Nepali-Automatic-Speech-Recognition

Welcome to the Nepali Whisper Model for Automatic Speech Recognition

Hey there! If you’re looking to transcribe Nepali audio into text, you’ve come to the right place. I’ve fine-tuned the Whisper small model specifically for Nepali speech recognition. Below, you’ll find all the details you need to get started, including where to find the datasets and notebooks!

🌟 Model Overview

• Model Name: Whisper Small
• Language: Nepali
• Current Error Rate: 30 (achieved after training for 30 epochs)

Where to Find the Good Stuff

• Dataset: I created a combined dataset for Nepali language transcription, which you can access here on Hugging Face.
• Training Notebook: Curious about how I trained this model? Check out my Kaggle Notebook to see all the steps I took!

🔧 How to Use This Model

Using this model is pretty straightforward! Here’s a quick guide to get you up and running:

Step 1: Install Required Libraries

Before we dive in, make sure you have the necessary libraries installed. Open your terminal or notebook and run:

!pip install torch transformers yt-dlp

Step 2: Load the Model

To get started with transcribing audio, you’ll need to load the Whisper model using the Hugging Face transformers library. Here’s how:

from transformers import pipeline

# Load the Whisper model for automatic speech recognition
pipe = pipeline("automatic-speech-recognition", model="amitpant7/Nepali-Automatic-Speech-Recognition")

Step 3: Transcribe an Audio File

If you have an audio file ready, you can transcribe it easily. Just point to your audio file (like audio.mp3) and run the following:

# Run inference on an audio file
result = pipe('audio.mp3')
print(result['text'])  # This will print the transcribed text

Step 4: Download and Transcribe YouTube Audio

Want to transcribe audio from a YouTube video? No problem! Use the following code to download the audio and transcribe it:

import yt_dlp

def download_youtube_audio(youtube_url, output_path="audio"):
    ydl_opts = {
        'format': 'bestaudio/best',
        'outtmpl': output_path,
        'postprocessors': [{
            'key': 'FFmpegExtractAudio',
            'preferredcodec': 'mp3',
            'preferredquality': '192',
        }],
    }

    with yt_dlp.YoutubeDL(ydl_opts) as ydl:
        ydl.download([youtube_url])
    return output_path

# Replace with the actual YouTube URL
youtube_url = "https://www.youtube.com/watch?v=H-ExNmHo2xI&pp=ygURYWJvdXQgeWFtYSBidWRkaGE%3D"
audio_file = download_youtube_audio(youtube_url)

# Now, run inference on the downloaded audio
result = pipe(audio_file)
print(result['text'])

📝 A Few Things to Keep in Mind

• The model works best with clear audio, so try to avoid noisy environments when recording.
• Performance may vary based on accents and pronunciation, but the model is designed to handle a variety of speech patterns.