Daytapes

Research · Transcription

Speech-to-text accuracy in 2026: what the published benchmarks say

Word error rate explained, the datasets behind the numbers, the Open ASR Leaderboard, vendor claims and the research on who transcription still gets wrong.

By Daytapes · Updated · Research-based

On clean English speech, the best speech-to-text systems in 2026 get roughly 5 words in 100 wrong. The top entries in the Open ASR Leaderboard paper average about 5.4% to 5.8% word error rate across eight English test sets. On the audio most people actually record, such as meetings, far-away mics, crosstalk and strong accents, error rates are much higher. Published research also shows they are not equal across speakers. A benchmark score tells you how a system did on one dataset, not how it will do on your kitchen-table conversation.

This is a research piece, not a test: we read the papers, leaderboards and vendor pages listed with this page (all accessed November 17, 2026). We haven't run our own accuracy tests, and we don't rank vendors. Where a number comes from a company about its own product, we say so.

What "word error rate" means

Word error rate (WER) is the standard score. You line up the machine's transcript against a careful human transcript of the same audio. Then you count three kinds of mistakes:

  • Substitutions: a wrong word ("site walk" heard as "side walk").
  • Deletions: a word that was said but is missing.
  • Insertions: a word that wasn't said but appears.

Add them up and divide by the number of words in the human transcript. Lower is better. NIST's long-running evaluations report exactly these counts (substitutions, deletions, insertions and total word errors), scored with NIST's SCLITE tool (NIST Hub-5 scoring readme). Because insertions count, WER can go above 100%. The Whisper paper reports values over 100 for its smallest models on some languages.

Two things WER hides:

  1. Formatting. Benchmarks usually strip punctuation and capital letters before scoring. The Open ASR Leaderboard uses Whisper's English text normalizer, for example (Open ASR Leaderboard paper). The Whisper authors point out that WER penalizes harmless formatting differences (Whisper paper). So a 5% WER doesn't mean 95% of your sentences come out readable as-is.
  2. Which words. Every error counts the same. Missing "um" and missing "not" both cost one word, but only one of them changes the meaning.

The datasets behind the numbers

A WER number only means something next to the dataset it came from. The common ones:

Dataset What it is Source
LibriSpeech About 1,000 hours of English read aloud from public-domain audiobooks; "test-clean" and harder "test-other" sets OpenSLR (Panayotov et al., 2015)
Common Voice Crowd-sourced read sentences in many languages; about 2,500 hours and 29 languages in the 2019 paper Ardila et al.
TED-LIUM 3 TED talks; 452 hours of training audio Hernandez et al.
AMI About 100 hours of recorded meetings, with close-talking and far-field microphones AMI corpus
Earnings-21 39 hours of company earnings calls, full of names and numbers Del Rio et al.
CHiME-6 Dinner-party conversations recorded with distant microphones Watanabe et al.
Switchboard About 260 hours of telephone conversations between strangers LDC

Read audiobooks are the easiest of these. Meetings and dinner parties are the hardest, and they're also the closest to what most people record.

What the independent leaderboard shows

The best-known independent comparison is Hugging Face's Open ASR Leaderboard. Its paper (v4, revised March 30, 2026) describes 86 systems across 12 datasets. Its English short-form ranking averages WER over eight test sets: AMI, Earnings22, GigaSpeech, LibriSpeech clean and other, SPGISpeech, TED-LIUM and VoxPopuli. It also reports speed as RTFx (seconds of audio transcribed per second of compute), because the most accurate system is not always fast or cheap enough to use.

The top of the paper's English table (Table 3, v4):

Rank System Average WER
1 Cohere Labs Transcribe 5.42%
2 Zoom Scribe v1 5.47%
3 IBM Granite Speech 4.0 1B 5.52%
4 NVIDIA Canary Qwen 2.5B 5.63%
5 IBM Granite Speech 3.3 8B 5.76%
6 Qwen3 ASR 1.7B 5.76%
7 ElevenLabs Scribe v2 5.83%

How to read it: the top seven are within half a percentage point of each other. That's about one word in 200. Differences that small can flip with a different test set or a different way of averaging. The live leaderboard also changes as new models are added, so check it yourself for today's order. What the table does show is that several systems from different companies, including open models you can run yourself, now perform about the same on these tests.

How far the numbers move with the audio

The Whisper paper (December 2022) is still one of the clearest public demonstrations of this. Its large-v2 model scored 2.7% WER on LibriSpeech test-clean. On harder audio it did much worse:

Test set (Whisper paper, Table 2) Whisper large-v2 WER
LibriSpeech test-clean (read audiobooks) 2.7%
Switchboard (phone calls) 13.8%
AMI, close-talking mics (meetings) 16.9%
AMI, single distant mic (meetings) 36.4%
CHiME-6 (dinner party, distant mics) 25.5%

Same model, very different results. A meeting recorded from one phone in the middle of the table is closer to the "distant mic" rows than the audiobook row. That's why mic placement matters more than any app choice. See How to record meetings in person.

Language matters just as much. On the FLEURS multilingual test, the paper reports large-v2 at 4.0% for Italian, 4.4% for English, 5.4% for Polish and 6.7% for Dutch, but 39.3% for Swahili. The authors tie the gap to how much training audio each language had. Whisper had no training data at all for 20 of FLEURS's 102 languages.

Are machines as good as people yet?

Closer than they were, depending on the test.

  • Phone calls, 2016. Microsoft researchers measured professional transcribers at 5.9% WER on Switchboard and 11.3% on the harder CallHome set. Their system scored 5.8% and 11.0%, and they called it human parity (Xiong et al.).
  • The pushback, 2017. IBM researchers got 5.5% and 10.3% on the same tests, and wrote that human performance "may be considerably better than what was earlier reported" (Saon et al.). How well humans do depends on how carefully they're asked to transcribe.
  • Podcast-style audio, 2022. The Whisper team compared their model with five transcription services on 25 recordings: one computer-assisted, four fully human. The computer-assisted service beat Whisper by 1.15 points of WER. The fully human services were "only a fraction of a percentage point better." The authors concluded Whisper was "very close to human-level accuracy" on that set.

The short version: on clear speech, good machine transcripts are now in the same range as a professional typist. On messy audio, a careful human still wins, and both get worse.

Accuracy isn't equal for everyone

Two peer-reviewed studies are worth knowing before you trust a transcript of someone else's words.

  • Accent and dialect. A 2020 study in PNAS tested five major commercial systems (from Amazon, Apple, Google, IBM and Microsoft) on interviews with 73 Black and 42 white speakers. The average WER was 0.35 for Black speakers and 0.19 for white speakers. In other words, roughly 35 errors per 100 words versus 19 (Koenecke et al., PNAS). The systems have improved since 2020. We found no follow-up that measured the same gap in 2026 systems, so we can't say how much of it remains.
  • Made-up text. A 2024 study of OpenAI's Whisper found that about 1% of transcriptions contained entire phrases or sentences that weren't in the audio. About 38% of those invented passages included explicitly harmful content. They happened more often for speakers with aphasia, especially around long pauses (Koenecke et al., "Careless Whisper"). A word error rate number doesn't capture this kind of failure.

What to do about it: before you quote someone, listen to the moment. Daytapes (our own product) keeps the audio with every transcript, which plays along line by line, for this reason: so you can hear what was actually said.

What the vendors claim (their own numbers)

Vendors publish their own benchmarks. These are useful, but the vendor chooses the test set, so read them as claims, not independent results.

  • Deepgram (Nova-3), February 12, 2025. Deepgram reports a median batch WER of 5.26% versus 10% for the next-best competitor it tested, and 6.84% for streaming. The test set was 2,703 files (81.69 hours) across domains such as meetings, phone calls, medical, finance and podcasts (Deepgram).
  • AssemblyAI, September 2026 data. AssemblyAI's benchmark page reports 4.99% English WER for Universal-3.5 Pro versus 5.03% for Cohere Transcribe. The same page shows ElevenLabs Scribe v2 ahead of AssemblyAI on accented English, verbatim transcripts and multilingual audio (AssemblyAI). It's to AssemblyAI's credit that its own page shows where it doesn't win.
  • OpenAI says its newer transcription models beat Whisper on the FLEURS benchmark, but we couldn't load the announcement page, so we don't quote its numbers here.

Notice that every vendor wins on its own chosen test. Deepgram's "next-best competitor" doubles its error rate, while AssemblyAI's page has the top systems within a few hundredths of a point. Both can be true on different audio.

What this means for you

  • Pick by your audio, not the leaderboard. If your recordings are meetings or family conversations, look at AMI and CHiME-type results, not LibriSpeech.
  • Fix the mic first. Going from a distant mic to a close one cut Whisper's AMI error rate by more than half (36.4% to 16.9%). No app upgrade comes close to that.
  • Test on your own recordings. Take a five-minute clip, transcribe it with two tools and count the mistakes in one paragraph. That beats any published number.
  • Keep the audio. Transcripts are drafts. Names, numbers and "not" are where errors hurt most. Keep the recording so you can check.

For how on-device Whisper compares with cloud services on cost and privacy, see Whisper vs. cloud transcription. For the meeting tools built on these models, see AI meeting note-takers compared.

Questions

What is a good word error rate?

It depends on the audio. The top systems on the Open ASR Leaderboard paper average about 5.4% to 5.8% WER across eight English test sets (paper version of March 30, 2026). Meetings, far-field audio and accents score much worse than clean read speech.

Is AI transcription as accurate as a human now?

On some tests, close. A 2016 Microsoft paper measured professional transcribers at 5.9% WER on Switchboard phone calls and its system at 5.8%. The Whisper paper (2022) found its best model very close to human transcription services on 25 podcast-style recordings. On hard audio, people still do better.

Why does my transcript look worse than the advertised accuracy?

Benchmarks are averages on specific datasets, often after removing punctuation and casing. Crosstalk, distance from the mic, noise, accents and names all raise the error rate, sometimes several times over.

Are vendor accuracy numbers trustworthy?

Treat them as the vendor's own claims. Each vendor picks the test set, and the winner changes with the dataset. Independent leaderboards and papers are better guides, and your own audio is the best test.

Do transcription apps make things up?

Sometimes. A 2024 study of Whisper found about 1% of transcriptions contained whole phrases or sentences that weren't in the audio, more often for speakers with long pauses. Check the audio before you quote.

Sources

  1. Open ASR Leaderboard paper (Srivastav et al., arXiv:2510.06961, v4 revised March 30, 2026) (accessed November 17, 2026)
  2. Hugging Face: Open ASR Leaderboard (live) (accessed November 17, 2026)
  3. Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision (Whisper paper, December 6, 2022) (accessed November 17, 2026)
  4. LDC: NIST Hub-5 English evaluation, scoring readme (May 11, 2000) (accessed November 17, 2026)
  5. OpenSLR: LibriSpeech ASR corpus (Panayotov et al., ICASSP 2015) (accessed November 17, 2026)
  6. Ardila et al., Common Voice: A Massively-Multilingual Speech Corpus (LREC 2020) (accessed November 17, 2026)
  7. Hernandez et al., TED-LIUM 3 (2018) (accessed November 17, 2026)
  8. University of Edinburgh: AMI Meeting Corpus (accessed November 17, 2026)
  9. Del Rio et al., Earnings-21 (Interspeech 2021) (accessed November 17, 2026)
  10. Watanabe et al., CHiME-6 Challenge (April 20, 2020) (accessed November 17, 2026)
  11. LDC: Switchboard-1 Release 2 (LDC97S62) (accessed November 17, 2026)
  12. Xiong et al., Achieving Human Parity in Conversational Speech Recognition (Microsoft, 2016) (accessed November 17, 2026)
  13. Saon et al., English Conversational Telephone Speech Recognition by Humans and Machines (IBM, March 6, 2017) (accessed November 17, 2026)
  14. Koenecke et al., Racial disparities in automated speech recognition (PNAS, March 23, 2020) (accessed November 17, 2026)
  15. Koenecke et al., Careless Whisper: Speech-to-Text Hallucination Harms (2024) (accessed November 17, 2026)
  16. Deepgram: Introducing Nova-3 (February 12, 2025; updated April 15, 2025) (accessed November 17, 2026)
  17. AssemblyAI: Benchmarks (September 2026 data) (accessed November 17, 2026)