Fine-Tuning OpenAI Whisper for Persian Speech Recognition

A complete and practical guide for fine-tuning OpenAI's Whisper model on Persian language datasets to achieve high-accuracy automatic speech recognition (ASR) for Farsi audio.

Project Overview

This repository provides a clear, reproducible pipeline for fine-tuning Whisper — OpenAI's powerful speech-to-text model — using Persian speech datasets.
The goal is to enhance Whisper’s performance specifically for Farsi transcriptions by training it on native Persian audio and text.

Key Features

• Fine-tunes Whisper on Persian speech for improved ASR results
• Step-by-step code using Hugging Face Transformers
• Uses WER (Word Error Rate) for accurate evaluation
• Supports Common Voice Persian dataset (or your own dataset)
• Designed for easy customization and GPU-accelerated training

Why Fine-Tune Whisper for Persian?

Although OpenAI's Whisper supports multiple languages, fine-tuning it on Persian-specific datasets can drastically improve transcription quality.
This is especially useful for:

• Voice assistant applications in Farsi
• Persian audio-to-text services
• Academic and commercial speech AI research

Requirements

• Python 3.8+
• PyTorch with GPU support
• Hugging Face transformers and datasets
• evaluate for metric calculation (Word Error Rate)