title: Product Requirements Document – Speakr
version: 2025-07-20
status: Draft
authors: David Jessup
Speakr is a privacy-first dictation hot-key utility for macOS (Windows/Linux later). In a single
keystroke, users can record speech, transcribe entirely on-device , and have the text typed
directly into any active input field. Speakr aims to be the fastest way for developers, writers,
and power-users to turn fleeting thoughts into code or prose without breaking flow, and
without sending audio to the cloud .
Switching to dedicated dictation apps breaks focus and incurs network latency.
Many corporate or offline environments forbid cloud speech services for privacy reasons.
OS-level dictation is unreliable for code, lacks custom hot-keys, and has high latency on older
hardware.
Opportunity : A lightweight, keyboard-driven tool that works anywhere text can be typed, requires
no network, and respects user privacy.
<= 3 s end-to-end latency for 5-second recordings on Apple Silicon (M-series).100% offline – no external network calls.Global hot-key works in background apps.
Support customisable models & hot-keys via UI.
Ship notarised universal macOS binary < 20 MB (excluding model).
Provide a clean upgrade path to Windows & Linux.
Real-time streaming (v1 may paste only after stop).
Mobile platforms.
Full grammar / punctuation correction.
Server-side sync or accounts.
Persona Needs / Pain-points
Dev Dana Insert comments/code quickly without losing keyboard context. Writer Will Draft snippets into any text editor without toggling apps. Privacy Peter Dictate confidential material offline, no data leaves device. Accessibility Ava Replace or augment typing due to RSI, keep workflow keyboard-first.
MoSCoW method: Must, Should, Could, Won’t (for now)
Priority Description
Must “As a user, I press <Opt> + ~ and my spoken words (≤30 s) are typed into the active field within ~3 s.” Must “As a user, the app asks for mic + Accessibility permissions on first run and explains why.” Must “As a user, I can change the hot-key in settings and be warned of conflicts.” Should “As a user, I can pick a smaller/faster model if my machine is slow.” Should “As a user, a subtle overlay shows ‘Recording… / Transcribing…’ states.” Could “As an advanced user, I can turn on auto-punctuation.” Could “As an advanced user, I can add bespoke words to the dictionary.” Won’t (v1) Live transcript shown word-by-word while speaking.
FR Description
FR-1 Global hot-key registers at app start and triggers record/transcribe/inject flow. FR-2 Audio capture uses 16 kHz mono via cpal, max configurable duration (default 10 s). FR-3 Transcription runs through Whisper (GGUF) via whisper-rs; language default EN. FR-4 Transcript is injected via synthetic keystrokes (enigo) into current focus. FR-5 If injection fails (secure field), fallback to clipboard-paste with user warning. FR-6 UI (tray or window) exposes: hot-key picker, model selector, auto-launch toggle. FR-7 App emits status events for UI overlay and logs (Recording, Transcribing, Error). FR-8 Settings persist locally (JSON in AppData, no cloud). FR-9 App auto-updates via GitHub Releases (optional in v1).
Category Requirement Metric / Acceptance
Latency End-to-end ≤ 3 s (M1, 5 s audio, small model) 95th percentile measured in telemetry log (local). Footprint Binary ≤ 20 MB; RAM ≤ 400 MB including model. du -sh and Activity Monitor/smoke tests.Reliability No crashes in 1-hour monkey test (500 invocations). CI integration test + manual QA. Security No outbound network sockets except auto-update domain (opt-out). Static analysis + firewall test. Compatibility macOS 13+. Intel macs may see doubled latency but functional. QA on Intel MBP (2020) & M1. Accessibility Follows macOS VoiceOver / high-contrast guidelines. Apple Accessibility Inspector score ≥ 85.
Metric Target
Time-to-text (P95) ≤ 3 s. Activation success rate ≥ 99% (hot-key triggers & types). Crash-free sessions > 99.5%. Daily active users (DAU) post-launch target: 1 k. % of transcripts requiring manual fix < 15% (optional feedback prompt).
Milestone Scope
M0 – Prototype spike Hot-key → record → transcribe → paste (CLI) M1 – MVP macOS app Tauri shell, settings window, notarised DMG M2 – Public beta Auto-update, error logs, model manager M3 – Windows/Linux alpha Replace injection backend, install bundles M4 – v1.0 GA Streaming (optional), website + docs
Should we bundle a small GGUF model or trigger a first-run download wizard?
How to handle non-Latin languages (auto-detect vs user-select)?
Do we sandbox the app on macOS or rely on hardened runtime?
Which licence (MIT vs GPL) given we embed Whisper weights?
Accept user telemetry opt-in for latency metrics?
Product Lead – @PMEngineering Lead – @TechLeadDesign – @UXSecurity – @SecQA – @QA
Reviews : Architecture (Tech), Security (Sec), Accessibility (UX).