Voice typing app for Linux

Hosted speech lets your Linux laptop dictate without running Whisper or Parakeet alongside your other apps.

Bulbul needs no local speech-model files or GPU for transcription. It sends audio to Groq, receives text, cleans it up through Groq, and types it into your focused app. Model inference does not depend on your laptop CPU/GPU, though platform requirements and network latency still matter.

The goal is everyday Linux voice typing, not asking you to build a model server first. Smaller CPU-friendly local models exist; choose them when offline speech matters more than avoiding local compute.

Use the package for your distribution, or the script that selects it.

This command downloads and runs the install script. Review scripts before executing them. The README documents .deb for Debian/Ubuntu/Mint, .rpm for Fedora/openSUSE, and AppImage elsewhere. You can instead download a signed release from the official GitHub repository.

After installation, add your Groq API key and grant microphone and input access where needed. This is no local-model setup, not zero setup.

Bulbul documents both X11 and Wayland support, using evdev for hotkeys and a uinput virtual keyboard for typing.

The .deb and .rpm installers configure input access. On Wayland, an AppImage can fall back to clipboard paste until uinput access is granted. Follow current platform docs rather than assuming every desktop environment behaves identically.

The README records verification on Mint/X11, Ubuntu/Wayland, and Fedora/Wayland. Basic typing support does not promise identical per-app integrations on every compositor.

Bulbul fits hosted dictation, Handy fits local offline speech, and OpenWhispr lets you choose.

Groq removes local speech inference as a bottleneck, not every possible bottleneck.

Groq documents a 216x real-time speed factor for Whisper Large v3 Turbo. This is a provider processing metric, not a Bulbul end-to-end benchmark or proof that every local model is slower. Upload time, connection quality, cleanup, service availability, and rate limits still affect when text appears.

The practical advantage is avoiding the speech-model download, memory, and compute workload. We do not claim identical latency on every device or zero RAM use.

Yes. Handy runs speech offline, and OpenWhispr supports local speech models.

Bulbul avoids the local speech-model memory and inference workload. It still runs an interface, records audio, and uses the network. No comparative battery or RAM savings have been measured here. Fully local speech is the better fit for offline use or a rule that audio must stay on-device; keep cloud cleanup disabled too.

Wispr Flow does not list Linux among its official platforms in our current check. Bulbul fits people who want Groq-backed dictation on Linux and Android without local speech models.

Bulbul supports configurable hotkeys; setup and provider limits still matter.

Use the hotkey shown in your installed Settings rather than relying on a shortcut that may have changed. Its README documents Shift + Alt + P for Polished mode. GNOME may need the AppIndicator extension for the tray icon; the README says dictation works without it.

Bulbul is free software with no subscription or Bulbul-imposed word cap. Groq has a rate-limited free tier; paid usage is billed to your own provider account. Speech transcription and text cleanup are separate requests, so speech pricing alone is not the whole cost.

Groq is not permitted to train or fine-tune models on your API inputs or outputs unless you explicitly grant permission or instruct it to do so.

Section 4.2 of the Groq Services Agreement says: "Groq is not permitted to use Inputs or Outputs for training or fine-tuning any AI Model Services or other models, unless explicitly granted permission or instructed by Customer."

Groq says inference content is not retained by default. Reliability or abuse investigations can temporarily log inputs and outputs for up to 30 days, or longer when legally required. Usage metadata is always retained and does not include customer inputs/outputs.

Groq documentation says all customers can enable Zero Data Retention (ZDR) in Data Controls to prevent content retention for reliability and abuse monitoring. The legal agreement refers to eligible customers. Check your account controls and current terms; Bulbul does not automatically enable ZDR, and this is not a promise that no data of any kind is ever stored.

This gives you a privacy-conscious hosted option without hosting the speech model on your laptop. It does not make cloud speech local: audio is sent to Groq and text is sent for cleanup. If your requirement is that audio never leave the device, a fully local workflow still has an advantage.

The general Groq Privacy Policy covers account, website, and usage information; it explicitly directs API Customer Data to the Services Agreement and DPA. These API-specific terms, not a blanket privacy slogan, support this comparison.

Product facts were checked against official sources on October 9, 2026. Provider performance is not an independent app comparison benchmark.