Voice & AI features

PsyOps Desktop runs every AI workload — voice command parsing, dictation transcription, and clinical text refinement — entirely on the practitioner’s Mac. No cloud APIs, no patient data leaves the device. The trade-off is that we ship the AI runtimes inside the app and need a recent Apple Silicon Mac to run them at acceptable speed.

What the AI features do

• Voice commands. Press the mic button and say something like “new appointment with Maria tomorrow at 3pm”. PsyOps parses the intent locally and opens the matching form prefilled — no confirmations to dismiss, no record created on a misheard command.
• Dictation in any clinical field. Speak directly into a session note, client field, or any rich-text editor in PsyOps. Dictation is transcribed locally with whisper.cpp and lightly polished by a small on-device language model so you get clean Markdown instead of raw speech-to-text artefacts.
• Clinical summaries. Long session notes get an LLM-generated title and summary so the client timeline shows useful context at a glance, without you having to write a heading every time.
• Speaker diarisation on imported session recordings — separates therapist from client across the transcript, so you can scan the conversation by speaker after the session.

What runs locally

PsyOps Desktop bundles or installs three components on first launch:

• Ollama serving qwen3.5:2b (~1.9 GB) for natural-language understanding and text refinement.
• whisper.cpp with the Whisper Medium model (~1.5 GB) for speech-to-text.
• NVIDIA Sortformer for speaker diarisation (~250 MB, bundled).

Total first-launch download is around 3.5 GB. Subsequent launches are fast — everything is cached on disk and reused.

Which Macs can run it

Voice & AI features require:

• Apple Silicon (M1, M2, M3, or M4 Mac). Intel Macs are not supported.
• macOS 14 (Sonoma) or later.
• ~6 GB free disk for the model files.
• An internet connection for the first-run download. After that, AI runs fully offline.

On Macs that don’t meet these requirements, the rest of PsyOps Desktop still works as expected — appointments, client notes, calendar sync, recording import — just without the voice button and the in-editor dictation. You can do everything with the keyboard and mouse.

Why on-device, not cloud

Therapy notes are some of the most sensitive data a person produces. PsyOps’ first principle is that none of it ever leaves the practitioner’s machine: no cloud transcription, no “our servers analyse your audio”, no opt-out toggles to misconfigure. The cost is hardware requirements; the benefit is that the ethical guarantee is structural, not policy.