Data processing: speech signal processing – linguistics – language – Speech signal processing – For storage or transmission
Reexamination Certificate
2000-10-18
2003-09-02
Knepper, David D. (Department: 2645)
Data processing: speech signal processing, linguistics, language
Speech signal processing
For storage or transmission
C704S206000, C704S210000, C704S225000
Reexamination Certificate
active
06615169
ABSTRACT:
FIELD OF THE INVENTION
The present invention generally relates to the field of coding and decoding synthesized speech and, more particularly, to an adaptive multi-rate wideband speech codec.
BACKGROUND OF THE INVENTION
Many methods of coding speech today are based upon linear predictive (LP) coding, which extracts perceptually significant features of a speech signal directly from a time waveform rather than from a frequency spectra of the speech signal (as does what is called a channel vocoder or what is called a formant vocoder). In LP coding, a speech waveform is first analyzed (LP analysis) to determine a time-varying model of the vocal tract excitation that caused the speech signal, and also a transfer function. A decoder (in a receiving terminal in case the coded speech signal is telecommunicated) then recreates the original speech using a synthesizer (for performing LP synthesis) that passes the excitation through a parameterized system that models the vocal tract. The parameters of the vocal tract model and the excitation of the model are both periodically updated to adapt to corresponding changes that occurred in the speaker as the speaker produced the speech signal. Between updates, i.e. during any specification interval, however, the excitation and parameters of the system are held constant, and so the process executed by the model is a linear time-invariant process. The overall coding and-decoding (distributed) system is called a codec.
In a codec using LP coding to generate speech, the decoder needs the coder to provide three inputs: a pitch period if the excitation is voiced, a gain factor and predictor coefficients. (In some codecs, the nature of the excitation, i.e. whether it is voiced or unvoiced, is also provided, but is not normally needed in case of an Algebraic Code Excited Linear Predictive (ACELP) codec, for example. LP coding is predictive in that it uses prediction parameters based on the actual input segments of the speech waveform (during a specification interval) to which the parameters are applied, in a process of forward estimation.
Basic LP coding and decoding can be used to digitally communicate speech with a relatively low data rate, but it produces synthetic sounding speech because of its using a very simple system of excitation. A so-called Code Excited Linear Predictive (CELP) codec is an enhanced excitation codec. It is based on “residual” encoding. The modeling of the vocal tract is in terms of digital filters whose parameters are encoded in the compressed speech. These filters are driven, i.e. “excited,” by a signal that represents the vibration of the original speaker's vocal cords. A residual of an audio speech signal is the (original) audio speech signal less the digitally filtered audio speech signal. A CELP codec encodes the residual and uses it as a basis for excitation, in what is known as “residual pulse excitation.” However, instead of encoding the residual waveforms on a sample-by-sample basis, CELP uses a waveform template selected from a predetermined set of waveform templates in order to represent a block of residual samples. A codeword is determined by the coder and provided to the decoder, which then uses the codeword to select a residual sequence to represent the original residual samples.
According to the Nyquist theorem, a speech signal with a sampling rate F
s
can represent a frequency band from 0 to 0.5 F
s
. Nowadays, most speech codecs (coders-decoders) use a sampling rate of 8 kHz. If the sampling rate is increased from 8 kHz, naturalness of speech improves because higher frequencies can be represented. Today, the sampling rate of the speech signal is usually 8 kHz, but mobile telephone stations are being developed that will use a sampling rate of 16 kHz. According to the Nyquist theorem, a sampling rate of 16 kHz can represent speech in the frequency band 0-8 kHz. The sampled speech is then coded for communication by a transmitter, and then decoded by a receiver. Speech coding of speech sampled using a sampling rate of 16 kHz is called wideband speech coding.
When the sampling rate of speech is increased, coding complexity also increases. With some algorithms, as the sampling rate increases, coding complexity can even increase exponentially. Therefore, coding complexity is often a limiting factor in determining an algorithm for wideband speech coding. This is especially true, for example, with mobile telephone stations where power consumption, available processing power, and memory requirements critically affect the applicability of algorithms.
In the prior-art wideband codec, as shown in
FIG. 1
, a pre-processing stage is used to low-pass filter and down-sample the input speech signal from the original sampling frequency of 16 kHz to 12.8 kHz. The down-sampled signal is then decimated so that the number of samples of 320 within a 20 ms period are reduced to 256. The down-sampled and decimated signal, with an effective frequency bandwidth of 0 to 6.4 kHz, is encoded using an Analysis-by-Synthesis (A-b-S) loop to extract LPC, pitch and excitation parameters, which are quantized into an encoded bit stream to be transmitted to the receiving end for decoding. In the A-b-S loop, a locally synthesized signal is further up sampled and interpolated to meet the original sample frequency. After the encoding process, the frequency band of 6.4 kHz to 8.0 kHz is empty. The wideband codec generates random noise on this empty frequency range and colors the random noise with LPC parameters by synthesis filtering as described below.
The random noise is first scaled according to
e
scaled
=sqrt[{exc
T
(
n
)
exc
(
n
)}/{
e
T
(
n
)
e
(
n
)}]
e
(
n
) (1)
where e(n) represents the random noise and exc(n) denotes the LPC excitation. The superscript T denotes the transpose of a vector. The scaled random noise is filtered using the coloring LPC synthesis filter and a 6.0-7.0 kHz band pass filter. This colored, high-frequency component is further scaled using the information about the spectral tilt of the synthesized signal. The spectral tilt is estimated by calculating the first autocorrelation coefficient, r, using the following equation:
r={s
T
(
i
)
s
(
i
−1)}/{
s
T
(
i
)
s
(
i
)} (2)
where s(i) is the synthesized speech signal. Accordingly, the estimated gain f
est
is determined from
f
est
=1.0
−r
(3)
with the limitation 0.2≦f
est
≦1.0.
At the receiving end, after the core decoding process, the synthesized signal is further post-processed to generate the actual output by up-sampling the signal to meet the input signal sampling frequency. Because the high frequency noise level is estimated based on the LPC parameters obtained from the lower frequency band and the spectral tilt of the synthesized signal, the scaling and coloring of the random noise can be carried out in the encoder end or the decoder end.
In the prior-art codec, the high frequency noise level is estimated based on the base layer signal level and spectral tilt. As such, the high frequency components in the synthesized signal are filtered away. Hence, the noise level does not correspond to the actual input signal characteristics in the 6.4-8.0 kHz frequency range. Thus, the prior-art codec does not provide a high quality synthesized signal.
It is advantageous and desirable to provide a method and a system capable of providing a high quality synthesized signal taking into consideration the actual input signal characteristics in the high frequency range.
SUMMARY OF THE INVENTION
It is a primary objective of the present invention to improve the quality of synthesized speech in a distributed speech processing system. This objective can be achieved by using the input signal characteristics of the high frequency components in the original speech signal in the 6.0 to 7.0 kHz frequency range, for example, to determine the scaling factor of a colored, high-pass filtered artificial signal in synthesizing the higher frequency components of the synthesized
Mikkola Hannu
Ojala Pasi
Rotola-Pukkila Jani
Vainio Janne
Knepper David D.
Nokia Corporation
Ware Fressola Van Der Sluys & Adolphson LLP
LandOfFree
High frequency enhancement layer coding in wideband speech... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with High frequency enhancement layer coding in wideband speech..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and High frequency enhancement layer coding in wideband speech... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3090429