v1.1.0 - Enhanced Keyword Handling & Audio Preprocessing

Release Notes - v1.1.0

Enhanced Keyword Handling & Audio Preprocessing 🚀

This release introduces powerful new keyword capabilities and audio preprocessing features, along with significant reliability improvements for production workloads.

✨ New Features

Keyword Spotting 2.0

• ✅ Negated keyword support: Use !keyword syntax to detect segments without specific terms
• 🎯 Improved accuracy: Better handling of homophones through fuzzy matching enhancements
• 📊 Enhanced logging: Detailed keyword processing metrics in request logs

Audio Preprocessing Pipeline

• 🎧 Smart audio enhancement: New preprocessing options including:
  • DC offset removal
  • Normalization & dynamic range compression
  • Noise reduction & silence trimming
  • Pre-emphasis filtering
• ⚙️ Configurable workflow: Enable/disable steps via environment variables
• 🖥️ UI integration: Toggle preprocessing in web interface

Gemma Improvements

• 🔒 Safer model loading: Better error handling for GPU memory constraints
• 📈 Performance metrics: Detailed memory usage tracking during initialization

🛠️ Improvements

• 💾 GPU memory management: Optimized VRAM allocation logging and fallback logic
• ✨ Type hint modernization: Better developer experience across codebase
• 📄 Documentation overhaul: Updated README with preprocessing guides and new API examples
• 🌐 UI enhancements: File preprocessing toggle and improved status displays

⚠️ Updated Known Issues

• Gemma-2b-it still requires >8GB GPU RAM for optimal performance
• AAC support continues to require FFmpeg preprocessing
• Keyword confidence thresholds may need adjustment for technical vocabulary

---

🚨 Migration Note
Add these new environment variables to your .env file for preprocessing controls:

AUDIO_ENABLE_PREPROCESSING=true
AUDIO_ENABLE_DC_OFFSET=true
AUDIO_ENABLE_NORMALIZATION=true
AUDIO_PRE_EMPHASIS=0.97

v1.0.0 - Initial Release

Release Notes - v1.0.0 (Initial Release)

The First Stable Release 🎉
Production-Ready Audio Transcription & Text Reconstruction Service

---

🚀 New Features

Core Functionality

• Whisper Model Integration: Full support for OpenAI's Whisper speech-to-text models (base, small, medium, large)
• Gemma Reconstruction: Text refinement using Google's Gemma-2b-it LLM
• Multi-Format Processing:
  • Direct file uploads via /transcribe endpoint
  • Remote URL processing via /pull endpoint

API Capabilities

• 🔐 API key authentication middleware
• 🚦 Rate limiting (10,000 requests/hour)
• 🌐 Multi-language support (Ukrainian/Russian primary focus)
• 🔍 Keyword spotting with confidence scoring
• ⏱️ Processing time metrics in all responses

Infrastructure

• 🧠 Smart model caching system with:
  • Automatic GPU/CPU fallback
  • Memory optimization
  • Concurrent request safety
• 📊 Detailed logging (app.log & request logs)
• 🐳 Production-ready Gunicorn configuration

---

⚠️ Known Issues

Performance

• Initial model load time can be slow (~30-60s for large Whisper models)
• Gemma models requires authorization for downloading
• Gemma-2b-it requires >8GB GPU RAM for optimal performance
• No native AAC audio support - requires FFmpeg preprocessing

Limitations

• Maximum file size hard-capped at 50MB
• Keyword spotting accuracy decreases with homophones
• UI only supports basic upload functionality

---

🛠️ Upgrade Guide

New Requirements

# Required system packages
sudo apt-get install ffmpeg python3-dev

Critical Configuration

# .env changes from pre-1.0 versions
API_KEY_ENABLED=true # Now required by default
MAX_CONTENT_LENGTH=52428800 # Explicit size limit

---

📦 Installation

# For first-time users
git clone https://github.com/yourorg/audio-transcription-service.git
cd audio-transcription-service
pip install -r requirements.txt

---

🙌 Acknowledgments

• OpenAI for Whisper speech recognition models
• Google Research for Gemma language models
• Flask & Torch communities for foundational libraries

---

First production release marks completion of core feature set. Subsequent releases will focus on performance optimization and expanded language support.