Translator for Atypical speech
Fine-tuning Whisper on 500 phrases to translate his brother's unique speech patterns
Deep Dive
A Reddit user is building a translator app for his profoundly deaf brother, whose Bengali speech is largely unintelligible to strangers but can be picked up in 2-3 months and is fully understood by close family. He's following LLM advice to fine-tune a Bengali Whisper model on a custom dataset of around 500 common phrases, recorded with his brother’s speech, aiming to translate atypical to typical speech for real-world navigation. He's seeking validation of his approach.
Key Points
- Fine-tuning Whisper on 500 Bengali phrases recorded by the user's deaf brother
- Brother's speech is understood by family after 2-3 months of exposure, indicating learnable patterns
- User created the phrase dataset with AI help and plans to build a real-time mobile translator
Why It Matters
Demonstrates how accessible AI can create personalized accessibility tools for rare communication needs.