenLabs at 6.77 percent ⁷⁰, behind Whisper, while the Voxtral paper's runs put Scribe far ahead of Whisper on Hindi.

Even hosting changes results. The same Whisper Large v3 model scores 4.5 percent on one provider ⁶ and 10.1 percent on another ⁷¹, because settings and pipelines differ.

Our own quick test on a Mac sits in the same range as the published numbers. On English FLEURS recordings, Whisper large-v3-turbo scored 3.92 percent ⁷², Parakeet TDT v3 scored 4.48 percent ⁷³, and Apple's built-in SpeechTranscriber scored 6.73 percent ⁷⁴. That is a small test with no number normalization, so treat it as a sanity check, not a ranking.

What the data cannot tell you

• Different tests do not mix. Each paper and leaderboard uses its own text normalization, so only compare figures from the same table. A vendor's own figure on its own model card is not an independent check.
• AA-WER is an index. One of its datasets is private, and the page does not name its language, so the blended score cannot be traced to a single recording type.
• Benchmarks are mostly read speech. FLEURS, LibriSpeech and Common Voice are far from real dictation, which has your accent, your microphone and your vocabulary.
• Leaderboards move. Figures change as systems are added or rescored, and several rows here are snapshots.
• Some groups have one row. Children, whispered speech and several dialects rest on a single study each, so they show direction, not a rate.
• Word error rate is not usefulness. It counts every word equally, ignoring punctuation, names and whether the error changes the meaning.

What this means for choosing how to write

If you dictate in English in a quiet room, almost any current top system is accurate enough that the choice should rest on speed, privacy, price and editing workflow, not on a decimal of WER. Gaps among the leaders on the AA-WER table, such as 2.6 percent for Gemini 3.5 Transcribe ¹¹ against 3.0 percent for Universal-3.5 Pro ⁷⁵, will not be visible in daily use.

Several things do change your results. The microphone matters:: a headset or close microphone avoids the jump seen between headset and distant recordings. The language or accent matters: if you write in a lower-resource language or dialect, test candidates on your own voice, and prefer a system that names your language. Your vocabulary matters: specialist terms and names raise error rates, so favor tools with custom dictionaries.

Before committing, read a paragraph aloud to a few tools and count the errors yourself. For the wider picture of how recognition engines work and where they fail, see our overview of speech recognition.