Technical explainer · Verified 2026-09-05

Run Whisper locally
hardware and setup routes

Local Whisper transcription is speech recognition performed on your own computer using downloaded Whisper model weights. You can run a C/C++ runtime, integrate a Python runtime, or use a desktop app that bundles a local engine. Hardware requirements depend on the model and runtime; once the selected model is downloaded, inference can work without an internet connection.

Official Whisper model size table
SizeParametersEnglish-only modelMultilingual modelRequired VRAMRelative speed
tiny39 Mtiny.entiny~1 GB~10x
base74 Mbase.enbase~1 GB~7x
small244 Msmall.ensmall~2 GB~4x
medium769 Mmedium.enmedium~5 GB~2x
large1550 MN/Alarge~10 GB1x
turbo809 MN/Aturbo~6 GB~8x

Values reproduced from the official Whisper README. Relative speeds measure English speech on an A100; language, speaking speed and hardware affect actual results. The large row does not identify the checkpoint behind the alias.

Budget memory and disk separately

Use the official VRAM table as a model comparison for the reference implementation: tiny and base are listed at ~1 GB, large at ~10 GB, and turbo at ~6 GB. Other runtimes have their own memory profiles. Disk is separate again; whisper.cpp lists tiny at 75 MiB and large-v3 at 2.9 GiB. Check both before downloading a candidate model.

Choose a route that fits your application

After checking your hardware, choose any of the independent setup routes below. whisper.cpp provides a C/C++ path with GGML models. faster-whisper provides a Python path using CTranslate2. A desktop app packages the model selection and runtime into an application workflow. Follow the linked runtime documentation for installation commands and choose the route that fits how you intend to use the transcript.

Prepare and verify an offline session

Download the model while connected, then test transcription with the local engine selected and the connection disabled. WhisperKit explicitly downloads models during first use, and local Whisper inference does not need a network connection after the weights are available. Choose representative recordings for the test. The download makes the model available offline. Review the transcript to check how well it handles your language, vocabulary and speech.

Set up local transcription

Check your hardware, choose one of the routes, then verify offline transcription.

  1. Check hardware and language. Compare your available memory with the table, then select an English-only or multilingual model. Check the chosen runtime's disk and memory requirements before downloading weights.
  2. Route A: use the whisper.cpp CLI. Follow the whisper.cpp repository's build and CLI instructions, download a compatible GGML model from its models documentation, and run transcription on a local audio file. Choose a quantized model if you want to evaluate lower disk and memory use.
  3. Route B: integrate faster-whisper in Python. Follow the faster-whisper setup instructions with Python 3.9 or greater. Choose a model and a CPU or GPU compute type. For the documented GPU path, check cuBLAS for CUDA 12 and cuDNN 9 for CUDA 12, then transcribe a local recording.
  4. Route C: select a desktop engine in WhisperJot. Select Jot Local Pro. On Mac, the default turbo download is approximately 1.5 GB, with compressed turbo at approximately 626 MB or English-only small.en at approximately 217 MB. Windows and Linux offer turbo at approximately 1.62 GB or quantized turbo at approximately 574 MB.
  5. Test your chosen route offline. After the model download finishes, disconnect and transcribe representative speech using the local engine. On Mac, WhisperJot also offers Transcribe Audio File from the menu bar; its result lands in History and can be exported as text.

How WhisperJot uses this

WhisperJot's macOS Jot Local Pro is WhisperKit-based. Its default large-v3-turbo download is approximately 1.5 GB; selectable alternatives are compressed large-v3-turbo at approximately 626 MB and small.en at approximately 217 MB. Non-English transcription requires large-v3-turbo because the .en option is English-only. WhisperKit's full catalog offers additional models beyond this product selection. Jot Local Pro runs on-device by default and works offline after the model download.

Explore WhisperJot →

Frequently asked questions

Does Whisper need internet?

A connection is needed to download model weights once. Inference then runs locally without a network connection. WhisperKit downloads its models during first use, so complete that setup before going offline. In an app with multiple engines, select the local engine for offline transcription.

Do I need a GPU to run Whisper locally?

No. Local runtimes include CPU execution, and faster-whisper documents CPU benchmarks. GPU VRAM figures from the official Whisper table are not universal minimum system RAM figures for every runtime. Choose a model and runtime together, then test on your hardware.

Can I transcribe an audio file in WhisperJot?

On Mac, use the menu bar's Transcribe Audio File action with an audio type the operating system can open. The result lands in History and can be exported as text. These file-import steps apply specifically to the Mac app. Windows and Linux support for this workflow is unconfirmed.

Sources

Primary documentation checked 2026-09-05. Specifications and runtime requirements can change.

Related explainers

Whisper model sizes

Compare Whisper model sizes using the official parameters, VRAM and relative speed table. Learn what English-only variants mean and how to choose.

whisper.cpp

Explore whisper.cpp, its GGML model disk sizes, integer quantization and Apple acceleration options, plus how WhisperJot uses GGML on Windows and Linux.

faster-whisper

Read faster-whisper's CTranslate2 speed and memory benchmarks with their model and precision settings, Python requirements, and when this runtime fits.

See also: Whisper glossary definition · System requirements · Pricing