AI Engineer Interview - Practical Task: Automatic Speech Recognition (ASR)

Overview

This repository contains a sample audio file for a practical interview question for the AI Engineer position. The goal is to assess your ability to implement an end-to-end ASR (Automatic Speech Recognition) system using modern AI techniques.

You are required to create a complete Google Colab notebook that processes the provided audio file and outputs the transcribed text. The notebook should demonstrate your skills in data handling, model selection/implementation, and inference.

Task Description

1. Input: Use the sample audio file provided in this repository: sample_audio.mp3 (or replace with the actual filename if different). This is a short voice recording in Persian for testing ASR capabilities.

2. Requirements:

   • Implement an ASR algorithm from scratch or using pre-trained models (e.g., via libraries like Hugging Face Transformers, SpeechRecognition, or Whisper by OpenAI). You may fine-tune if necessary, but keep it efficient for Colab.
   • The notebook must include:
     • Loading and preprocessing the audio file (e.g., handling sampling rate, noise reduction if needed).
     • The core ASR model/inference pipeline.
     • Output: The transcribed text from the audio.
     • Evaluation: Optionally, include metrics like Word Error Rate (WER) if you have a ground truth transcript (not provided; you can assume or generate one for demo).
   • Use Python and standard libraries/tools available in Colab (e.g., install dependencies like transformers, torch, librosa, etc., via !pip install).
   • Ensure the notebook is self-contained, runnable in Colab, and well-documented with comments explaining each step.
   • Handle edge cases: Audio format compatibility, potential errors, and runtime efficiency (should run in under 5-10 minutes on Colab's free tier).

3. Deliverables:

   • A single Google Colab notebook (.ipynb file AND its shareable LINK).
   • The final output cell should print the transcribed text clearly, e.g.:

     Transcribed Text: "This is a sample voice for ASR testing."
     

   • Submit the notebook via [email/link/pull request] as instructed in the interview process.

Guidelines

• Time Limit: Complete this by the required deadline that your AI mentor/interviewer told to simulate real-world problem-solving.
• Originality: Avoid copying entire code blocks from tutorials; demonstrate understanding by customizing the implementation.
• Best Practices: Use version control (e.g., commit your notebook to a fork of this repo if possible), modular code, and error handling.
• Resources: You can reference open-source models like Whisper, Wav2Vec, or CTC-based approaches. No proprietary APIs without free access.
• Questions? If anything is unclear, ask during the interview.

Getting Started

1. Fork or clone this repository.
2. Open Google Colab: Go to colab.research.google.com and upload/create your notebook.
3. Download the sample_audio.mp3 from this repo and upload it to your Colab session (or access via GitHub URL).
4. Implement and test your ASR pipeline.

Good luck! We're excited to see your solution.