Documentation
Parleq's speech recognition runs entirely on-device. The optional cleanup pass runs either on-device — deterministic corrections with Lightweight, or full LLM rewriting with an on-device model (Gemma 4 E4B or the lighter Qwen3-4B) — or against a cloud LLM provider you configure with your own credentials. Pick the one that fits your situation:
Two on-device tiers: Lightweight (deterministic, built on Concord — no download, no RAM floor) and an on-device LLM via Apple MLX (full rewriting) — choose Gemma 4 E4B (~4 GB download, 12 GB+ RAM) or the lighter Qwen3-4B (~2.4 GB download, 8 GB+ RAM). No API key, no cloud account, no per-call cost — cleanup runs entirely on your Mac with no network boundary.
Setup guide →
Direct Google AI Studio API key. The fastest path to first-paste — about five minutes from sign-up to dictation. A free tier† is available without a credit card; check Google's pricing page for current quotas.
Setup guide →
Direct OpenAI API key (api.openai.com). GPT-4o and GPT-4.1 families for cleanup; o1 / o3 / o4-mini reasoning models for context-aware reference turns. Simple API key auth.
Setup guide →
Anthropic Claude or OpenAI GPT-OSS via your AWS account. Three auth modes: AWS SSO (recommended for org accounts), static IAM access keys, or scoped Bedrock API keys.
Setup guide →
Gemini models — plus Claude Haiku and Sonnet — on GCP, with IAM, audit logs, and data residency. Two auth modes: gcloud Application Default Credentials, or service-account JSON.
Setup guide →
GPT-4o and GPT-4.1 families on Microsoft's contract, plus o1 / o3 / o4-mini reasoning models. Two auth modes: resource API key, or Microsoft Entra ID via az login.
Setup guide →
† Third-party pricing, free tiers, and quotas are set by the providers — AWS, Google, Microsoft, Okta, and others change them without notice. Cost statements here were accurate when written; verify current rates on the provider's pricing page before you build.
Enterprise & teams
Back cloud cleanup with one corporate sign-in instead of per-user API keys, and deploy Parleq as a governed app via your existing MDM. Every employee signs in with the account they already have; nothing to distribute or rotate. Setting this up at home without an IT department? The DIY guide is for you.
One corporate OIDC sign-in backs cloud cleanup for every employee — every user signs in with the account they already have; no API keys to distribute or rotate. Federates into AWS Bedrock (AssumeRoleWithWebIdentity) and Google Cloud Vertex AI (Workforce Identity Federation), with IT-admin setup playbooks for each leg.
Read more →
Complete schema reference for every managed-configuration key — provider pinning, model allowlists, auth-mode lockdown, destination pins, the 9 enterprise OIDC federation pins, feature toggles, retention limits, and update-feed override. Pin the issuer, client ID, and cloud provider so users sign in but can't re-point the app.
Read more →
Deploying Parleq to a fleet, step by step — the MDM profile → upload → verify workflow, common patterns for Azure-only and Bedrock-only orgs, auth-mode and compliance-first lockdown, and the signed-app posture your security review needs.
Read more →
Wiring OIDC sign-in or Sign-in-with-Google at home, without an IT department. A path-picker decision table, a flagship walkthrough for Sign in with Google → Vertex, the gotchas that burn hours, and how to drive the whole thing with an AI assistant alongside you.
Read more →
Open Settings → Cleanup → Provider and pick None — paste raw ASR (skip cleanup). Parleq will paste the raw transcript exactly as the on-device speech model emitted it, and no audio or text ever leaves your Mac. Useful when transcript content must never touch a third party — e.g., for compliance-strict environments or when you want to manually edit the ASR output yourself. If you want cleanup but with no cloud account, use Settings → Cleanup → On-device instead — cleanup runs in-process with no network boundary.
All provider secrets are stored in the macOS Keychain — never in ~/.parleq/config.json. Audio is held in process memory and discarded as soon as it's transcribed; no audio file is ever written to disk.