Filterbank and PLP Features
Filterbank and Perceptual Linear Prediction (PLP) features are ways of summarizing a speech signal into compact, perceptually meaningful numbers that machine learning models can use.
Overview
They matter because they let speech recognizers focus on the parts of sound humans actually hear, discarding irrelevant detail.
Deep Dive
To turn raw audio into features, the signal is split into short frames and passed through a bank of overlapping filters spaced on the mel scale, which mimics the ear's nonlinear frequency sensitivity. Summing the energy in each filter produces log-mel filterbank features, the dominant input for modern deep speech models. PLP, developed by Hynek Hermansky, adds more psychoacoustics: it applies the bark-scale critical bands, an equal-loudness curve weighting frequencies as the ear does, and a cube-root intensity-to-loudness compression, then fits an all-pole (linear prediction) model to smooth the spectrum. The result is a low-dimensional representation robust to speaker and channel differences. MFCCs are a close cousin that adds a cosine transform to decorrelate the filterbank outputs.
Technical Insight
The key idea is perceptual warping: linear hertz is remapped to mel or bark scales so filters are narrow at low frequencies and wide at high ones, matching cochlear resolution. PLP's equal-loudness preemphasis and cube-root compression model how the ear's loudness perception is nonlinear. The final linear-prediction step fits a smooth spectral envelope, capturing vocal-tract shape while suppressing pitch harmonics that vary between speakers.
Strategic Impact
Access and reach
It improves accessibility through transcription, narration, and voice interfaces.
Cost and budget
Media teams can ship polished audio faster with smaller budgets.
Speed and scale
Customer-facing systems can process spoken interactions at larger scale.
The Future of Filterbank and PLP Features
Deep neural networks increasingly prefer raw log-mel filterbanks over heavily engineered PLP or MFCC features, because the network learns its own transformations better than hand-designed decorrelation. The frontier is learnable front ends like SincNet and wav2vec that operate on raw waveforms. Still, mel filterbanks remain ubiquitous as a stable, low-cost input, and the perceptual principles behind PLP continue to inform how engineers design and interpret these learned representations.
Real-World Implementation
Computing 40 log-mel filterbank features per frame as the input to a speech-to-text neural network
Using PLP features in noise-robust voice command systems for cars
Speaker recognition pipelines that rely on perceptually warped spectral features
Keyword spotting on low-power devices where compact filterbank features reduce computation
Risks & Guardrails
Voice misuse and impersonation risks increase when consent is missing.
Accuracy can drop across accents, dialects, or noisy environments.
Synthetic audio can be mistaken for authentic speech without clear labeling.
Implementation Roadmap
Obtain explicit consent for voice capture, cloning, and reuse.
Test quality across diverse speakers and background conditions.
Define when a human must review or approve outputs.
Label synthetic audio and keep provenance records for accountability.
Keep Exploring
Free newsletter
Keep up with AI in 3 minutes a day
One short email each weekday with the three AI stories that actually matter. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Filterbank and PLP Features quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Next guide
Linear Probing and Frozen Feature Evaluation
Frequently asked questions
What is Filterbank and PLP Features?
Filterbank and Perceptual Linear Prediction (PLP) features are ways of summarizing a speech signal into compact, perceptually meaningful numbers that machine learning models can use. They matter because they let speech recognizers focus on the parts of sound humans actually hear, discarding irrelevant detail.
Why are speech filterbanks spaced on the mel or bark scale instead of linear hertz?
The mel and bark scales approximate human cochlear resolution, which is finer at low frequencies and coarser at high frequencies.
What does the linear prediction step in PLP accomplish?
PLP fits an all-pole model to produce a smooth spectral envelope, capturing vocal-tract characteristics while suppressing speaker-dependent pitch harmonics.
Which feature type is the dominant input for modern deep speech recognition models?
Log-mel filterbank features are the standard, compact, perceptually meaningful input for today's deep speech models.
How do MFCCs differ from raw log-mel filterbank features?
MFCCs apply a discrete cosine transform on top of log-mel filterbank energies to decorrelate them into compact coefficients.
What perceptual element does PLP include that basic mel filterbanks do not?
PLP adds an equal-loudness preemphasis and cube-root intensity-to-loudness compression to better model human loudness perception.