Whisper Large V3 Turbo: Offline Dictation in 99 Languages on Mac
Most cloud-based voice dictation tools run Whisper on their servers. They charge for the API call, send your audio to their infrastructure, and return the transcript. Resonant runs the same model — Whisper Large V3 Turbo — on your Mac, offline, with no account and no audio leaving your machine.
The most trusted speech model in the world
OpenAI released Whisper in 2022 and it changed the offline speech recognition landscape. Before Whisper, accurate local transcription required expensive, narrow models trained on specific domains. Whisper demonstrated that a single general-purpose model, trained at scale on diverse audio, could achieve accuracy that rivaled cloud services.
Whisper Large V3 Turbo is the latest distilled version. It has 809 million parameters but only 4 decoder layers instead of the full 32 in Whisper Large V3 — which makes it substantially faster while retaining nearly all the accuracy. It supports 99 languages and defaults to English when no language is specified.
The weight of global adoption matters here. Whisper is the model that powers dictation tools, transcription services, accessibility software, and developer integrations worldwide. Its accuracy characteristics are well-documented across thousands of real-world use cases.
99 languages, fully offline
Whisper Turbo covers languages that no other model in Resonant handles: Arabic, Hindi, Vietnamese, Thai, Hebrew, Turkish, Indonesian, Ukrainian, Farsi, Bengali, Tamil, and dozens more. If you dictate in a language not served by a dedicated model — Parakeet for European, SenseVoice for East Asian, Zipformer for Japanese and Korean — Whisper Turbo is your model.
You can specify your language in Resonant's settings so Whisper doesn't have to detect it on every dictation. For bilingual workflows, leaving it on auto-detect works well for most major language pairs.
When to use Whisper Turbo
Your primary language isn't English or a major European language. Whisper Turbo's multilingual coverage is its defining feature. For languages where no dedicated model exists, it's the right choice.
You want the most battle-tested model available. Whisper has been evaluated more extensively than any other offline ASR model. If you want accuracy you can trust across diverse content, Whisper delivers.
You dictate in multiple languages. Whisper handles language-switching reasonably well and covers a broader set than Parakeet.
How to enable it in Resonant
Go to Settings → Transcription and select “Whisper Large V3 Turbo”. The model is about 1 GB and downloads in the background while Resonant keeps running. Once downloaded, set your preferred language in the language selector below the model list if you don't need auto-detection.
Your audio is processed entirely on-device. INT8 quantization keeps the model compact enough to run efficiently on any Apple Silicon Mac.
Download Resonant and run Whisper locally.