Whisper large-v3 is a multilingual speech-recognition and speech-translation model in the Whisper family, with 1550 M parameters. It changes the spectrogram input and adds a Cantonese language token compared with earlier large models. Its model card reports reduced errors across a wide variety of languages, while the project cautions that accuracy remains uneven across languages.
Large-v3 uses 128 Mel frequency bins instead of 80 and adds a new language token for Cantonese. Its card describes training on 1 million hours of weakly labeled audio and 4 million hours of pseudo-labeled audio collected using large-v2, for 2.0 epochs. This training recipe belongs to large-v3. The original Whisper series has a different training history, so use the recipe for the checkpoint you are evaluating.
The large-v3 card states: “The large-v3 model shows improved performance over a wide variety of languages, showing 10% to 20% reduction of errors compared to Whisper large-v2.” That describes the reported comparison, not a fixed improvement for every recording. The project warns that performance varies by language and that predictions can include text that was not spoken in the audio.
The large-v3 card lists 99 language codes in its metadata, and the tokenizer provides the language list. Coverage does not imply equal accuracy across those languages. The official GitHub README says Whisper code and model weights use the MIT License. The large-v3 model-card frontmatter instead lists Apache-2.0; the model-card metadata and repository therefore give conflicting license information for this checkpoint.
WhisperJot uses Whisper in Jot Local Pro, its local engine for long-form accuracy. On macOS it runs through WhisperKit with large-v3-turbo as the default model. On Windows and Linux it loads whisper.cpp-format GGML models through a .NET runtime. Auto mode picks between Jot Local Pro and the Parakeet-based Jot Local per utterance. Both local engines work offline after their one-time model download.
Explore WhisperJot →The model card reports 10% to 20% reduction of errors across a wide variety of languages. It does not promise that reduction for every language or recording. Evaluate your own audio, especially when comparing technical terms, accents or background noise.
The large-v3 card explicitly describes a new Cantonese language token. Its metadata lists 99 language codes. Cantonese is supported, and accuracy varies widely across the model's supported languages. Test representative recordings in the language you intend to transcribe, including your usual vocabulary.
No. WhisperJot's macOS Jot Local Pro defaults to large-v3-turbo through WhisperKit. Windows and Linux also offer turbo GGML choices. The full large-v3 model and the pruned turbo checkpoint are distinct models with different decoder designs. Check the selected checkpoint when comparing their behavior.
Primary documentation checked 2026-09-05. Specifications and runtime requirements can change.
Understand Whisper large-v3-turbo: fewer decoder layers, reference speed, translation limits, compressed variants and its role in WhisperJot Jot Local Pro.
Compare Whisper model sizes using the official parameters, VRAM and relative speed table. Learn what English-only variants mean and how to choose.
Compare Whisper and Parakeet by model size, language scope, architecture and license, with task-based guidance and WhisperJot's per-utterance Auto mode.
See also: Whisper glossary definition · System requirements · Pricing
