Hardware & Performance Automatic GPU-First · No Silent CPU

GPU or CPU for local transcription

Leave device selection set to Automatic on fresh installations: Stellar Recorder detects supported GPUs first, avoids hidden CPU fallbacks on GPU failure, and uses CPU only when no GPU exists or when you deliberately select CPU.

Device Routing & Error Transparency Schema
Hardware & Performance Automatic GPU-First · No Silent CPU
01 · Default Policy

Automatic Selection

Fresh installs and legacy profiles without a device lock both request Automatic mode.

02 · GPU First

Parallel Matrix Execution

Supported GPU hardware handles speech inference batches significantly faster than CPU.

03 · Strict Gate

Never Hide GPU Errors

Incompatible drivers or runtime crashes halt visibly with setup remediation guidance.

04 · Explicit CPU

Predictable Fallback

CPU executes only when no GPU is detected or when chosen intentionally in settings.

Stellar Recorder - Local Mode Execution & Review Workspace
GPU-First · Explicit CPU Choice
Local Mode transcript review showing completed GPU transcription and executive summary

Why GPUs accelerate speech transcription

Neural speech models execute massive streams of parallel matrix math. Desktop GPUs process many steps simultaneously, whereas CPUs execute fewer concurrent threads - causing CPU runs (especially under the Slower, More Precise 100-language model) to take noticeably longer.

What happens when a GPU is unsupported or fails

Older GPUs or installations missing required runtime components (such as Visual C++ x64 redistributables) cannot run accelerated kernels. Rather than silently falling back to CPU and turning a 2-minute job into an unpredictable wait, Stellar Recorder surfaces explicit remediation steps and lets you decide whether to fix drivers or switch explicitly to CPU.

When CPU - or a private LAN speech server - makes sense

Laptops without discrete or supported GPUs can select CPU deliberately alongside the Faster (25-language) model. Alternatively, teams with a dedicated workstation GPU on their network can opt into a remote speech server you configure (keeping in mind recording audio uploads to that configured endpoint).

Hardware checklist

  • Ensure Microsoft Visual C++ x64 runtime is installed before model activation.
  • Prefer the Faster (25-language) model over the Slower, More Precise (100-language) model if your hardware only supports CPU execution.
  • Inspect visible driver or model errors directly rather than forcing retries.