1994年以前的speech coder的小结
1994年以前的speech coder的小结 -------------------------------Speech coding before 1994----------------------------------------
Speech quality is claissified into four general categories:
1)broadcast--above 64 kbits/s
2)Toll or network (200-3200Hz)--above 16 kbits/s
3)Communication--above 4.0 kbits/s
4)Synthetic--below 4.0 kbits/s
Object Mesurement:
1)signal-to-noise (SNR)
2)segmental SNR (SEGSNR)
3)articulation index
4)log spectral distance
5)the Euclidean distance
Subjective Mesurement:
Diagnostic Rhyme Test(DRT)--an intelligiblity measure where the subject's task is to recognize one of two possible words in a
set of rhyming pairs.
Diagnostic Aceptablitity Mesure(DAM)--based on results of test methods evaluating the quality of a communication system based
on teh acceptableility of speech as perceived by trained normative listener.
Mean Opinion Score(MOS)--involves 12 to 24 listeners who are instructed to rate phonetically balanced records according to a
five-level quality scale.
Waveform coders:
A.Scalar and vector quantization
1)Scalar Quantization
pulse-Code Modulation(PCM)--a memoryless proces that quantizes amplitudes by rounding off each sample to one of a set of
discrete values.
Adaptive PCM(APCM)--uniform quantizer. step size is estimated from past coded speech samples.(A 7-bit log quantizer for
speech achieves the performance of a 12-bit uniform quantizer)
Differential PCM(DPCM)--utilizes the redundancy in the speech waveform by exploiting the correlation between adjacent
samples.(better than PCM for rate at and below 32 kbits/s)
Adatvie DPCM(ADPCM)--the step size in DPCM is adaptive.
Delta Modulation(DM)--a sub-class of DPCM where the difference is encoded only with 1 bit.
Adaptive DM(ADM)-the step size in DM is adaptive.
standards:
G.721 CCITT standard(1988)---ADPCM 32-kbits/s
G.723 ---ADPCM 24 and 40 kbits/s (the performance of ADPCM degrades quickly for rates below 24 kbits/s)
2)vector quantization
--consists of an N-dimensional quantizer and a codebook. The incoming data are formed into a N-dimesional vector, then is mapped by quantizer to an entry in the codebook.
Full searched (F-VQ)--the codebook is fully searched for each incoming.
Tree-structured vector quantizer--the codebook is searched in "tree" way.(a degradation fo 1 db in the SNR compared with F- VQ)
Mulistep VQ--consist of a cascade of two or more quantizers, each one encoding the error or residual of the previous quantizer.(1 dB better in the SNR compared to F-VQ)
LBG--an iterative codebook design algorithm:inital guess for the codebook and then interative improvement by using a large
number of training vectors.
Gain/Shape VQ(GS-VQ)--normalizing the vectors fo the codebook and encoding the gain separately.
(0.7 db improvement compared to the F-VQ)
Adaptive codebooks(A-VQ)--the codebook is adaptive forward or backword.
B.sub-Band and Transform Coders
1)Sub-Band Coders(SBC)--the signal band is divided into frequency sub-bands using a bandk of bandpass filters.
standard:
AT&T voice store-and-forward standard--used for voice storage at 16 or 24 kbits/s and consits of five-band nonuniform tree-
structured QMF bank in conjunction with APCM coders. A silence compression alogrithm is also part of the standard.
CCITT G.772--for 7-kHz audio at 64 kbits/s for ISDN teleconferencing, based on two-band sub-band/ADPCM coder. Low frequency
suband is quantized at 48 kbits/s while the high-frequency sub-band is coded at 16 kbits/s.
2)Transform Coders(TC)--the transform components of a unitary transform are quantized at the transmitter and decoded and
inverse-transformed at the receiver. The bit-rate reduction lies in the fact that unitary transform tend to generate near-
uncorrelated transform components which can be coded independently.
several siscrete transform:
Discrete Cosine Transform(DCT) (near optimal)
Discrete Fourier Transform(DFT)
Walsh-Hadamard Transform(WHT)
kARHUNEN-lOEVE tRANSFORM(kLT) (optimal)
Adaptive transform coder(ATC)--encodeds the transform components using adaptive quantization and bit assignment rules.
//from here, I omit many examples....
Speech coding using sinusoidal analysis-synthesis models--relies on sinusoidal representations of the speech waveform.
A. speech Analysis-synthesis using the short-Time Fourier Transform
Time-varying spectral analysis can be performed using the short-time Fourier transform(STFT).
B.Sinusoidal Transform Coding(STC)--using unitary sinusoidal transforms implies that speech waveform si represented by a set of narrowband functions.(based on the fact that voiced speech is typically highly periodic and hence it can be represented by a constraned set of sinusoids)
C.The Multiband Excitation Coder(MBE)--relies on a model that treats the short-time speech spectrum as the product of an excitation spectrum and a vocal tract envelope
improved Multiband Excitation Coder(IMBE)--quantizeing the MBE model parameters.
standard:
Australian mobile staellite standard(AUSSAT) and the International Mobile Satellite(Inmarsat_M) employ IMBE that operates at 6.4 kbits/s
Vododer Methods.
--speech-specific coder.The performance of vocoders generally degrades for nonspeech signals. Rely on speech-specific
analysis-synthesis which is mostly based on the source-system model.
A.The Channel and the Formant Vocoder
relies on representing the speech spectrum as the product of vocal tract and excitation spectra.
B.Homomorphic Vocoders--vocal tract and the ecxitation log-magnutude spectra can be combined additively to produce the speech log-magnutude spectrum.
C. Linear-Predictive Vocoders(LPC)--predict the sample by uisng a linear comibation of last samples.
a)The calssical two-state excitation model
standard:
LPC-10--usins a 10th-order predictor to estimate the vocal-tract parameters.
b)mixed excitation model
LPC combined with others..?
C)Residual excited linear prediction(RELP)--encodes the residual of LPC, and allot more bits for the perceptually important components.(the quality of RELP coder at rates above 4.8 kbits/s is higher than the analogous two stated LPC)
Analysis-by Synthesis Linear Predictive Coders
--system parameters are determined by linear prediction and the excitation sequence is determinded by closed-loop or analysis-by-synthesis optimaization
A.Multipulse-Excited Linear Prediction(MPLP)--forms an excitation sequence which consists of multiple nonuniformly spaced pulses. Both amplitude and locations of the pluses are determined one pluse at a time such that the weighted mean squared error is minimized.(produced good quality speech at rates as low as 10 kbits/s)
B.Regular Pulse Excitation Coder(RPE)--the pulses in the RPE coder are uniformly spaced and therefor their position are determined by specifying the location k of the first pulse within the frame and the spacing between nonzero pulse.
C.Code Excited Linear Prediction(CELP)--encodes the excitation using a codebook of Guassian sequences. THe book contains 1024 vectors and each vector si 40 sampels(5 ms) long. A gain factor scales the excitation vector and the excitation samples are filter by the long- and short-term synthesis filters. The optimum vecotor is selected such that the perceptually weighted MSE is minimized.
