Speak It

Speak naturally into any text field on the web. The extension transcribes your speech in real-time and formats it based on where you're typing. Emails get proper greetings. Slack stays casual. Twitter keeps it short.

The key difference: instead of storing your messages to learn your style, it stores statistics - sentence length, formality scores, phrase frequencies. A style fingerprint, not a transcript.

Voice-to-Text Anywhere

• Works on any website with text inputs
• Keyboard shortcut: Alt+Shift+S to toggle recording
• Real-time transcription with live preview
• Platform-aware formatting (Gmail, Slack, Twitter, Notion, Google Docs, etc.)

Style Learning

• Learns how you write without storing what you write
• Extracts statistics: average sentence length, formality score, contractions usage
• Stores common greetings, sign-offs, and phrases
• Style profile improves over time (diminishing influence prevents drift)

Meeting Mode

• Captures tab audio from Google Meet, Zoom, or any browser-based meeting
• Mixes in your microphone for full conversation capture
• Speaker diarization (labels who said what)
• Live transcript in side panel
• Auto-saves meeting notes to web app

Context Detection

• Detects platform automatically (no manual switching)
• Adjusts formality, capitalization, and structure per platform
• Preserves your words while cleaning up grammar and filler words

What Gets Stored

CREATE TABLE user_style_profiles (
  user_id VARCHAR(255) PRIMARY KEY,
  avg_sentence_length FLOAT,
  formality_score FLOAT,        -- 0 (casual) to 1 (formal)
  uses_contractions BOOLEAN,
  emoji_frequency FLOAT,
  top_phrases JSON,             -- ["sounds good", "let me know", ...]
  greetings JSON,               -- ["Hey", "Hi", ...]
  signoffs JSON,                -- ["Thanks", "Cheers", ...]
  sample_count INT,
  updated_at TIMESTAMP
);

What Never Gets Stored

• Message content
• Transcripts
• Audio recordings
• Conversation history

How It Updates

Each time you accept a formatted message:

1. Statistics are extracted from the approved text
2. New stats are blended with existing profile
3. Influence decreases as sample count increases (1st message = 100%, 100th message = ~1%)
4. Profile stabilizes over time

Voice-to-Text Mode

// Capture mic at 16kHz
const stream = await navigator.mediaDevices.getUserMedia({
  audio: { sampleRate: 16000, echoCancellation: true, noiseSuppression: true }
});

// Convert Float32 to Int16 PCM for Deepgram
processor.onaudioprocess = (event) => {
  const input = event.inputBuffer.getChannelData(0);
  const pcm = new Int16Array(input.length);
  for (let i = 0; i < input.length; i++) {
    const sample = Math.max(-1, Math.min(1, input[i]));
    pcm[i] = sample < 0 ? sample * 0x8000 : sample * 0x7fff;
  }
  socket.send(pcm.buffer);
};

Meeting Mode Audio Mixing

// Tab audio (others' voices) + Mic audio (your voice)
const mixLength = Math.min(tabPcm.length, micChunk.length);
for (let i = 0; i < mixLength; i++) {
  // Mic slightly louder since it's your voice
  const mixed = (tabPcm[i] * 0.6) + (micChunk[i] * 0.8);
  finalPcm[i] = Math.max(-32768, Math.min(32767, Math.round(mixed)));
}

Why Mic Capture is in Content Script

Chrome offscreen documents can't show permission prompts (they're invisible). Mic capture silently fails. The fix: capture mic from the content script (runs on visible page), stream chunks to offscreen document for mixing with tab audio.

Every platform implements text input differently:

// Notion fallback
navigator.clipboard.writeText(text);
showMessage('Press Cmd+V to paste');

const url = new URL('wss://api.deepgram.com/v1/listen');
url.searchParams.set('model', 'nova-2');
url.searchParams.set('language', 'en');
url.searchParams.set('smart_format', 'true');
url.searchParams.set('punctuate', 'true');
url.searchParams.set('diarize', 'true');        // Speaker labels
url.searchParams.set('utterances', 'true');
url.searchParams.set('interim_results', 'true'); // Live preview
url.searchParams.set('encoding', 'linear16');
url.searchParams.set('sample_rate', '16000');
url.searchParams.set('channels', '1');

function buildStylePrompt(profile, context) {
  const formality = profile.formality_score > 0.7 ? "formal" :
                    profile.formality_score < 0.3 ? "casual" : "balanced";

  return `Format this transcript for ${context}.

User's writing style:
- Tone: ${formality}
- Average sentence length: ~${Math.round(profile.avg_sentence_length)} words
- ${profile.uses_contractions ? "Use contractions naturally." : "Avoid contractions."}
- Preferred greetings: ${profile.greetings.slice(0, 3).join(", ")}
- Preferred sign-offs: ${profile.signoffs.slice(0, 3).join(", ")}

Rules:
1. ONLY add punctuation and paragraph breaks
2. Remove filler words: um, uh, like, basically, you know
3. Keep EVERY other word exactly as spoken
4. Do NOT rewrite, rephrase, or improve their language`;
}

speak-it/
├── manifest.json           # Extension config (Manifest V3)
├── background.js           # Service worker, message routing
├── content.js              # UI overlay, text insertion, mic capture
├── content.css             # Overlay styles
├── offscreen.html          # Audio processing document
├── offscreen.js            # Tab capture, Deepgram streaming, mixing
├── popup.html              # Extension popup UI
├── popup.js                # Popup logic
├── meeting-panel.html      # Side panel for Meeting Mode
├── meeting-panel.js        # Meeting UI logic
├── meeting-panel.css       # Meeting panel styles
└── icons/                  # Extension icons

• No message storage: Your words are processed and discarded
• Statistics only: Style profiles contain patterns, not content
• Local-first: Extension works offline for basic transcription
• You control your data: Delete your profile anytime

• Notion requires manual paste (Cmd+V)
• Google Docs text insertion is slow (character-by-character)
• Meeting Mode requires mic permission on each new domain
• Deepgram free tier has usage limits

The repo is private right now.

If you're a recruiter or hiring manager and want to walk through the code or architecture in more detail, I'm happy to do a live session.