Updated the TTS inference and made it work

[AdvityaDua]

Description
This PR significantly upgrades the interview backend by migrating speech-to-text inference to Groq's high-performance API, enhancing the voice analysis scoring system, and introducing LLM-based CV evaluation.

Key Changes
🎤 Speech-to-Text (ASR)
Groq Integration: Replaced the local Whisper implementation with Groq-hosted Whisper (whisper-large-v3). This removes the need for local model loading and significantly reduces latency.
Dependency Removal: Removed ffmpeg normalization steps as Groq handles audio processing, making the backend lighter and easier to deploy.
Robustness: Added safety checks for empty audio, "None" hallucinations, and safe Groq client initialization.
📊 Voice Analysis
Scaled Scoring: The
VoiceAnalyzer
now returns scores scaled out of 10 (for Fluency, Clarity, Confidence, Pace) alongside the original raw weighted scores. This aligns with the frontend's requirement for a 0-10 scale.
Improved Metrics:
WPM Calculation: Now uses the actual transcript (if provided) for more accurate Speech Rate (WPM) calculation.
Pitch & Energy: Integrated librosa.pyin for robust pitch tracking and RMS for energy/clarity analysis.
Error Handling: Added an explicit analysis_ok flag and structured error codes for better frontend feedback.
📄 CV Evaluation
LLM Scorer: Implemented LLMScorer using Groq to intelligently evaluate CVs against Job Descriptions.
Heuristic Fallback: Added graceful fallbacks to heuristic scoring if the LLM API is unavailable.
Structured Output: Ensures JSON parsing reliability for LLM responses.
🧪 Testing
Added comprehensive unit tests for LLMScorer and CVEvaluationEngine.
Added debug_groq.py and debug_voice_import.py utilities for environment verification.