Added in v0.17.0
Faster Whisper on Linux
Faster Whisper is an optional local speech engine in Vocalinux. It runs Whisper models through CTranslate2 with int8 quantization on your CPU and types what you say into the focused app. It is for people who want Whisper models and the full language catalog on a machine without a GPU, using a different runtime from the default whisper.cpp.
CPU first
int8 on the CPU by default. No PyTorch needed for the CPU path.
On-device
After the one-time model download, recognition runs locally. Audio stays on your machine.
Pinned models
Models load only from checksum-verified local files, never from an unpinned upstream revision.
What is Faster Whisper?
faster-whisper is a reimplementation of OpenAI's Whisper on top of CTranslate2, an inference engine for transformer models. It runs converted Whisper models and supports reduced precision such as int8, which lowers memory use on the CPU.
Vocalinux added Faster Whisper as an optional engine in v0.17.0, together with Parakeet. whisper.cpp remains the default. Vocalinux stores Faster Whisper models under ~/.local/share/vocalinux/models/ and checks each file against a pinned checksum before loading it.
Install and select Faster Whisper
curl -fsSL https://raw.githubusercontent.com/VocaHQ/vocalinux/main/install.sh -o /tmp/vl.shbash /tmp/vl.sh --auto --engine=faster_whisper- 1
Install the Faster Whisper engine
Faster Whisper is an optional extra. Run the installer and choose Faster-Whisper in the engine menu, or pass --engine=faster_whisper. If the faster-whisper package fails to install, the installer falls back to whisper.cpp and says so.
- 2
Open Settings → Speech Model
Open Settings and go to Speech Model. Set your Language, then expand Advanced to see the Engine and Model size pickers.
- 3
Pick Faster Whisper and a size
Set Engine to Faster Whisper and choose a Model size. Settings suggests one for your hardware. Picking only selects the model; nothing downloads yet.
- 4
Confirm the download
If the model is not on disk, click Download and confirm. Every file is checked against a pinned checksum before the engine loads it.
- 5
Dictate
Hold your dictation shortcut (Right Alt by default on new installs), speak, and release. Text is typed into the focused app.
To start once with Faster Whisper from a terminal, run vocalinux --engine faster_whisper.
Model sizes
The Model size picker offers these sizes. The .en variants are English only; if your language is English, Settings suggests the .en variant of the recommended size. There is no large-v3.en.
| Size | Model IDs | Parameters | Description |
|---|---|---|---|
| Tiny | tiny, tiny.en | 39M | Fastest, lowest accuracy |
| Base | base, base.en | 74M | Fast, good for basic use |
| Small | small, small.en | 244M | Balanced speed and accuracy |
| Medium | medium, medium.en | 769M | High accuracy, slower |
| Large v3 | large-v3 | 1550M | Highest accuracy, slower |
Which size Settings suggests
- CPU only, under 16 GB RAM: Tiny
- CPU only, 16 GB RAM or more: Base
- CUDA GPU detected, under 8 GB RAM: Base
- CUDA GPU detected, 8 GB RAM or more: Small
The suggestion never changes your model or starts a download. If dictation feels slow, go one size down; if it makes too many mistakes and speed is fine, go one size up. The Whisper model guide explains the size trade-offs in more depth.
Compute options and GPUs
Vocalinux chooses the device and precision for you; there is no compute-type setting. On the CPU it loads the model with int8. If PyTorch is installed and reports a CUDA GPU, it loads the model on CUDA with float16 instead.
The installer's Faster Whisper option is the CPU CTranslate2 path and does not install the CUDA toolkit. If you want GPU dictation on AMD, Intel or NVIDIA, whisper.cpp with Vulkan is the supported route; see GPU acceleration.
Languages
Faster Whisper uses the same language catalog as the other Whisper engines, including Auto-detect. It also supports Follow keyboard layout, which switches the dictation language with your active layout, and custom dictionary terms that bias recognition toward names and jargon. See languages by engine, multilingual dictation and custom dictionary.
Package availability
| Install method | Faster Whisper | Notes |
|---|---|---|
| install.sh / source install | Yes | Menu option or --engine=faster_whisper |
| AUR (vocalinux) | Yes | Listed as an optional dependency: python-faster-whisper |
| AppImage (and AUR vocalinux-bin) | No | Ships whisper.cpp only |
| Flatpak | No | Ships whisper.cpp only |
| Snap | No | Ships whisper.cpp only |
| .deb / .rpm | No | Thin packages; whisper.cpp on CPU |
| Nix flake | No | faster_whisper extra is not packaged |
Faster Whisper vs whisper.cpp, Whisper and VOSK
All four are local engines. Choose by hardware and by how you installed Vocalinux. For the full table, see the engine comparison or Parakeet for the other CPU engine added in v0.17.0.
| Engine | Pick it when |
|---|---|
| whisper.cpp | The default. Ships in every package, uses a Vulkan GPU on AMD, Intel and NVIDIA, and offers quantized and English-only variants. Start here. |
| Faster Whisper | You installed with install.sh, from source or from the AUR and want a Whisper model on a CPU-only machine, running through CTranslate2 instead of whisper.cpp. |
| Whisper (OpenAI) | You already work with PyTorch and NVIDIA CUDA and accept a much larger install. |
| VOSK | Older or low-RAM machines. Small models are about 40 MB, with one model per language. |
Troubleshooting
The installer finished, but whisper.cpp is selected
If faster-whisper fails to install or import, the installer falls back to whisper.cpp. Read the installer output, then re-run it with --engine=faster_whisper.
Another language is not recognized
The .en models are English only. Pick the size without .en, or set Language to Auto-detect with a multilingual model.
Dictation feels slow
Go one size down. On CPU-only machines, Tiny or Base is the usual starting point. Larger sizes take longer per sentence on a CPU.
Anything else
Run vocalinux --debug and check ~/.local/share/vocalinux/vocalinux.log, then see the troubleshooting guide.
Faster Whisper FAQ
Is Faster Whisper the same model as Whisper?
It runs the same Whisper model family converted to the CTranslate2 format. Vocalinux downloads those conversions from pinned Hugging Face revisions. The sizes and English-only variants match the Whisper names you already know.
Does Faster Whisper use my GPU?
On a typical install it runs on the CPU with int8. Vocalinux switches to CUDA with float16 only when PyTorch is installed and reports a CUDA GPU. The installer's Faster Whisper option does not install CUDA. For Vulkan GPU acceleration on AMD, Intel or NVIDIA, use whisper.cpp.
Can I change the compute type?
There is no setting for it. Vocalinux picks int8 on the CPU and float16 on CUDA automatically.
Does audio leave my machine?
No. Faster Whisper is a local engine. After the model download, recognition runs on your machine. Only the Remote API engine and VocaGateway send audio to an endpoint you configure.
Why is Faster Whisper not in my Engine list?
Settings lists only engines whose libraries are installed. The AppImage, Flatpak, Snap, .deb and .rpm packages do not include faster-whisper. Re-run the installer with --engine=faster_whisper, or on Arch install python-faster-whisper.
Try Faster Whisper
Install Vocalinux with the Faster Whisper engine, or keep whisper.cpp and switch later from Settings → Speech Model.