Jarvis

Jarvis routes system-wide push-to-talk speech on macOS through local Whisper, NVIDIA Parakeet via Sherpa-ONNX, or cloud transcription APIs without recording telemetry or requiring user accounts. The utility was built as a zero-subscription alternative to commercial dictation apps, combining speech recognition with an Ollama bridge or remote LLM endpoints for contextual cleanup. Users trigger it inside any active text field by holding the function key, converting speech into formatted code comments, shell commands, or pull request drafts.

Why Jarvis instead of Wispr Flow

Wispr Flow pairs fast voice capture with cloud-based context inference, but it routes microphone audio through central servers and demands a monthly subscription. For developers working in proprietary repositories, terminal sessions, or production consoles, sending audio to hosted endpoints creates compliance and privacy risks. Jarvis replicates that push-to-talk experience without the cloud dependency. It runs speech recognition on Apple Silicon hardware using local Whisper models or NVIDIA Parakeet, inserting transcribed text directly into any focused window while keeping audio strictly in local memory.

The difference lies in developer workflow control. Wispr Flow locks formatting behind fixed cloud prompts that cannot adapt to technical syntax. Jarvis exposes prompt engineering directly, letting engineers define templates for conventional commits, markdown task lists, or incident logs. It also links to local Ollama runtimes, allowing models like Llama 3 to remove filler words and reformat technical prose entirely on-device without token costs or data leakage.

What it does well

• On-device speech recognition via Whisper and NVIDIA Parakeet keeps proprietary code discussions, secrets, and architecture reviews contained to local machines.
• Local LLM post-processing through Ollama removes filler words and structures dictation without external cloud dependencies or API charges.
• Customizable prompt templates allow engineers to define transformation rules for pull request summaries, documentation notes, and issue updates.
• Universal push-to-talk injection operates across any active macOS window, pasting text directly into code editors, IDEs, and terminal sessions.
• Dual-mode architecture supports low-latency cloud inference through Deepgram, Google Gemini, or AWS Bedrock when local hardware compute is constrained.