Speech-to-Text on Intel Iris Xe

I built a self-hosted speech-to-text service around integrated GPUs I already owned. The four parts follow the work from hardware to the words that reach the editor:

1. Giving Idle GPUs a Job: Intel Iris Xe, Proxmox passthrough, OpenVINO, Kubernetes placement and the limits of a speed comparison.
2. Building and Releasing Container Images: custom Docker images, model exports, GitLab CI, digest promotion and what CVE scanning actually blocks.
3. Testing Whisper Accuracy with FLEURS and Common Voice: reference transcripts, word error rate, frozen release gates, quiet speech and the risks behind better averages.
4. When the Browser Stops Hearing: Open WebUI microphone capture, live preview, final transcription, recovery and the boundaries of browser tests.

I use Russian, English, Polish and German. The measurements belong to the recordings and releases described in each part; everyday dictation and controlled tests answer different questions. Published parts appear below as the series is released.