SpeechKit modes and data paths

SpeechKit uses three strict modes. Choose the smallest mode that matches the task so the expected processing and output stay clear.

​

Dictation

Dictation turns speech into text. It does not perform LLM rewriting, execute utilities, or generate spoken replies. Use it for notes, forms, messages, and other direct transcription tasks.

​

Assist

Assist turns speech or text into one bounded result. That result may come from a deterministic replacement, a utility, or an LLM, and may optionally be read aloud. Use it when you want transformation or a single answer rather than a live conversation.

​

Voice Agent

Voice Agent is a realtime audio conversation. It is intended for interactive brainstorming, support, and follow-up questions where turn detection and spoken output matter.

​

Hands-Free and customization

Hands-Free controls activation, microphone capture, automatic end-of-speech, and optional speaker output over the three modes. Words teach SpeechKit terms to recognize; Replacements apply deterministic text, command, snippet, synonym, or template transformations. Meeting sits beside the three modes rather than inside them. It records a call on its own pipeline and shares only the speech-to-text transcriber with Dictation, so call audio never reaches the dictation overlay, history, or clipboard shortcuts.

​

Local and external providers

Provider choice determines the data path: SpeechKit ships adapters for local whisper.cpp and Piper, a self-hosted whisper-server, Ollama and other OpenAI-compatible endpoints, and the external providers OpenAI, Groq, Deepgram, AssemblyAI, Hugging Face, OpenRouter, Azure Speech, and Microsoft Foundry. Google speech-to-text, Google text-to-speech, and Gemini Live were removed; configurations that still name them fail closed instead of silently selecting a different provider. SpeechKit does not make an external provider local. Review provider configuration and privacy terms before sending sensitive audio.

​

Network Scope

Network Scope is the boundary you set once and the backend enforces everywhere — provider assembly, model downloads, updates, telemetry, server connections, and integrations. Choosing a stricter scope suspends incompatible settings instead of deleting them, and Settings explains what is disabled and why.

​

Retention Scope

Retention Scope is the second, independent axis: it answers what outlives the work that produced it. An install with no network access still writes transcripts to disk. Set both scopes deliberately.