OhmsVoice

OhmsVoice turns speech into clean, pasteable text on your iPhone and iPad. Tap record, say what you mean, tap stop. The transcript is copied to the clipboard, ready to paste into an email, a ticket, or a document. HOW IT WORKS Transcription runs entirely on the device. The default engine is the open-source Parakeet model (NVIDIA, CC-BY-4.0), installed once with your consent as a 483 MB download that is pinned to a verified revision and checksummed. It supports 25 languages with automatic punctuation and capitalization, and you can delete it any time in Settings. Apple's speech recognizer is available as an optional engine, off by default, with on-device recognition required. Your audio never leaves the device with either engine, and there is no silent fallback: if the engine you chose isn't ready, OhmsVoice tells you and stops. After transcription you choose what happens to the words. Raw Transcript copies them exactly as spoken. Clean (On-Device) removes filler words, applies spoken punctuation commands, and fixes capitalization — no API key, no network. The other presets rewrite the transcript with a language model you supply. SIX PRESETS • Clean (On-Device) — fillers out, spoken commands and capitalization applied, fully local. • Clean & Concise — removes filler words and fixes grammar, keeping your voice. • Executive Email — turns spoken notes into a subject line and bullet points. • Meeting Minutes & Action Items — summary, decisions, owners, next steps. • Developer Bug / Jira Ticket — problem, repro steps, expected vs actual, proposed fix. • Raw Transcript — exactly what you said, with no network call. BRING YOUR OWN KEY OhmsVoice is not a subscription and does not resell anyone's model. Cloud presets remain blocked until you save your own API key and explicitly approve the named provider: OpenAI, Anthropic, OpenRouter, Google Gemini, xAI, or GitHub Models. Running one then sends transcript text — never audio — directly to that provider under your account and its privacy and retention terms. The provider receives your key, transcript, and ordinary request metadata. Keys live in the Keychain; deleting one revokes that provider's stored approval. Ollama stays on 127.0.0.1:11434. WHAT LEAVES THE DEVICE Exactly three things can use the network: the one-time, consented, checksum-verified speech model download from huggingface.co, whose host can receive ordinary HTTPS connection metadata; transcript text sent to a provider only after you save its key, explicitly approve it, and run a cloud preset; links you choose to open. No account, no telemetry, no analytics, no Ohm server. Speech model: NVIDIA Parakeet TDT 0.6B v3, © NVIDIA, CC-BY-4.0. Runtime: FluidAudio, Apache-2.0.