THEORY EXAMINATION (SEM–VIII) 2016-17 SPEECH PROCESSING

SECTION A – Basic Concepts of Speech Processing

Section A contains short conceptual questions related to speech signals, signal processing, and speech analysis techniques. These questions focus on the fundamental concepts used in speech processing systems. 

Question (a): What is Pitch?

Answer:
Pitch is the perceptual characteristic of sound that determines whether a sound is perceived as high or low. In speech processing, pitch corresponds to the fundamental frequency of the speech signal generated by the vibration of vocal cords.

When the vocal cords vibrate rapidly, the pitch becomes high. When the vibration is slow, the pitch becomes low.

Pitch plays an important role in:

Speech recognition

Speaker identification

Speech synthesis

It also helps differentiate between male and female voices since male voices generally have lower pitch compared to female voices.

Question (b): Explain Acoustic Phonetics

Answer:
Acoustic phonetics is the branch of phonetics that studies the physical properties of speech sounds. It focuses on analyzing sound waves produced during speech.

Acoustic phonetics examines:

Frequency of speech signals

Amplitude of sound waves

Duration of speech sounds

Spectral characteristics

By studying these properties, researchers can understand how speech signals are generated and how they can be analyzed and processed digitally.

Question (c): Why is Sampling Required?

Answer:
Sampling is required to convert an analog speech signal into a digital signal so that it can be processed by computers.

Speech signals are naturally continuous signals. However, digital systems require discrete signals. Sampling captures the signal amplitude at regular intervals.

According to the Nyquist theorem, the sampling frequency must be at least twice the highest frequency present in the signal to avoid distortion.

For example, telephone speech signals are typically sampled at 8 kHz.

Question (d): Define Channel Vocoder

Answer:
A channel vocoder is a speech processing system used for speech compression and analysis.

It divides the speech signal into multiple frequency channels using band-pass filters. Each channel extracts information about the energy of the signal within that frequency band.

Instead of transmitting the entire speech signal, the vocoder transmits only the extracted parameters such as:

Amplitude

Pitch

Frequency band information

This reduces the amount of data required to transmit speech signals.

Question (e): What is Frequency Domain?

Answer:
The frequency domain represents a signal in terms of its frequency components instead of time.

In speech processing, analyzing signals in the frequency domain helps identify characteristics such as:

Pitch

Harmonics

Formants

Mathematical techniques like the Fourier Transform are used to convert signals from time domain to frequency domain.

This analysis helps understand how different frequencies contribute to speech signals.

Question (f): Define Correlation Function with Example

Answer:
The correlation function measures the similarity between two signals or between a signal and a delayed version of itself.

In speech processing, correlation functions are often used for pitch detection.

Example:
If a speech signal has periodic patterns, the correlation function will show peaks at time intervals corresponding to the pitch period.

Thus, correlation analysis helps detect repeating patterns in speech signals.

Question (g): What is a Filter?

Answer:
A filter is a device or algorithm used to remove unwanted components from a signal or isolate specific frequency ranges.

In speech processing, filters are used to:

Remove background noise

Enhance speech clarity

Extract important frequency components

Common types of filters include:

Low-pass filter

High-pass filter

Band-pass filter

Band-stop filter

Filters play a critical role in improving the quality of speech signals.

Question (h): Differentiate Between Speech and Silence

Speech processing systems must detect silence segments to improve processing efficiency and reduce data storage.

Question (i): Define Convolution with Example

Answer:
Convolution is a mathematical operation used to combine two signals to produce a third signal.

In speech processing, convolution is used to model how speech signals pass through systems like filters or the vocal tract.

Example:
If a speech signal passes through a filter, the output signal is the convolution of the input signal and the filter impulse response.

Convolution helps analyze how systems affect speech signals.

Question (j): What is Linear Predictive Coding (LPC)?

Answer:
Linear Predictive Coding is a technique used to represent speech signals efficiently.

It works by predicting the current speech sample based on a linear combination of previous speech samples.

LPC extracts parameters that describe the vocal tract characteristics.

Applications of LPC include:

Speech compression

Speech synthesis

Voice transmission systems

LPC significantly reduces the amount of data required to represent speech signals while maintaining intelligibility.

SECTION B – Intermediate Concepts of Speech Processing

Section B focuses on speech signal modeling, pitch detection, and speech parameter analysis. 

Question: Sampling and Quantization of Speech Signals

Answer:
Sampling and quantization are essential steps in converting analog speech signals into digital form.

Sampling captures the signal amplitude at regular intervals. This converts a continuous-time signal into a discrete-time signal.

Quantization converts sampled amplitudes into discrete numerical levels so they can be stored digitally.

For example, in digital audio recording, the microphone captures analog speech signals which are then sampled and quantized before being stored in digital format.

Question: Digital Models for Speech Signals

Answer:
Digital speech models represent speech signals mathematically.

One commonly used model is the source-filter model, which assumes speech production involves:

A sound source (vocal cords)

A filter (vocal tract)

The vocal tract shapes the sound produced by the vocal cords to generate different speech sounds.

This model is widely used in speech synthesis and speech recognition systems.

Question: Applications of Speech Processing

Speech processing has many real-world applications, including:

Speech recognition systems

Voice assistants like Siri or Alexa

Speech synthesis systems

Speaker identification

Automated call centers

Hearing aids

These technologies improve human-computer interaction and accessibility.

Question: Short-Term Pitch Detection

Short-term pitch detection determines the pitch of speech signals within small time frames.

The process includes:

Dividing speech signals into short frames

Computing correlation values

Identifying peaks corresponding to pitch periods

Pitch detection helps identify whether speech is voiced or unvoiced.

SECTION C – Advanced Speech Processing Concepts

Section C questions require deeper understanding of speech analysis and synthesis techniques. 

Question: Speech Synthesis and LPC

Speech synthesis refers to generating artificial speech using computers.

Speech synthesis systems convert text into speech signals using several stages such as:

Text analysis

Phoneme generation

Speech waveform generation

Linear Predictive Coding plays a crucial role in speech synthesis because it models the vocal tract and produces natural-sounding speech.

LPC uses mathematical equations to estimate predictor coefficients that represent speech signals efficiently.

Question: Short-Time Fourier Analysis

Short-Time Fourier Transform (STFT) analyzes how the frequency components of speech signals change over time.

Speech signals are non-stationary, meaning their properties change over time.

To analyze such signals, STFT divides them into short frames and computes Fourier transforms for each frame.

This allows visualization of speech signals using spectrograms, which show frequency variation over time.

Question: Autocorrelation, NMSE, and Formant Estimation

Autocorrelation Method

Autocorrelation measures similarity between a signal and delayed versions of itself. It is commonly used for pitch detection.

Normalized Mean Square Error (NMSE)

NMSE measures the difference between predicted and actual speech signals. It is used to evaluate the accuracy of speech models.

Formant Estimation

Formants are resonance frequencies of the vocal tract. They help identify vowel sounds and are important in speech recognition systems.

Conclusion

Speech processing combines signal processing techniques with linguistic knowledge to analyze, synthesize, and recognize speech signals. Concepts such as sampling, pitch detection, filtering, and linear predictive coding are essential for building modern speech technologies.

These technologies power systems such as voice assistants, automated translation tools, and speech recognition systems.