100% local · No cloud · No audio upload

Will it run on
your machine?

Celerando voice input — fully offline, four models to choose from, three hardware tiers from office laptop to GPU workstation. This page tells you exactly what you need.

4
Models
26
Well-recognised languages
3
Acceleration paths
0
Cloud calls
Self-Check

Which tier am I in?

Three clicks — and you'll know which model and which acceleration path match your machine. Real-time recommendation.

Input
Describe your machine
Values update the result instantly.
Result
Tier: Recommended
CPU
RAM — 8 GB
481632
Enough for speech recognition, Smart Modes and the translator
GPU
→ What you can use
Rough estimate based on hardware — which models your edition unlocks is listed under Editions. The app checks your graphics card itself and suggests suitable models.
01 — Whisper & Parakeet

Which model fits you?

Both run locally. Pick by language needs and available hardware: Whisper is part of every edition — free with Whisper Small or Base in your system language, from Pro with every model and 26 well-recognised languages. Parakeet is included from Business. In the app you pick a model from one list; W stands for Whisper, P for Parakeet.

Default
Whisper
Whisper.net 1.9.1 — ggml
Maximum language coverage. From Base to Turbo.
  • 🌍 26 languages in good quality
  • 📐 Three model sizes — Base, Small and Turbo
  • ⚡ Vulkan on any graphics card, CUDA optional
  • 🛡 Battle-tested open-source model
CPUVulkanCUDA
Recommended for maximum language coverage — in every edition
Optional
Parakeet
NVIDIA NeMo — ggml
Low latency. Technical terms as hotwords.
  • 🇪🇺 17 European languages, auto-detection
  • ⚡ Low latency, streaming-ready
  • 🎮 Vulkan on AMD, Intel and NVIDIA from 2 GB, otherwise CPU
  • 📖 Personal Dictionary works inside the recognition
CPUVulkan
Recommended for European languages & specialist vocabulary — fast on the CPU too
02 — Hardware

Three tiers. Clear choice.

From office laptop to GPU workstation. Find the card that best matches your machine.

Tier 1
Minimum requirements
Just enough to run
CPU
x86-64 with AVX2 (Intel Core from Haswell 2013, AMD Ryzen)
RAM
4 GB
GPU
— (CPU is enough)
Storage
100 MB + model (Whisper Base 141 MB)
OS
Windows 10 build 19041 (May 2020) or Windows 11
e.g. office laptop, occasional notes
Sweet Spot
Tier 2
Recommended
Comfortable everyday use
CPU
Multi-core x86-64 with AVX2
RAM
8 GB
GPU
Graphics card with 2 GB of graphics memory or more — AMD, Intel or NVIDIA (Vulkan)
Storage
1–3 GB for models
OS
Windows 10/11, current
e.g. daily dictation, mixed models
Tier 3
High-performance
Long dictation sessions, Smart Modes and translator on the GPU
CPU
Modern x86-64 with AVX2
RAM
16 GB
GPU
8 GB of graphics memory or more for Smart Modes and the translator, 12 GB for TranslateGemma 12B
Storage
up to 15 GB
OS
Windows 10/11 current + recent GPU driver
e.g. power user with GPU workstation
03 — Models

Four models.
Whisper & Parakeet.

Whisper
OpenAI · 3 models · 26 languages
CPUVulkanCUDA
Base
74 M ~141 MB
Smallest download, for low-end PCs
Recommended
Small
244 M ~466 MB
Good accuracy, runs without a graphics card
Turbo
809 M ~574 MB
Most accurate Whisper size, graphics card recommended
Parakeet
NVIDIA NeMo · 1 model · 17 languages
CPUVulkan
parakeet-tdt-0.6b-v3
~638 MB 1 file · q8_0

17 European languages in good quality with automatic language detection — DE, EN, ES, FR, IT, PT, NL, PL, RU, UK, SK, CS, RO, HR, BG, FI, HU. Low latency, streaming-ready. Words from the Personal Dictionary feed straight into the recognition as hotwords.

04 — Smart Modes · Optional

Local LLMs, when dictation should take shape.

Optionally a local language model cleans up, formats, or restructures the transcribed text — “phrase as an email”, “outline as bullet points”.

Note: With a graphics card with 8 GB of graphics memory or more, Gemma 3 4B runs on the card alongside speech recognition, usually in under a second on a fast card. Without one, you pick the same model in its more compact CPU version — then the CPU takes a few seconds per sentence. Voice input itself keeps working unchanged either way.
Gemma 3 4B Instruct
Q8_0
Size
~4.1 GB
VRAM
~5.3 GB
Backend
Vulkan
Gemma 3 4B Instruct
Q4_K_M
Size
~2.5 GB
VRAM
—
Backend
CPU
Privacy

Your voice never leaves this machine.

🔒
Fully local
All models run on your device. No external processing.
☁
No cloud
Audio and text data never leave your machine.
⬇
One-time download
Models are downloaded once straight from Hugging Face (only the optional CUDA runtime comes from the Menterium CDN) — offline-capable afterwards.
05 — FAQ

Answers in one sentence.

Download now.

14-day free trial. Models download once when voice input is activated.