added voice recording shortcut, added markdown for responses for both…

[karansingh7773-rathore]

🎙️ Voice Recording & Gemini Model Selection Features

📋 Overview

This PR introduces the highly anticipated Voice Recording feature with live transcription powered by Deepgram, and seamless integration with the Gemini API to analyze the transcribed discussions. It also brings dynamic Gemini Model Selection directly to the user interface, alongside major stability improvements for API rate limits and WebSockets.

✨ New Features

1. Global "Record Voice" Shortcut (Ctrl+Shift+R)

• Added a global keyboard shortcut (Ctrl+Shift+R) that can be triggered from anywhere to instantly capture system and microphone audio.
• The UI dynamically displays a live transcription feed of the ongoing conversation in real-time.

2. Deepgram Live Transcription Integration

• Integrated the Deepgram WebSocket API (wss://api.deepgram.com/v1/listen) with the nova-3 model for state-of-the-art live Speech-to-Text (STT).
• Captures and merges both internal system desktop audio and external microphone input.
• Added live, multi-language translation support directly via Deepgram’s API features.

3. Gemini Audio Analysis Pipeline

• Once recording is stopped, the final highly-accurate Deepgram transcript is immediately forwarded to the selected Gemini model.
• Automatically generates AI responses and problem-solving solutions based purely on the spoken conversation.

4. Dynamic Gemini Model Selector

• Added a dedicated settings interface for users to select their preferred Gemini model.
• Includes support for the newest high-throughput, free-tier models:
  • Gemini 2.5 Flash (5 RPM · 20 RPD)
  • Gemini 3 Flash (5 RPM · 20 RPD)
  • Gemini 2.5 Flash Lite (10 RPM · 20 RPD)
• Auto-persists the user's model selection and language targets to localStorage.

🛠️ Stability & Performance Fixes

• Deepgram Audio Stream Buffering: Implemented a pendingChunks queue to buffer all early audio chunks locally. Once the Deepgram socket.onopen fires, the queue is instantaneously flushed. This guarantees that critical WebM Container Headers are never silently dropped before the connection opens, ensuring 100% reliable transcription decoding.
• Robust Mutex for Concurrent Toggles: Shifted the recording status check to a synchronous React ref (isRecordingRef) with a hard mutex (isTogglingRef) wrapped in a try/finally block. This strictly prevents the toggleRecording execution from being double-triggered, eliminating the Gemini 429 Too Many Requests API spam bug.
• Micro-recording Guard: Implemented a minimum duration tracking metric (1 second). Accidental double-taps on the shortcut will still halt the recording correctly but will gracefully skip submitting the negligible audio payload safely, saving valuable RPM/RPD limits.