| Property | Value |
|---|---|
| Base Model | openai/whisper-large-v3 |
| Training Method | LoRA Fine-tuning |
| Task | Hindi Speech Recognition |
| Dataset | Common Voice 13.0 Hindi |
| Model Author | kasunw |
What is whisper-large-v3-hindi?
Whisper-large-v3-hindi is a specialized automatic speech recognition (ASR) model fine-tuned specifically for Hindi language processing. Built upon OpenAI's Whisper large-v3 architecture, this model leverages LoRA (Low-Rank Adaptation) training techniques to optimize performance for Hindi speech recognition while maintaining computational efficiency.
Implementation Details
The model implementation utilizes the PEFT (Parameter-Efficient Fine-Tuning) framework and Transformers library. It's designed to run with FP16 precision on compatible hardware and includes optimizations for batch processing of audio segments.
- Supports processing of 30-second audio chunks
- Implements batch processing with size 16
- Includes timestamp generation capability
- Optimized for both CPU and GPU deployment
Core Capabilities
- Hindi speech-to-text transcription
- Efficient processing of long audio files through chunking
- Timestamp generation for word alignment
- Support for both inference and fine-tuning workflows
Frequently Asked Questions
Q: What makes this model unique?
This model combines the robust capabilities of Whisper large-v3 with specialized Hindi language optimization through LoRA fine-tuning, making it particularly effective for Hindi ASR tasks while maintaining memory efficiency.
Q: What are the recommended use cases?
The model is ideal for Hindi speech transcription tasks, including subtitle generation, voice command processing, and general speech-to-text applications focusing on Hindi language content. It's particularly suitable for applications requiring batch processing of audio files.