(SEM VII) THEORY EXAMINATION 2024-25 SPEECH PROCESSING

KEC078 – SPEECH PROCESSING

Complete Solved Question Paper

SECTION A

(Attempt all questions – 2 × 10 = 20 marks)

Q1(a) Key assumptions of acoustic theory of speech production

Speech is produced by an excitation source (voiced/unvoiced).

Vocal tract acts as a linear, time-varying filter.

Source and filter are independent.

Speech signal is quasi-stationary over short intervals.

Q1(b) Acoustic phonetics

Acoustic phonetics is the branch of phonetics that studies speech sounds as physical signals, analyzing properties like frequency, amplitude, spectrum, and duration.

Q1(c) Purpose of short-time energy

Detects speech vs silence                              Identifies voiced/unvoiced regions

Useful for endpoint detection

Q1(d) AMDF vs Autocorrelation

Q1(e) Purpose of spectrographic displays            Visualize frequency content over time

Identify formants, pitch, transitions                           Analyze speech dynamics

Q1(f) Properties of STFT

Time-frequency representation                                 Based on windowing

Trade-off between time & frequency resolution       Overlapping windows improve accuracy

Q1(g) Homomorphic vocoder

A system that separates source and vocal tract components using cepstral analysis.
Used in: speech coding, speech recognition, speaker identification.

Q1(h) Formant estimation using homomorphic methods

Compute log spectrum                                               Apply inverse FFT → cepstrum

Separate low quefrency (vocal tract)                           Peaks correspond to formants

Q1(i) Prediction error in LPC

Difference between actual speech sample and predicted sample using past samples.

Q1(j) Significance of NMSE in LPC                          Measures prediction accuracy

Lower NMSE = better LPC model                                Used to evaluate LPC performance

SECTION B

(Attempt any three – 10 × 3 = 30 marks)

Q2(a) Speech production mechanism & acoustic phonetics

Speech is produced when airflow from lungs excites the vocal cords (source).
The vocal tract shapes the sound by resonance (filter).
Acoustic phonetics analyzes this sound in terms of spectral features, formants, pitch, and energy.

Q2(b) Short-time zero-crossing rate (ZCR)

ZCR counts how often signal crosses zero in a frame.              High ZCR → unvoiced sounds

Low ZCR → voiced sounds                                                        Helps in speech/silence classification

Q2(c) Short-time Fourier analysis

Speech is non-stationary                                                           Divides signal into short frames

FFT applied on each frame                                                        Enables time-frequency analysis

Q2(d) Homomorphic processing – benefits & limitations

Benefits

Separates source & filter                                                           Robust spectral analysis

Limitations                                                                               Computationally expensive

Phase unwrapping issues                                                           Less suitable for real-time systems

Q2(e) LPC & lossless tube model relationship

LPC models vocal tract as all-pole filter                         Tube model represents tract as concatenated tubes

LPC coefficients approximate tube resonances (formants)

SECTION C

(Attempt ONE from each question – 10 × 5 = 50 marks)

Q3(a) Acoustic theory of speech production

Speech is modeled as:                  Speech=Source×Vocal Tract×RadiationSpeech = Source × Vocal\ Tract × RadiationSpeech=Source×Vocal Tract×Radiation

Voiced → periodic excitation        Unvoiced → noise excitation

Vocal tract acts as filter                 Foundation for LPC, vocoders, speech synthesis

Significance                                       Simplifies speech modeling

Enables efficient speech coding         Used in ASR & TTS systems

Q3(b) Lossless tube models             Vocal tract modeled as uniform tubes

Each tube represents tract segment   Reflection coefficients simulate articulator movement

Accurately models formant frequencies

Q4(a) Short-time energy & average magnitude

Short-time energy                            E(n)=∑∣x(n)∣2w(n)E(n)=\sum |x(n)|^2 w(n)E(n)=∑∣x(n)∣2w(n)

Average magnitude                         M(n)=∑∣x(n)∣w(n)M(n)=\sum |x(n)| w(n)M(n)=∑∣x(n)∣w(n)

Uses                                                   Endpoint detection

Speech activity detection                   Voiced/unvoiced analysis

Q4(b) Autocorrelation vs AMDF for pitch

Q5(a) Spectrographic displays                         Time on x-axis

Frequency on y-axis                                             Intensity as color

Shows formants, pitch, harmonics                       Widely used in speech analysis & linguistics

Q5(b) FFT-based filter banks

Input signal → FFT                                               Frequency bins grouped into bands

Each band acts as filter                                        Efficient implementation for MFCC extraction

Q6(a) Complex cepstrum computation

Steps:                                                                   Compute FFT

Take log magnitude + phase                               Unwrap phase

Apply inverse FFT                                                 Used to separate excitation and vocal tract.

Q6(b) Homomorphic system

A system where convolution becomes addition using logarithmic transformation.
Application: source-filter separation in speech.

Q7(a) Autocorrelation method in LPC

Steps:                                                                   Frame blocking

Compute autocorrelation                                    Apply Levinson-Durbin algorithm

Obtain LPC coefficients

Q7(b) Principles of linear predictive analysis

Current sample predicted from past samples     Minimizes prediction error

Models vocal tract as all-pole filter                     Widely used in speech compression