abar-uwc/united-med-asr-indic

United Medical ASR Indic (Accelerated Training)

This model is a fine-tuned version of distil-whisper/distil-large-v3.5 for multilingual medical speech recognition across Indian languages, trained using HuggingFace Accelerate for optimal performance.

Performance Optimizations

• Multi-GPU Training: Distributed training across available GPUs
• Mixed Precision: FP16 training for 2x speed improvement
• Gradient Accumulation: Effective large batch sizes
• Memory Optimization: Gradient checkpointing and efficient data loading
• Batch Processing: Locations processed in optimized batches

Model Description

• Base Model: distil-whisper/distil-large-v3.5
• Languages: Multiple Indian languages (Hindi, English, Bengali, Tamil, Telugu, Gujarati, Kannada, Malayalam, Marathi, Punjabi, Urdu)
• Domain: Medical ASR for Indic languages
• Training: Accelerated sequential fine-tuning with batch processing

Usage

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch

processor = WhisperProcessor.from_pretrained("abar-uwc/united-med-asr-indic")
model = WhisperForConditionalGeneration.from_pretrained("abar-uwc/united-med-asr-indic")

# The model will automatically detect and transcribe in the appropriate language