How Accurate Is Whisper in 2026? WER Benchmarks & Sources

Whisper Large-v3 achieves 2.7% WER on the LibriSpeech test-clean benchmark (clean read English audiobook audio) and 5.2% WER on the LibriSpeech test-other benchmark (varied audio conditions). On real-world English audio — meetings, podcasts, phone calls, interviews — WER ranges from 8% to 12%. WER is calculated as (Substitutions + Deletions + Insertions) / total reference words, the standard ASR accuracy metric defined by NIST. Lower is better.

On English audio, Whisper Large-v3 (~8-12% real-world WER) is roughly tied with Google Chirp 2 / USM-2 (~8-11%) and Deepgram Nova-3 (~7-10%), and slightly behind AssemblyAI Universal-2 (~7-10%) and Speechmatics Ursa (~7-10%). All five sit within 2-3 percentage points on the Hugging Face Open ASR Leaderboard composite score. Whisper's advantages: 99+ languages out of the box, MIT license (free to self-host), large open ecosystem. Commercial API advantages: custom vocabulary boosting, native streaming, built-in speaker diarization. For raw transcription accuracy alone on English, Whisper is competitive with the best.

Whisper Large-v3 (1.5B parameters) is the current most accurate, achieving 2.7% WER on LibriSpeech clean. Large-v2 is slightly less accurate (~3–5%). The Tiny, Base, Small, and Medium models trade accuracy for speed — Tiny achieves only 10–15% WER but runs 32× real-time on a GPU.

Yes. Spanish is a Tier 1 language for Whisper with 3–6% WER on clean audio — near-parity with English. French, Italian, Portuguese, German, and Dutch perform similarly. Lower-resource languages (Vietnamese, Thai, Welsh) have significantly higher WER.

Whisper accuracy degrades with: noisy audio (+5–15% WER), strong accents (+5–10%), phone audio vs studio (+5–10%), multiple overlapping speakers (+5–10%), technical/domain vocabulary (no custom vocab support), and long silences (Whisper occasionally hallucinates text during silence).

Whisper transcribes all speech but does not natively identify speakers (no diarization). For speaker labels, you need to combine Whisper with tools like pyannote-audio or use WhisperX, which adds forced alignment and diarization. Commercial tools built on Whisper (VexaScribe, TurboScribe) include diarization.

Yes. Whisper is released under the MIT license, which permits unrestricted commercial use. You can self-host, modify, and include it in products you sell. OpenAI also offers a paid Whisper API ($0.006/min) for those who don't want to self-host.

Yes. Once the model is downloaded, Whisper runs entirely locally with no internet connection required. This makes it suitable for privacy-sensitive applications, offline environments, and air-gapped systems. Model sizes range from 39MB (Tiny) to 3GB (Large-v3).

Word Error Rate (WER) is the standard ASR accuracy metric defined by NIST: WER = (Substitutions + Deletions + Insertions) / Number of words in reference. Substitutions are wrong words, Deletions are missed words, Insertions are extra words. WER of 5% means 95% of words are correct relative to a ground-truth transcript. For Chinese, Japanese, and Korean, Character Error Rate (CER) is used instead because word boundaries are ambiguous. Whisper's 2.7% WER on LibriSpeech is measured by OpenAI in the original Whisper paper (cdn.openai.com/papers/whisper.pdf), and reproduced on the Hugging Face Open ASR Leaderboard.

The most reliable Whisper benchmarks come from: (1) The Hugging Face Open ASR Leaderboard (huggingface.co/spaces/hf-audio/open_asr_leaderboard) — composite score across 8 datasets including LibriSpeech, TED-LIUM, AMI, GigaSpeech, Earnings-22, CallHome, CommonVoice, and SPGISpeech. (2) The original Whisper paper (Radford et al., OpenAI 2022) — covers LibriSpeech, FLEURS, and zero-shot evaluation on 14 datasets. (3) Papers With Code WER leaderboards (paperswithcode.com/task/speech-recognition). Vendor blog claims should be cross-checked against these independent sources.

Large-v3 (1.5B parameters, ~2.7% WER on LibriSpeech clean) gives maximum accuracy at 1× real-time on a modern GPU. Large-v3-turbo (809M parameters, released October 2024) is a distilled version that runs at 8× real-time with only a 0.3-0.7 percentage-point WER increase — roughly 3-4% WER on LibriSpeech clean. Use Large-v3 when accuracy is paramount and processing time is tolerable. Use Large-v3-turbo for production workloads where latency or throughput matters. Both are MIT-licensed and free to self-host.

Yes — three documented failure modes. (1) Hallucination on silence: during long pauses or silent segments, Whisper sometimes invents plausible-sounding text. Mitigation: use Voice Activity Detection (VAD) preprocessing to skip silence. (2) Repeated tokens: Whisper occasionally gets stuck repeating the same phrase on certain audio patterns. Less frequent in v3 than v2; faster-whisper's `repetition_penalty` parameter helps. (3) Language detection errors: Whisper can confuse linguistically similar languages — Ukrainian misclassified as Russian, Catalan as Spanish, Welsh as English. Mitigation: specify the language explicitly via the `language` parameter rather than relying on auto-detection. These are open known issues tracked in the openai/whisper GitHub repository.

Depends entirely on audio length and cleanliness. In Novascribe's July 2026 benchmark of 904 audio files, GPT-4o Transcribe beat Whisper-1 on LibriSpeech clean read speech (3.1% vs 4.7% WER) but collapsed on longer, harder audio: Earnings21 financial calls 43.8% (Whisper-1: 9.7%), TED-LIUM long prepared speech 27.1% (Whisper-1: 5.0%), AMI meetings 40.9% (Whisper-1: 26.4%). GPT-4o Mini follows the same pattern at half the price. For voice memos and short clean clips, GPT-4o models are more accurate and cheaper. For anything longer than a few minutes — podcasts, meetings, earnings calls, lectures, interviews — Whisper-1 outperforms GPT-4o by 4-5×.

Whisper-1, decisively. Novascribe's July 2026 benchmark measured Whisper-1 at 9.7% WER on the Earnings21 financial-call benchmark while GPT-4o Transcribe scored 43.8% and GPT-4o Mini scored 44.2% on the same audio. The pattern repeats on TED-LIUM (Whisper-1: 5.0% vs GPT-4o: 27.1%) and AMI meetings (Whisper-1: 26.4% vs GPT-4o: 40.9%). Whisper-1 is OpenAI's oldest transcription model but was trained on 680,000 hours of long-form real-world audio, while the GPT-4o transcription models appear optimized for short conversational turns. If your audio is longer than ~2 minutes, choose Whisper-1 or a competitor like AssemblyAI Universal-3.5 or Deepgram Nova-3, not the GPT-4o transcription line.