Frequently Asked Questions

VoxRelay captures audio playing through a selected Windows playback source, recognizes speech locally with a Whisper-compatible model, and displays floating subtitles over other applications. English is transcribed directly, while supported non-English speech is translated into English.

No. Auto mode uses NVIDIA CUDA when a supported GPU and driver are available and otherwise selects CPU processing. CUDA uses the larger Large-v3/float16 model for higher accuracy and lower latency. CPU uses the smaller Small/int8 model and is expected to be slower and somewhat less accurate.

Right-click the VoxRelay notification-area icon and open Processing Device. Auto is recommended, NVIDIA CUDA requires a supported GPU, and CPU runs without CUDA. VoxRelay saves the choice and reloads the appropriate model without restarting.

No. Captured audio, language detection, transcription, and translation run locally on the Windows computer. Audio is processed in memory and is not sent to Max Power IT, Hugging Face, or another speech-processing service.

VoxRelay downloads the selected CUDA or CPU model from a revision-pinned copy on maxpowerit.com and verifies each file against its expected size and SHA-256 hash. If that server is unavailable, VoxRelay automatically uses the original Hugging Face repository as a backup. These are file-download requests; captured audio and subtitle text are not uploaded. After the model is cached, normal transcription and translation remain local.

No. The default Max Power IT download and public Hugging Face backup can be used without an account. An unauthenticated-request warning means a backup download is subject to lower Hub rate limits; it does not mean VoxRelay is using an online transcription service.

VoxRelay captures a Windows playback or loopback source through WASAPI. This is normally the audio sent to speakers, headphones, an HDMI output, or compatible capture hardware. It does not normally capture a microphone.

Right-click the VoxRelay notification-area icon, open Select Source, and choose the desired device. Choose Refresh Sources after connecting or enabling new audio hardware.

Only when compatible hardware and its Windows driver expose the ARC/eARC signal as a selectable playback or loopback source. A TV's ARC/eARC connector alone is not a Windows audio source. HDCP-protected content and output-only interfaces may not provide capturable audio.

VoxRelay analyzes confident speech and rechecks the language every 30 seconds. English is transcribed; supported non-English speech is translated into English. A possible change must be detected confidently more than once before VoxRelay switches the remembered language.

Translated captions begin with a source-language label and may include the detector's confidence, such as [Translated from French 94%]. Direct English transcription does not receive a translated-from label.

The 30-second interval determines when analysis is due, but a change also requires clear speech and repeated confident detections. Silence, short clips, music, overlapping speakers, and uncertain audio can delay the switch.

Speech models can infer familiar phrases from ambiguous audio such as silence, music, applause, noise, or television transitions. VoxRelay filters common false phrases, non-speech audio, low-confidence segments, and repeated overlap, but no speech model can eliminate every false caption. Use the latest app version and the cleanest available playback source.

Voice-activity and confidence filters intentionally reject probable non-speech and unreliable model output. Very quiet speech, distant dialogue mixed under music, crosstalk, and heavily compressed audio may fall below those thresholds.

Drag inside the transparent subtitle area to move it. Move the pointer near an invisible edge or corner until a resize pointer appears, then drag. The overlay has a minimum size to keep subtitle text readable.

VoxRelay is designed as floating text plus a notification-area controller. It does not create a conventional visible window or taskbar button. Right-click the VoxRelay tray icon to show or hide subtitles, select a source, choose Auto/CUDA/CPU processing, refresh sources, or quit.

Confirm speech is playing through the selected source, refresh the source list, check Windows Volume Mixer routing, and verify that the audio is not muted. Very quiet input and non-speech content are intentionally ignored.

The selected speech model may require a substantial download and storage. The CUDA Large-v3 model is several gigabytes; the CPU Small model is smaller. VoxRelay displays the download percentage and transferred size in a temporary model-setup window. Download time depends on the internet connection, security scanning, and disk speed. Later launches reuse the cached model.

The standard Store launch does not save captured audio or an SRT history. Audio is discarded after processing. Other distributions can support an explicitly configured SRT output path; that setting writes subtitle text only to the location chosen by the user.

No. VoxRelay does not require an account and does not include advertising, analytics, tracking, or telemetry SDK. Public support-board posts and emails are information users voluntarily provide outside the desktop app.

Microsoft Store builds update through the Store. To uninstall, quit VoxRelay from its tray menu, then use Settings > Apps > Installed apps. Cached model files may remain in the user's VoxRelay model directory or Hugging Face cache and can be removed separately when no local application needs them.