Local Speech Recognition Faster (25) · More Precise (100)

Set up local speech models

Run speech recognition locally on your computer. Choose between the default Faster model (25 languages) for quick turnaround and the optional Slower, More Precise model (100 languages), with local speaker separation where entitled.

Audio and video files processed through a local speech chip into speaker-separated transcripts
Model Selection & Device Routing
Faster (25) · More Precise (100) Local-First Design
01 · Model Choice

Faster (25) or Slower & More Precise (100)

Start with the default Faster model for 25 languages, or download the Slower, More Precise model covering 100 languages.

02 · Device Gate

GPU-First Automatic Check

Uses a supported GPU first, including on migrated older profiles without an explicit CPU override.

03 · Visible Error

No Silent CPU Switch

GPU runtime errors show exact setup fixes instead of silently falling back to CPU.

Stellar Recorder - Capture Setup & Local Mode Status
Faster (25) · More Precise (100) · GPU-First
Local Mode status indicator and audio preflight controls inside Stellar Recorder

Deterministic GPU-first execution

Automatic device selection checks for a supported GPU first and uses CPU only when no GPU is detected or you intentionally choose CPU. If GPU initialization fails, the app surfaces remediation steps instead of unpredictably slowing down on CPU.

Separation of speech and meeting summaries

Installing speech models enables transcription only. Meeting summaries stay optional and disabled until you enter your own OpenAI-compatible text endpoint and model.

On-device speech processing flow
Media files flowing into a local speech processor and speaker transcript rows

How it works step by step

  1. Open Model Setup in Stellar Recorder and check hardware prerequisites (including Microsoft Visual C++ x64 runtime).
  2. Keep Automatic device selection (GPU-first) or explicitly select CPU when running on CPU-only hardware.
  3. Select the Faster model (25 languages) or download the Slower, More Precise model (100 languages) as needed.
  4. Check the Privacy Meter to verify speech inference reads as local on your PC.

Limits and boundaries to know

  • Initial setup requires internet access to download model weights; no pre-bundled models ship in the installer.
  • Running the Slower, More Precise (100-language) model on CPU can take longer than the duration of the recording.
  • Speech model language settings are independent from the 100-language desktop UI selector.