Text-to-speech generation in Prompt Engineering / GenAI - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

What is the main role of the vocoder in text-to-speech systems?

In text-to-speech (TTS) systems, the vocoder is a key component. What does it do?

A It analyzes the input text to detect emotions.

B It converts text into phonemes for pronunciation.

C It converts the acoustic features into audible speech waveforms.

D It translates speech back into text.

💡 Hint

Think about how the system creates sound from intermediate data.

The vocoder takes acoustic features like spectrograms and generates the actual sound wave, making speech audible.

❓ Predict Output

intermediate

What is the output shape of the mel spectrogram in this TTS preprocessing code?

Given the code below that converts audio to a mel spectrogram, what is the shape of mel_spectrogram?

Prompt Engineering / GenAI

import numpy as np
import librosa

audio, sr = librosa.load('audio.wav', sr=22050)
mel_spectrogram = librosa.feature.melspectrogram(y=audio, sr=sr, n_mels=80, hop_length=256, n_fft=1024)
output_shape = mel_spectrogram.shape

A (number_of_frames, 80)

B (80, number_of_frames)

C (1024, 80)

D (256, number_of_frames)

💡 Hint

Check the documentation for librosa.feature.melspectrogram output dimensions.

The mel spectrogram has shape (n_mels, time frames). Here, n_mels=80, so shape is (80, number_of_frames).

❓ Model Choice

advanced

Which model architecture is best suited for generating natural-sounding speech waveforms in TTS?

Among these options, which model architecture is designed specifically to generate high-quality speech waveforms in text-to-speech systems?

A WaveNet

B BERT

C ResNet

D Tacotron 2

💡 Hint

Consider which model is a neural vocoder producing raw audio.

WaveNet is a neural vocoder that generates raw audio waveforms sample-by-sample, producing natural speech sounds.

❓ Hyperparameter

advanced

Which hyperparameter directly controls the time resolution of mel spectrogram frames in TTS preprocessing?

In mel spectrogram extraction, which hyperparameter affects how often frames are sampled over time?

A n_fft

B n_mels

C win_length

D hop_length

💡 Hint

Think about the step size between frames in the spectrogram.

The hop_length sets the number of audio samples between successive frames, controlling time resolution.

❓ Metrics

expert

Which metric best evaluates the naturalness of synthesized speech in TTS systems?

When assessing how natural synthesized speech sounds, which metric is most appropriate?

A Mean Opinion Score (MOS) from human listeners

B Word Error Rate (WER) from speech recognition on synthesized audio

C BLEU score comparing synthesized text to reference text

D Mean Squared Error (MSE) between predicted and target spectrograms

💡 Hint

Naturalness is subjective and often measured by human judgment.

MOS is a subjective score collected from human listeners rating speech naturalness, making it the best metric.