faster-whisper Hotkey

a minimalist push-to-talk style transcription tool built upon cutting-edge ASR models.

Hold the hotkey, Speak, Release ==> And baamm in your text field!

In the terminal, in a text editor, or even in the text chat of your online video game, anywhere!

Features

• User-Friendly Interface: Allows users to set the input device, transcription model, compute type, device, and language directly through the menu.
• Fast: parakeet-ultra gives instant transcription, granite-speech-5.0-470m-turboctc-nc comes pretty close.
• LLM correction: Optionally try to repair broken transcriptions text via any OpenAI-compatible API endpoint.

Current models

To help with choosing your model, you can look at AA-AgentTalk score.

CPU speed for a 40 s clip (best compute type per model, see Tips):

On GPU every model is super fast (all < 3.5 s for the same clip), so that's not particularly useful to benchmark that. But anyway, detailed evaluation results can be found in test_audio_data/transcription_results.md.

• (NEW) moondream/parakeet-ultra:

  • 25 languages
  • Transcription only
  • Post-trained version of nvidia/parakeet-tdt-0.6b-v3: same 0.6B architecture, but:
    • Lower word error rate on every benchmark
    • Runs with Moondream's Photon engine --> 3x faster!
  • CPU/GPU (I let GPU as option, but there is no point using it... Anyway!)

• Qwen/Qwen3-ASR-1.7B-hf:

  • 30 languages
  • Transcription only
  • Automatic language recognition
  • CPU/GPU

• ibm-granite/granite-speech-4.1-2b:

  • 6 source languages (en, de, es, fr, ja, pt)
  • Transcription and translation (bidirectional to/from English, plus English→Italian and English→Mandarin)
  • No automatic language recognition
  • Punctuation and capitalization
  • CPU/GPU

• ibm-granite/granite-speech-4.1-2b-nar:

  • 5 languages (en, de, es, fr, pt)
  • Transcription only
  • No automatic language recognition
  • No capitalization for proper nouns
  • CPU/GPU

• ibm-granite/granite-speech-5.0-470m-turboctc-nc:

  • English only
  • Transcription only
  • No capitalization or punctuation in output --> LLM-correction can be helpful
  • CPU/GPU

• CohereLabs/cohere-transcribe-03-2026:

  • 14 languages
  • Transcription only
  • Officially, no automatic language recognition, but it still works pretty great
  • CPU/GPU
  • Quite smart, deals well with hesitation and stutters

• nvidia/canary-1b-v2:

  • 25 languages
  • Transcription and translation (bidirectional to/from English)
  • No automatic language recognition
  • CPU/GPU

• mistralai/Voxtral-Mini-3B-2507:

  • 8 languages
  • Transcription only
  • Automatic language recognition
  • Smart (it even guesses when to put some quotes, etc.) and seems less error-prone for non English native speakers
  • GPU only

• Systran/faster-whisper:

  • Many languages
  • Transcription only
  • CPU/GPU

Installation

see https://docs.astral.sh/uv/ for more information on uv. uv is fast :)

Requires Python 3.11.14

From PyPi

• As a pip package:

  uv venv --python 3.11.14
  source .venv/bin/activate
  uv pip install faster-whisper-hotkey
  

• or as an tool, so that you can run faster-whisper-hotkey from any venv:

  uv tool install faster-whisper-hotkey --python=3.11.14
  

From source

1. Clone the repository:

   git clone https://github.com/blakkd/faster-whisper-hotkey
   cd faster-whisper-hotkey
   

2. Install the package and dependencies:

• as a pip package:

  uv venv --python 3.11.14
  source .venv/bin/activate
  uv pip install .
  

• or as an uv tool:

  uv tool install .
  

Usage

1. Run the tool using one of these methods:

   • If installed from PyPi or from source as a package/tool:

     faster-whisper-hotkey
     

   • Or using uv run (handles the venv and dependencies automatically):

     cd faster-whisper-hotkey
     uv run faster-whisper-hotkey
     

2. Go through the menu steps.

3. Once the model is loaded, focus on any text field.

4. Then, simply press the hotkey (PAUSE, F4, F8 or INSERT) while you speak, release it when you're done, and see the magic happening!

When the script is running, you can forget it, the model will remain loaded, and it's ready to transcribe at any time.

Configuration File

The script automatically saves your settings to ~/.config/faster_whisper_hotkey/transcriber_settings.json.

Limitations

• granite-nar: punctuation comes from the model, but capitalization is a post-processing pass (sentence starts + English "I"), so proper nouns like names stay lowercase.

• granite-turboctc: no punctuation or capitalization in output (by design of the non-autoregressive architecture).

  Use the autoregressive granite variant if you need full capitalization, or the LLM correction option to fix proper nouns.

• Using window type detection to send appropriate key strokes, we unfortunately can't see "sub windows". So for example, the VSCodium/VSCode terminal isn't supported for now. No clue if we can workaround this.

• Windows supported is not planned. That said, you can use eutychius's branch which seems working fine. See this comment for instructions.

Tips

• If you you pick a multilingual faster-whisper model, and select en as source while speaking another language it will be translated to English, provided you speak for at least few seconds.
• If you pick parakeet-ultra, you can even use multiple languages during your recording!
• For models whose weights are natively bfloat16 (the precision menu marks the native one), you can still pick float32: it upcasts the weights, which costs extra memory, but is far faster on CPUs without good native bf16 support. If your transcriptions are slow, try float32.

Acknowledgements

Many thanks to:

• the developers of faster-whisper for providing such an efficient transcription inference engine
• NVIDIA for their blazing fast parakeet and canary models
• Moondream for parakeet-ultra and their Photon local inference engine
• Mistral for their impressively accurate model Voxtral-Mini-3B model
• Cohere for their cohere-transcribe-03-2026 model
• IBM for their granite-speech models
• Qwen for their Qwen3-ASR-1.7B model
• and to all the contributors of the libraries I used

Also thanks to wgabrys88 and MohamedRashadthat for their huggingface spaces that have been helpful!

And to finish, a special mention to @siddhpant for their useful broo tool, who gave me a mic <3