voice

On-device speech transcription runs locally using Qwen3 models via the ominix-api, converting audio to text for offline accessibility without network dependency. It requires Apple Silicon hardware and a reachable ominix-api instance for model execution.

Preset-voice text-to-speech synthesis operates locally on Apple Silicon using local models with emotion and speed controls. It functions offline and automatically falls back to the macOS Say command if the ominix-api is unavailable.

Local model lifecycle management supports listing, downloading, loading, and unloading speech models through ominix-api commands. This controls resource allocation for on-device transcription and synthesis during local production workflows.

When the ominix-api is unavailable, speech synthesis uses a transparent fallback to the built-in macOS Say command. The system also attempts auto-discovery via environment variables, discovery files, and common ports before initiating this fallback.

Downloading local ASR and TTS models requires network access to retrieve the model files initially. Once downloaded and loaded, on-device speech transcription and synthesis operate completely offline without any network dependency.

Discovery for local speech processing auto-detects the ominix-api by checking environment configurations, reading a discovery file, and scanning common ports. It includes health checks to ensure the API is reachable before executing transcription or synthesis tasks.