Skip to content
Back to search results

Linear Predictive Coding

Introduced by
Freddy Burg, John Makhoul
Year
1975
Type
Predictive speech coding and analysis
Family
Signal processing, Speech modeling
Font size

What does it do?

Linear Predictive Coding (LPC) is a powerful signal processing technique for modeling and compressing speech by assuming each speech sample can be predicted from a linear combination of previous samples.

Pioneered by Burg and Makhoul in the 1970s, LPC is the foundation of speech codecs, speech synthesis, speaker recognition, and speech enhancement. LPC exploits the time-correlated structure of speech to achieve high compression ratios and enable efficient parameter extraction.

The intuition behind the method

Speech sounds repeat patterns: vibrating vocal cords create pitch periodicity, and the vocal tract shape filters this vibration into phonemes. LPC captures these regularities by predicting each speech sample from the previous few samples (typically 8–16). If the prediction is accurate, only the prediction error (residual) needs to be transmitted, achieving compression. LPC models the vocal tract as a linear system, making it highly efficient for speech.

History

Burg introduced LPC in the context of geophysics and spectral estimation in the 1960s. Makhoul brought LPC to speech processing in 1975, founding modern speech coding. Government research (NSA, DARPA) accelerated LPC codec development in the 1980s. CELP (Code-Excited Linear Prediction) combined LPC with vector quantization and became standard for mobile speech codecs (GSM, EVRC). Modern codecs (AMR, OPUS) still use LPC principles with extensions.

When to use it

Use LPC for speech compression and coding (telephony, VoIP), speaker verification and recognition, voice quality assessment, speech synthesis (TTS), and speech enhancement. LPC is particularly effective at low bit rates (< 4 kbps) and for clean speech. For wideband audio (music) or very low bit rates (< 1 kbps), more sophisticated codecs may be preferred.

Limitations and considerations

  • Designed for speech; music and general audio require higher model order or alternative methods.
  • Assumes linear vocal tract and steady-state sound production; transients (plosives, fricatives) are not modeled accurately.
  • Pitch estimation is challenging, especially for female speakers and noisy conditions; pitch errors propagate into synthesis.
  • Low-frequency residual (>1 kHz detail) is often quantized aggressively; speech intelligibility depends on accurate high-frequency representation.
  • Individual speaker variation (accent, voice quality) not captured in basic LPC; speaker normalization requires additional processing.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Highly efficient compression: speech can be represented with 12–24 parameters per 20 ms frame (~600 bits/sec), enabling 50:1+ compression.
  • Represents vocal tract physics: LPC coefficients correspond to vocal tract resonances (formants), making them interpretable and robust to noise.
  • Fast computation: LPC coefficients can be estimated in real time on embedded hardware using Durbin recursion or lattice algorithms.
  • Provides clear separation of excitation (pitch, gain) and filter (vocal tract); enables speech modification (pitch shifting, time stretching, voice conversion).
  • Well-established and standardized: LPC is used in military and cellular speech codecs; decades of refinement ensure reliability.

Applications

  • Speech coding and telephony: fundamental technique in cellular (GSM, 3G, 4G) and internet telephony (VoIP, Skype) codecs.
  • Speaker recognition and verification: LPC cepstral coefficients (cepstral analysis of LPC) are standard for speaker identification systems.
  • Voice conversion and speech synthesis: LPC enables morphing voice characteristics and generating synthetic speech.
  • Speech enhancement and denoising: LPC can model speech; subtraction of noise estimate from LPC residual improves SNR.
  • Forensic audio analysis: LPC features reveal speaker characteristics and emotional state in legal audio recordings.
  • Music instrument modeling: LPC variants (formant synthesis) approximate string and wind instrument acoustics.

How it works

How it works

1

Record or digitize a speech signal at 8–16 kHz (speech bandwidth); frame the signal into overlapping windows (typically 20 ms, 50% overlap).

2

Apply pre-emphasis filter (high-pass) to flatten the speech spectrum; apply Hamming window to reduce edge effects.

3

Estimate the autocorrelation function of the windowed signal; compute LPC coefficients by solving the normal equations (Yule-Walker, Durbin, or lattice method).

4

Represent each speech frame as LPC coefficients (typically 10–12 scalars) plus a gain and fundamental frequency (pitch). Synthesize speech by filtering white noise or a pitch impulse through the LPC filter (all-pole filter).

5

Extract speech features: LPC coefficients reflect vocal tract shape; pitch reflects voicing; gain reflects loudness. Use features for speaker recognition, speech coding, or synthesis. Compute spectral envelope by evaluating LPC filter frequency response.

Common pitfalls

  • Using insufficient LPC order (< 8): low-order models fail to capture formants and produce muffled, synthetic speech.
  • Over-aggressive residual quantization: LPC coefficients may be accurate, but quantized residuals degrade quality.
  • Neglecting pre-emphasis; without high-pass filtering, low-frequency energy dominates and LPC fitting emphasizes incorrect aspects.
  • Assuming constant frame rate; voice activity detection and frame dropping at silence improve codec efficiency but require careful implementation.
  • Ignoring listener subjective quality; MOS (Mean Opinion Score) and listening tests are critical for speech codec tuning.

Frequently asked questions

What is the optimal LPC order for speech?

Typical LPC order is 10–12 for speech at 8 kHz and 16–20 for wideband (16 kHz) speech. A rule of thumb: order ≈ 2 + fs/1000, where fs is sampling rate in Hz. Higher order captures more detail but requires more coefficients and risks overfitting. Listening tests often guide the choice.

How do I extract speaker identity from LPC coefficients?

Use cepstral analysis of LPC (inverse FFT of log LPC spectrum) to extract formant positions and fine spectral structure characteristic of a speaker. These cepstral features (typically 12–13 coefficients) are more discriminative than raw LPC and are used in speaker recognition systems.

Why is pitch estimation hard in LPC?

LPC models the vocal tract filter; pitch is determined by excitation (vibration rate), not the filter. Pitch estimation requires separate autocorrelation analysis or harmonic analysis of the residual signal. Noisy or weak pitch signals (female speakers, whisper) are particularly challenging; errors in pitch estimation directly degrade synthesis quality.

Can LPC compress music and general audio?

LPC is designed for speech and performs poorly on music because music has rich harmonic content and complex spectral structure. Higher LPC order (50+) can model music roughly, but dedicated audio codecs (MP3, AAC) achieve better quality and compression. LPC remains the standard for speech-only applications.

How is LPC related to CELP codecs?

CELP (Code-Excited Linear Prediction) enhances LPC by replacing a simple pitch excitation with a vector-quantized code from a codebook. This improves quality at low bit rates (< 8 kbps). CELP is the basis of many modern cellular speech codecs (GSM-EFR, EVRC, CDMA); it's essentially LPC plus sophisticated excitation modeling.

Related methods

Sources

  1. Makhoul, J. (1975). Linear prediction: A tutorial review. Proceedings of the IEEE, 63(4), 561–580.

    Open DOI reference
    Copied.
  2. Rabiner, L. R., & Schafer, R. W. (1978). Digital Processing of Speech Signals. Prentice-Hall.

    ISBN: 978-0132136029
    Copied.
  3. Haykin, S. (2002). Adaptive Filter Theory (4th ed.). Prentice Hall.

    ISBN: 978-0130901262
    Copied.