Logo Lanfrica

mohdali-dev/BaltiVoice-ASR

Domaine:

natural language processing

Type de record:

model
Créateur:
moh
Hôte:
First public ASR system for Balti (bft) - low-resource Tibetic language. Fine-tuned Whisper-small | WER: 30% | Live Demo # 🎙️ BaltiVoice ASR ### First AI-Powered Automatic Speech Recognition for the Balti Language ### *Speak Balti. Get Transcription. Preserve Culture.* --- ## 🌟 Why BaltiVoice? The Mission ### 🗣️ The Language: Balti (بلتی) Balti is a Tibetic language spoken by approximately 400,000 to 450,000 people in Gilgit-Baltistan, Pakistan, and Ladakh, India. With roots in Classical Tibetan, it carries centuries of oral history, poetry, and cultural identity. | Feature | Detail | | --- | --- | | **Script** | Primarily Perso-Arabic (Nastaliq); historically Tibetan. | | **Status** | 🔴 **Critically Low-Resource**: No standard digital tools, no ASR models, limited NLP datasets. | | **Risk** | Without digital representation, the language risks fading from the modern technological landscape. | ### 💔 The Problem: A Digital Silence Despite its rich heritage, Balti suffers from "Digital Neglect." - ❌ **Zero Public ASR**: Until now, there was no open-source speech-to-text model for Balti. - ❌ **No Standard Data**: Researchers had no validated audio-text pairs to build upon. - ❌ **Accessibility Barrier**: Native speakers could not use voice assistants, transcription tools, or educational apps in their mother tongue. > *"When a language lacks digital tools, it becomes invisible to the future. I built BaltiVoice to ensure Balti has a seat at the AI table."* ### 🚀 Why I Built This Model As an AI/ML Engineer, I wanted to bridge the gap between cutting-edge technology and underserved communities. This project is about more than fine-tuning Whisper—it’s about preservation through code. - **Preserve Oral History**: By digitizing speech, we create a permanent archive of Balti pronunciation and dialects. - **Prove Low-Resource Viability**: Demonstrating that high-quality ASR is possible with just ~16 hours of data using transfer learning. - **Empower Future Builders**: This dataset and model serve as a foundation for: - 🎓 **Education**: Tools to teach Balti literacy. - 🏥 * …