Which model size should I pick for offline speech-to-text?

Start with the compact model — it covers everyday speech-to-text on an 8 GB machine without a GPU. If you have 16 GB of RAM and a dedicated GPU, the balanced model gives noticeably cleaner transcripts at five times the speed. The premium and ultra models are worth it only if you have 4 GB or more of free VRAM and process long recordings.

Does offline mode really work without any internet connection?

Yes. After the one-time model download, WhisPaste runs whisper.cpp entirely on your machine. Aeroplane mode, an air-gapped workstation, or a flaky hotspot do not affect transcription quality at all — the only thing that needs the network is the optional auto-update check, which you can disable.

How accurate is offline speech-to-text compared to cloud APIs?

For everyday voice input the offline balanced and premium models match what most cloud APIs produce. Specialist domains — heavy accents, very noisy environments, rare technical vocabulary — sometimes benefit from a larger cloud model. WhisPaste lets you switch per session, so you can keep offline as the default and reach for cloud only when a recording really needs it.

Offline speech-to-text — when the audio must never leave your machine.

When you want speech-to-text on your laptop but the recording cannot leave the device — because of a client contract, a regulated industry, or simple personal preference — WhisPaste runs the entire transcription pipeline locally with whisper.cpp, so the audio is processed where it was captured and nothing is uploaded.

Choosing a mode: local by default, cloud on demand

WhisPaste defaults to local mode. The model runs on your CPU or GPU, the audio stays on the machine, and the transcript is delivered to the cursor without a network round-trip. If you need maximum speed for a long recording, you can switch a single setting to a cloud provider — but that is an explicit choice per session, not a hidden fallback.

What offline mode can — and what it cannot

Offline Whisper handles 99 languages, mixed-language input, accents, background noise, and natural speech with pauses and restarts. What it cannot do: real-time streaming to a server, speaker diarisation across many channels, or model sizes that exceed your available RAM. The compact model fits in roughly 300 MB of GPU VRAM, the premium turbo model in about 2.6 GB.

RAM and CPU profile

The minimum is 8 GB of RAM and a 64-bit CPU on Windows 10, macOS 10.15 or Ubuntu Linux — enough for the compact model on CPU. With 16 GB of RAM and a GPU with 2–4 GB of VRAM, the balanced model transcribes a one-minute clip in seconds. Apple Silicon shares memory between CPU and GPU, so an 8 GB Mac already handles the balanced model.

Which model size should I pick for offline speech-to-text?: Start with the compact model — it covers everyday speech-to-text on an 8 GB machine without a GPU. If you have 16 GB of RAM and a dedicated GPU, the balanced model gives noticeably cleaner transcripts at five times the speed. The premium and ultra models are worth it only if you have 4 GB or more of free VRAM and process long recordings.
Does offline mode really work without any internet connection?: Yes. After the one-time model download, WhisPaste runs whisper.cpp entirely on your machine. Aeroplane mode, an air-gapped workstation, or a flaky hotspot do not affect transcription quality at all — the only thing that needs the network is the optional auto-update check, which you can disable.
How accurate is offline speech-to-text compared to cloud APIs?: For everyday voice input the offline balanced and premium models match what most cloud APIs produce. Specialist domains — heavy accents, very noisy environments, rare technical vocabulary — sometimes benefit from a larger cloud model. WhisPaste lets you switch per session, so you can keep offline as the default and reach for cloud only when a recording really needs it.