Can I build my own Wispr Flow with AI?

Build a browser-based AI dictation tool inspired by Wispr Flow (web only, not a system-wide desktop app).

**Objective:** Let a user hold a record button, transcribe speech with an API, optionally clean up the text, and copy it for use elsewhere.

**Target user:** A writer or knowledge worker who wants voice-to-text in the browser without a Wispr subscription.

**Core functionality:**
- Push-to-talk or click-to-record microphone capture
- Transcribe audio via speech-to-text API
- Optional "Clean up" that fixes punctuation and filler words via LLM
- Copy transcript to clipboard in one click
- History of recent dictations
- Simple tone presets (email, notes, casual)

**Pages/screens:**
- Dictation studio (main)
- History list
- Settings (API keys server-side, default cleanup preset)

**User flows:**
1. User opens the app and grants mic permission
2. User records a thought
3. Transcript appears; user optionally runs Clean up
4. User copies text into another app
5. User revisits a past dictation from History

**UI direction:** Calm productivity UI, large record button, transcript as the hero element, minimal chrome. Keyboard shortcut for start/stop if the browser allows.

**Data model:**
- User: { id, email } (optional)
- Dictation: { id, userId?, rawText, cleanedText?, durationMs, createdAt }

**Authentication:** Optional for a personal local-history tool; recommended if syncing across devices.

**APIs/integrations:** Speech-to-text API; optional LLM for cleanup. Keep keys on the server.

**Responsive requirements:** Must work on mobile Safari/Chrome with mic permissions.

**Error states:** Mic permission denied; empty recording; API timeout with retry.

**Explicit exclusions:** No OS-wide hotkey, no paste-into-any-app, no menu-bar native client, no offline on-device model in v1.

**Suggested implementation order:**
1. Mic recording in browser
2. Upload/transcribe via STT API
3. Copy + history
4. LLM cleanup presets
5. Mobile polish and error states