CELP - Code Excited Linear Prediction

By: Preethi Ramkumar

Meaning of CELP – “Code Excited Linear Prediction”, is a lossy compression algorithm used for low bit rate (e.g.,4800 bps in U.S. Fed-Std-1016) speech coding. CELP coding, at 4800 bps, breaks the performance barrier of most Government standards, providing Consortium ratings of "very good" intelligibility and "excellent" quality, comparable to 32,000 bps (CVSD).

 Like all vector quantization techniques, CELP coding is a frame-oriented technique that breaks a sampled input signal into blocks of samples (i.e., vectors) that are processed as one unit. CELP coding is based on analysis-by-synthesis search procedures, perceptually weighted vector quantization (VQ), and linear prediction (LP). A 10th order LP filter is used to model the speech signal's short-term spectrum, or formant structure. Long-term signal periodicity, or ` 1111111, is modeled by an adaptive code book VQ. The residual from the short-term LP and pitch VQ is vector quantized using a fixed stochastic code book. The optimal scaled excitation vectors from the adaptive and stochastic code books are selected by minimizing a time varying, perceptually weighted distortion measure that improves subjective speech quality by exploiting masking properties of human hearing.

The CELP coder's computational requirements are dominated by the two code book searches. The computational complexity and speech quality of the coder depend upon the search sizes of the code books. Any subset of either code book can be searched to fit processor constraints, at the expense of speech quality.

Other Related Definitions:

“…CELP is a relatively new digital voice algorithm that permits high quality | | digital voice communications over normal dial-up telephone lines.” [Totse]

“…A study of a new type of speech coding algorithm that in many ways fills a void in the capabilities of present generation speech coders. Previously studied waveform coding schemes tend to have a knee in the speech quality / bit rate curve such that for rates substantially below 10 kb/s, the quality of the reproduced speech falls off rapidly. At rates below 3 kb/s, vocoders which produce synthetic quality speech are the only alternative. The Code Excited Linear Prediction (CELP) coding scheme studied here tends to fill in the gap between waveform coders and vocoders at rates around 5 kb/s.” [P. Kabal]

“…CELP (Code Excited Linear Prediction) - By using predictive techniques such as those used in ADPCM combined with more advanced prediction techniques that predict some of the periodicity (or long-term correlations) we could quantize the predictor error with just a few bits per sample. However, for bit rates at 8 kb/s and lower, we only have a fraction of bits per sample available. To solve this dilemma we use Vector Quantization (VQ) or codebook techniques. By having a codebook both at the encoder and decoder side only the indices have to be transmitted. In CELP coders codebooks are used to quantize the predictor error. At the decoder, the corresponding codebook vector is looked up and filtered through the inverse predictor filters to produce the reconstructed output speech. Further improvement can be achieved by applying a postfilter.” [Lucenttls]

“…A new speech coding algorithm, pole-zero CELP (code excited linear prediction), is proposed in which the short-time spectral envelop of the speech wave is modeled with a pole-zero filter. Hence, zeros which occur in the speech spectrum (e.g., during nasalized speech sounds) are more accurately modeled than with the traditional all-pole filter. A linear prediction analysis of the speech signal to obtain a pole-zero filter representation of the spectral envelope results in a set of nonlinear simultaneous equations. Previously, complex iteration schemes have been used to solve for the filter coefficients from these equations. However, knowledge of the excitation in the pole-zero CELP coder leads to a linear but suboptimum solution for the filter coefficients. As expected, it shows improved spectral matching in regions of anti-resonant nasalized sounds. In addition the high frequency portion of the speech signal is matched better. Informal listening tests shown a slight preference for speech coded with the zero filter included and a perceptually weighted error criterion. ” [ IEEE ]

“…The essence of CELP techniques, which is an analysis-by-synthesis approach to codebook search, is retained in LD-CELP. The LD-CELP however, uses backward adaptation of predictors and gain to achieve an algorithmic delay of 0.625 ms. Only the index to the excitation codebook is transmitted. The predictor coefficients are updated through LPC analysis of previously quantized speech. The excitation again is updated by using the gain information embedded in the previously quantized excitation. The block size for the excitation vector and gain adaptation is five samples only. A perceptual weighting filter is updated using LPC analysis of the unquantized speech. ” [Lucent Technologies]

Related Links:

CMU - CELP: FS-1016 Code Excited Linear Prediction coder
Totse - Info on CELP: Code Excited Linear Prediction
Reports - Code Excited Linear Prediction Coding of Speech at 4.8 kb/s
Lucenttls - CODE EXCITED LINEAR PREDICTION (CELP)
FAQ - Information on 2400 bps LPC and 4800 bps CELP speech coders
CodingCompression - Code excited linear prediction (CELP) voice coder
Otolith - Linear Predictive Coding (LPC)
Cogs - Code Excited Linear Predictor (CELP
Purdue - Speech Compression (Coding)
Arun Kumar and Allen Gersho, Fellow, IEEE - LD-CELP Speech Coding with Nonlinear Prediction
Digital source coding of speech signals - Speech coding algorithms
SpeechAnalysis - CELP coders
Digital Voice Systems - Voice Coding Overview

Technical Resources:

Vocal - GSM 06.60 Enhanced Full Rate (EFR) Vocoder Algebraic-Code-Excited Linear Predictive (ACELP)

Products and Solutions:

Audiocoding
Bell-labs
Bdti
Citeseer
Cpupedia
Eas
Javvin

Blogs, News, feeds…

http://www.hydrogenaudio.org/forums/lofiversion/index.php/t22313.html
http://www.ee.cityu.edu.hk/~cfchan/demo.html
http://www.isr.umd.edu/ISR/accomplishments/007_CELP/

Books:

Telecommunications : analog to digital conversion of radio voice by 4,800 bit/second code excited linear prediction (CELP) by SuDoc GS

See Also:

CELP Resources

 

(C) Copyright Birds-Eye.Net, All rights reserved.
It is against the law to reproduce this content or any portion of it in any form without the explicit written permission of Birds-Eye Network Services, LLC. Federal copyright law (17 USC 504) makes it illegal, punishable with fines up to $100,000 per violation plus attorney's fees.