Source-linked AI summary
Techniques for Feature Extraction In Speech Recognition System : A Comparative Study
Urmila Shrawankar, V M Thakare
TL;DR
Speech recognition must handle substantial variability in speech signals, motivating compact feature representations. The paper reviews feature-extraction approaches and reports method-specific advantages, including LPC over PLP in difficult conditions and IPS over MFCC in isolated-word recognition under clean and reverberant conditions.
Problem
Speech signals vary substantially, making direct recognition difficult and motivating feature representations that reduce variability.
Source: §III · p. 2Method
The paper reviews speech-feature extraction techniques, their strengths and weaknesses, and combined approaches including LPC–MFCC subsystems.
Source: §XV · p. 6, §XIX · p. 8Results
LPC has reduced word error rates versus PLP in difficult conditions, while IPS outperforms MFCC in isolated-word recognition under clean and reverberant conditions.
Source: §XVIII · p. 8, §XVII · p. 8Takeaways & Limitations
The conclusion calls for new hybrid methods aimed at better performance in robust speech recognition.
Source: §XIX · p. 8
Abstract
from arXiv · showhide
The time domain waveform of a speech signal carries all of the auditory information. From the phonological point of view, it little can be said on the basis of the waveform itself. However, past research in mathematics, acoustics, and speech technology have provided many methods for converting data that can be considered as information if interpreted correctly. In order to find some statistically relevant information from incoming data, it is important to have mechanisms for reducing the information of each segment in the audio signal into a relatively small number of parameters, or features. These features should describe each segment in such a characteristic way that other similar segments can be grouped together by comparing their features. There are enormous interesting and exceptional ways to describe the speech signal in terms of parameters. Though, they all have their strengths and weaknesses, we have presented some of the most used methods with their importance.
I. INTRODUCTION
Speech feature extraction reduces variable audio into parameters that can distinguish speech segments, while digital signal processing provides the front end for speech analysis and related tasks.
I. INTRODUCTION
- Speech characteristics are fairly stationary over short 5–100 ms periods, and changes over time reflect different speech sounds.The short-term amplitude spectrum is identified as the information basis for extracting speech features.Source: §I · p. 1
- Speaker, speaking-rate, content, and acoustic-condition variability makes feature analysis important to overall speech-recognition performance.Source: §I · p. 1
- The listed feature methods include LPC, LPCC, PLP, MFCC, FFT-based power-spectrum analysis, MEL, RASTA, and delta coefficients.Source: §I · p. 1
II. DIGITAL SIGNAL PROCESSING (DSP) TECHNIQUES
- Digital signal processing forms the speech-recognition system’s core and supports analysis, synthesis, coding, recognition, enhancement, voice modification, speaker recognition, and language identification.Source: §II · p. 2
III. FEATURE EXTRACTION
Feature extraction reduces speech-signal variability, and the section presents LPC as a model-based approach that predicts current samples from past samples.
III. FEATURE EXTRACTION
- Feature extraction reduces speech variability by discarding some source information and using short-term spectral analysis, including frequency warping on a Bark- or mel-like scale.The passage relates this analysis to the cochlea’s quasi-frequency analysis and its nonlinear frequency scale.Source: §III · p. 2
IV. LINEAR PREDICTIVE CODING
- LPC models speech production with an all-pole filter and estimates framewise predictor coefficients by minimizing squared differences between original and estimated speech.The described frames are normally 20 ms long, and the model includes gain G.Source: §IV · p. 2
- The LPC coefficients use order p, with p=10 for LPC-10 and p=18 for the improved algorithm; the parameters are computed using Levinson–Durbin recursion.The recursion is specified for the autocorrelation method (Deller et al., 2000).Source: §IV · p. 2
- LPC analysis also makes voiced/unvoiced and pitch decisions, but recognition uses cepstral coefficients rather than the high-variance predictor coefficients.Pitch, gain, and predictor coefficients vary across frames; the passage also lists bit rate, delay, complexity, and objective performance as evaluation parameters, and several LPC variants.Source: §IV · p. 3
V. MEL FREQUENCY CEPSTRAL COEFFICIENTS (MFCC)
MFCCs represent log filter-bank energies in the cepstral domain using a mel-frequency scale, with dynamic coefficients added to capture spectral change over time.
V. MEL FREQUENCY CEPSTRAL COEFFICIENTS (MFCC)
- MFCC is widely used, but its dependence on spectral form makes it noise-sensitive; periodicity-based methods are suggested as a possible remedy, although speech also has aperiodic content.About 20 coefficients are common in ASR, while 10–12 are often considered sufficient for speech coding.Source: §V · p. 3
- MFCC uses a mel-frequency approximation that is linear below 1 kHz and logarithmic above it, motivated by reduced auditory frequency selectivity at higher frequencies.The features are described as the cepstrum of log filter-bank energies.Source: §V · p. 3
- The MFCC computation applies a DCT to log filter-bank energies after the frame’s DFT spectrum is filtered.The described variables include the frame DFT, filter response, transform window size, and number of filters.Source: §V · p. 3
- First- and second-difference coefficients capture changes over time and are concatenated with static coefficients in the frame’s final feature output.Source: §V · pp. 3–4
VI. LFCC SPEECH FEATURES (LFCC-
LFCC omits mel warping and uses a linearly spaced filter bank, while HFCC-E separates filter bandwidth from spacing and derives bandwidth from ERB.
VI. LFCC SPEECH FEATURES (LFCC-
- LFCC-FB40 skips mel-frequency warping and uses 40 equal-width, equal-height linear filters, each 164 Hz wide, spanning 133–6857 Hz.Source: §VI · p. 4
- LFCC applies a DFT, filters the magnitude spectrum, logarithmically compresses filter-bank outputs, and applies a DCT; the described computation keeps the first 13 parameters.Source: §VI · p. 4
VII. HFCC-E
- HFCC-E decouples filter bandwidth from spacing, derives bandwidth from ERB rather than the mel scale, and retains mel-based filter center frequencies.The passage reports E-factor values of 4, 5, and 6.Source: §VI · p. 4
VIII. PURE FFT
The section describes FFT-based spectral analysis and contrasts a direct FFT representation with feature methods shaped by auditory-system knowledge.
VIII. PURE FFT
- Pure FFT features retain comparatively more speech information than methods using auditory-system knowledge, but extra information at higher frequencies may be less salient for recognition.The passage gives high sampling rates such as 44.1 kHz as an example.Source: §VIII · p. 4
IX. POWER SPECTRAL ANALYSIS (FFT)
- A speech power spectrum describes frequency content over time and is computed by first applying a discrete Fourier transform.Source: §IX · p. 4
- A real-valued speech signal permits an efficient real-point FFT, whose output contains magnitude and phase information.Source: §IX · p. 5
X. PERCEPTUAL LINEAR PREDICTION (PLP)
PLP modifies the short-term speech spectrum with psychophysically based transformations to model human hearing. Its described pipeline uses Bark-spaced filters and converts the transformed spectrum into cepstral coefficients.
X. PERCEPTUAL LINEAR PREDICTION (PLP)
- PLP modifies the short-term speech spectrum through psychophysically based transformations, aiming to describe human hearing more accurately than pure linear prediction.Source: §X · p. 5
XI. PLP SPEECH FEATURES (PLP-FB19)
- The PLP parameters use a Bark-spaced bank of 18 filters covering 0–5000 Hz; equal-loudness pre-emphasis is applied on the down-sampled signal, and intensity-loudness compression is performed.Source: §XI · p. 5
- After inverse DFT yields an equivalent autocorrelation function, autoregressive modeling and conversion to cepstral coefficients produce the PLP features.Source: §XI · p. 5
XII. MEL SCALE CEPSTRAL ANALYSIS (MEL)
Mel-scale analysis warps the short-term spectrum on the Mel scale and smooths it in the cepstral domain, unlike PLP’s Bark warping and all-pole modeling. RASTA filtering band-passes features to address channel distortions and short-term noise variation.
XII. MEL SCALE CEPSTRAL ANALYSIS (MEL)
- Mel-scale analysis uses Mel warping and direct inverse-DFT cepstral smoothing, whereas PLP uses Bark warping and an all-pole model to smooth the modified power spectrum.Source: §XII · p. 5
XIII. RELATIVE SPECTRA FILTERING (RASTA)
- RASTA band-passes feature coefficients: its high-pass portion alleviates convolutional channel noise, while low-pass filtering smooths frame-to-frame spectral changes.Source: §XIII · pp. 5–6
XIV. RASTA-PLP
- RASTA-PLP combines PLP’s speaker-difference-minimizing spectral warping with RASTA’s filtering of subband energy to smooth short-term noise and remove constant channel offsets [Herm90] [HermM94].Source: §XIV · p. 6
XV. COMBINED LPC & MFCC
The section describes a combined LPC–MFCC recognition system and a separate Matching Pursuit feature approach. The combined system accepts results confirmed by both subsystems, while Matching Pursuit selects dictionary atoms and summarizes their frequency and scale parameters.
XV. COMBINED LPC & MFCC
- The combined LPC–MFCC system trains separate neural-network subsystems and recognizes in parallel, confirming a result only when both agree; the cited author claims this decreases error rate [К.R. Aida–Zade, C. Ardil and S.S. Rustamov, 2006].Source: §XV · p. 6
- With 16 kHz sampling, 160-sample frame steps, and 400-sample frames, feature-vector dimension depends on frame dimensions and speech duration; scaling durations to a common scale is proposed as a partial remedy.Source: §XV · p. 6
XVI. MATCHING PURSUIT (MP)
- Matching Pursuit selects dictionary waveforms that best correlate with signal structure, stopping at a coefficient threshold or atom-count limit, or by signal-to-residual energy ratio.Source: §XVI · p. 7
- Matching Pursuit features record atom frequency and scale, then summarize each parameter by its mean and standard deviation; averaging is described as enhancing robustness.Source: §XVI · p. 7
XVII. INTEGRATED PHONEME SUBSPACE (IPS)
The section addresses speech recognition in reverberant environments, where long room responses complicate estimation of reverberant effects. It presents Integrated Phoneme Subspace (IPS) features and reports better isolated-word recognition than conventional MFCC under clean and reverberant conditions.
XVII. INTEGRATED PHONEME SUBSPACE (IPS)
- When a room impulse response exceeds the short-time DFT length, reverberation has additive and multiplicative effects in the power spectrum, complicating estimation in time or frequency.Source: §XVII · p. 7
- IPS is a data-driven feature method based on the logarithmic mel-frequency filter-bank domain, using linear, time-invariant transform matrices.The method is proposed for speech affected by reverberation.Source: §XVII · p. 7
- The authors report that IPS outperforms conventional MFCC in isolated-word recognition under clean and reverberant conditions, with optimal phoneme-subspace dimensions differing.The subspace dimensions were selected using an MDL-based experiment.Source: §XVII · p. 8
XVIII. CONCLUDING HIGHLIGHTS
The highlights report feature-method comparisons with outcomes that vary by condition and task, alongside differences in processing and representation. They describe LPC, FFT, MFCC, PLP, ICA, LDA, DCT, and Mel-scale methods.
XVIII. CONCLUDING HIGHLIGHTS
- The highlights state that noisy observed signals do not show good DCT performance without noise-suppression methods.Source: §XVIII · p. 8
- ICA applied in the time or time-frequency domain is reported to perform well on phoneme-recognition tasks.Source: §XVIII · p. 8
- LDA applied to time-frequency speech data is reported to outperform combined temporal and spectral linear discriminants for continuous-digit recognition.Source: §XVIII · p. 8
- Mel-scale cepstral analysis and PLP both modify short-term spectra using psychophysical transformations, but warp frequency with Mel and Bark scales, respectively.Source: §XVIII · p. 8
- PLP uses an all-pole model to smooth the modified power spectrum, whereas Mel-scale cepstral analysis uses cepstral smoothing.Source: §XVIII · p. 8
- LPC is reported to reduce word error rates in difficult conditions compared with PLP.Source: §XVIII · p. 8
- FFT-based features are described as linear and fast, retain information for later analysis, and make spectral noise effects readily interpretable.Source: §XVIII · p. 8
- PLP outperforms MFCC in specific conditions; MFCC reduces frequency information, models loudness with a logarithm, and is described as relatively fast and simple.PLP's RASTA filter is described as improving robustness to linear spectral distortions.Source: §XVIII · p. 8
XIX. CONCLUSION
The conclusion reviews feature-extraction methods and their pros and cons, noting that authors report improvements from some combined methods. It calls for new hybrid methods aimed at better performance in robust speech recognition.
XIX. CONCLUSION
- The authors call for new hybrid feature methods to achieve better performance in robust speech recognition, noting reported improvements from combinations of techniques.Source: §XIX · p. 8