Audio transcription (Whisper) in Prompt Engineering / GenAI - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output of Whisper transcription code snippet

What is the output of the following Python code that uses Whisper to transcribe a short audio file?

Prompt Engineering / GenAI

import whisper
model = whisper.load_model('small')
result = model.transcribe('audio_sample.wav')
print(result['text'])

A "[Noise] Unintelligible speech detected."

B "Hello, this is a test audio for transcription."

C SyntaxError: invalid syntax

D FileNotFoundError: [Errno 2] No such file or directory: 'audio_sample.wav'

💡 Hint

The model transcribes the audio content into text stored in the 'text' key of the result dictionary.

The Whisper model's transcribe method returns a dictionary with a 'text' key containing the transcription of the audio file. If the file exists and is clear, it outputs the spoken words as text.

❓ Model Choice

intermediate

Choosing the right Whisper model for fast transcription

You want to transcribe a large number of short audio clips quickly with reasonable accuracy. Which Whisper model should you choose?

A tiny

B large-v2

C medium

D large

💡 Hint

Smaller models run faster but with less accuracy.

The 'tiny' model is the fastest and smallest Whisper model, suitable for quick transcription of many short clips with acceptable accuracy.

❓ Hyperparameter

advanced

Effect of temperature parameter in Whisper transcription

In Whisper's transcribe method, what is the effect of increasing the 'temperature' parameter from 0.0 to 1.0?

A Decreases transcription speed significantly

B Improves transcription accuracy by reducing errors

C Increases randomness in transcription, possibly producing more diverse but less stable text

D Switches the model to a different language automatically

💡 Hint

Temperature controls randomness in text generation.

Higher temperature values increase randomness in the output, which can lead to more varied but less consistent transcriptions.

❓ Metrics

advanced

Evaluating Whisper transcription quality

Which metric is most appropriate to measure the accuracy of Whisper's transcriptions compared to ground truth text?

A Word Error Rate (WER)

B Mean Squared Error (MSE)

C Accuracy Score

D F1 Score

💡 Hint

This metric counts word-level differences between predicted and true text.

Word Error Rate (WER) measures the number of word insertions, deletions, and substitutions needed to match the transcription to the reference, making it ideal for speech recognition evaluation.

🔧 Debug

expert

Debugging a Whisper transcription error

You run the following code but get a RuntimeError: CUDA out of memory. What is the best way to fix this error?

Prompt Engineering / GenAI

import whisper
model = whisper.load_model('large')
result = model.transcribe('long_audio.wav')
print(result['text'])

A Run the code without a GPU by setting device='cpu' in load_model()

B Increase the batch size to process more audio at once

C Use a higher temperature value in transcribe()

D Switch to a smaller model like 'medium' or 'small' to reduce GPU memory usage

💡 Hint

Large models use more GPU memory; smaller models use less.

The 'large' Whisper model requires a lot of GPU memory. Using a smaller model reduces memory needs and avoids the out-of-memory error.