Local Models

Documentation Index

Fetch the complete documentation index at: https://hyperwhisper.com/docs/llms.txt

Use this file to discover all available pages before exploring further.

Local models let HyperWhisper transcribe fully offline — once a model is downloaded, audio never leaves your device and no internet connection is required to transcribe. HyperWhisper ships two families of speech-to-text models:

• Whisper — OpenAI’s general-purpose multilingual models, multiple sizes from Tiny to Large v3
• Parakeet — NVIDIA’s models, fast and accurate for English (v2) and 25 European languages (v3)

For local post-processing (cleanup, formatting), variants of Google’s Gemma 4 are also available on Windows x64.

​

Downloading models

Open Model Management from the app, click Download on any entry, and watch the circular progress indicator. You can cancel mid-stream with the × button if needed. Downloaded models are stored on disk and stay there until you remove them.

​

Does local mean offline?

Yes. After a model finishes downloading, HyperWhisper transcribes entirely on your machine with no network calls. This is the strongest privacy guarantee — your audio is never uploaded anywhere. See Data Privacy for details.

​

GPU vs CPU

Local transcription engines use your GPU for acceleration when available, and fall back to CPU automatically if you don’t have a dedicated GPU or there isn’t enough VRAM. Local Gemma post-processing is available on Windows x64 only: it uses CUDA on NVIDIA GPUs and falls back to CPU on x64 AMD, Intel, non-GPU, or GPU-constrained systems. The VRAM values listed below are the recommended GPU memory for full GPU acceleration. If your GPU has less than the recommended VRAM, the model still runs — it just uses CPU (or partial GPU offload), which is slower but works on any modern machine.

​

Whisper models

​

Parakeet models

NVIDIA Parakeet models are typically faster than equivalent-size Whisper models and very accurate for the languages they support. Parakeet v3 covers: English, German, Spanish, French, Italian, Portuguese, Dutch, Polish, Romanian, Swedish, Danish, Finnish, Norwegian, Czech, Slovak, Hungarian, Croatian, Slovenian, Bulgarian, Ukrainian, Greek, Lithuanian, Latvian, Estonian, Catalan, Basque.

​

Local Gemma post-processing models

Local Gemma models clean up and format transcript text after transcription. They run offline after download, but they are separate from speech-to-text models and are currently available only in the Windows x64 app.

​

Which model should I pick?

​

When to switch to cloud instead

Local models are great for privacy, offline use, and zero per-minute cost. But cloud providers are noticeably more accurate than even the largest local Whisper model — especially for accented English, technical vocabulary, and noisy audio. If you find local accuracy isn’t cutting it, try a HyperWhisper Cloud tier — there’s no markup over the underlying provider and a Pro license includes 5,000 credits to start. See Providers for tiers and pricing.

​

Boosting local accuracy

You can squeeze more accuracy out of any local model by:

• Adding custom vocabulary for product names, jargon, and colleagues’ names
• Reducing background noise — see Best Practices
• Speaking at a natural pace (not too fast, not too slow)

​

Removing models

To free up disk space, click the trash icon next to any downloaded model in Model Management. The file is removed immediately and can be re-downloaded any time.