Curiosity · AI Model

OpenAI Whisper v3 (large-v3)

Whisper large-v3 is OpenAI's open-weight automatic speech recognition (ASR) model. It covers 99 languages with strong accuracy on accented and noisy audio, and is released under an MIT-style licence, which is why it anchors most open-source transcription pipelines — from podcast tooling to meeting-notes apps.

Model specs

Vendor
OpenAI
Family
Whisper
Released
2023-11
Context window
30 tokens
Modalities
text, audio
Input price
n/a
Output price
n/a
Pricing as of
2026-04-20

Strengths

  • Open-weight and MIT-licensed — fully self-hostable
  • 99 languages with strong accented-speech robustness
  • Timestamped word-level output
  • Mature ecosystem — whisper.cpp, faster-whisper, Hugging Face

Limitations

  • 30-second chunking — long audio needs a segmenter
  • Hallucinates on silent or background-only audio — use VAD upstream
  • Speaker diarisation must be added separately
  • Real-time streaming needs faster-whisper or a distilled variant

Use cases

  • Podcast and video transcription
  • Meeting notes and action-item extraction
  • Subtitle generation across languages
  • Voice-note capture for productivity tools

Benchmarks

BenchmarkScoreAs of
Common Voice multilingual WER≈10.4%2023
LibriSpeech test-clean WER≈2.0%2023

Frequently asked questions

What is Whisper v3?

Whisper large-v3 is OpenAI's third-major-version open-weight speech-to-text model. It supports 99 languages and is released with MIT-licensed weights, so it can be self-hosted for free and fine-tuned.

How accurate is Whisper?

On clean English audio (LibriSpeech test-clean) Whisper large-v3 achieves roughly 2% word-error-rate. On multilingual Common Voice it averages around 10–11% WER, which is state-of-the-art among open-weight ASR models.

Can Whisper do real-time streaming?

Not out of the box — Whisper is trained on 30-second windows. For real-time, use faster-whisper, whisper.cpp with low-latency buffering, or a distilled variant like Distil-Whisper. Deepgram Nova-3 and AssemblyAI offer purpose-built streaming.

Is Whisper free to use?

Yes — the model weights are MIT-licensed and self-hostable for free. OpenAI also offers a hosted Whisper endpoint via its audio API for a per-minute fee.

Sources

  1. OpenAI — Whisper repo — accessed 2026-04-20
  2. Hugging Face — openai/whisper-large-v3 — accessed 2026-04-20