Miraa Alternative

An open-source alternative to Miraa, a Japanese transcription and translation app. Multilingual audio analysis and alignment system for transcription, translation, and visualization.

🚀 What it does

• Processes music into aligned, translated text and explained using a LLM
• Supports en-jp transcription and translation
• Visualizes timing, confidence, and alignment results
• Searches through Japanese dictionaries to find definitons of words
• Separates audio using demucs and masks for better separation
• Auto downloading from YouTube for better input

🧠 How it works

• Splits audio into stems (vocal & instrumental)
• Audio preprocessing & segmentation (VAD-based)
• Speech recognition & translation using pretrained models
• Using Genius API to get lyrics
• Saves data to a .json file for easy viewing (debugging)
• HTML-based interactive visualization (soon)

🛠 Tech Stack

• Python
• PyTorch / torchaudio
• Speech & translation models
• Transformers
• HTML / JS for dashboards
• External APIs & web scraping

📊 Results

• Improved alignment accuracy through iterative refinement (fine-tuning)
• Explanation of Japanese songs and their meaning (for studying the language)
• Robust handling of noisy real-world audio (actual songs)
• Scalable pipeline design (soon)

🧪 Status

Actively iterating and experimenting

---

Flowchart

This diagram shows how the program functions:

Backend

---