Data processing: speech signal processing – linguistics – language – Speech signal processing – For storage or transmission
Reexamination Certificate
2001-07-02
2002-10-22
Dorvil, Richemond (Department: 2641)
Data processing: speech signal processing, linguistics, language
Speech signal processing
For storage or transmission
C704S220000, C704S200100
Reexamination Certificate
active
06470312
ABSTRACT:
BACKGROUND OF THE INVENTION
(1) Field of the Invention
The present invention relates to a speech coding apparatus, a speech processing apparatus, and a speech processing method. More particularly, the present invention relates to a speech coding apparatus which encodes speech signals into lower bitrate signals (e.g., less than 4 kb/s) by applying Analysis-by-Synthesis (AbS) vector quantization techniques to source speech signals containing a plurality of periodic components within a fixed time interval, based on an appropriate speech production model. The present invention also relates to a speech processing apparatus and a speech processing method which performs speech coding with AbS vector quantization techniques, based on a speech production model.
(2) Description of the Related Art
The code-excited linear prediction (CELP) method has been known as a speech coding technique which encodes telephone voice signals with a spectrum ranging from 0.3 to 3.4 kHz into a bitstream with a rate of 4 to 16 kb/s. CELP is widely used in digital mobile communications systems and enterprise communications systems.
What CELP coders actually transmit are: linear predictive coding (LPC) coefficients representing resonance characteristics of the human vocal tract, and parameters representing excitation signals (sound source data) consisting of periodic pitch components and noise components. The CELP algorithm uses an LPC synthesis filter H(z) of equation (1) as a model of the human vocal tract, assuming that the input signal to that filter (sound source signal) can be divided into its periodic pitch components and noise components. The former components represent the periodicity of voices, while the latter the randomness.
H
(
z
)
=
1
1
-
∑
i
=
1
p
α
i
·
z
-
i
(
1
)
Subsequently, the filter coefficients characterizing the LPC synthesis filter, as well as the pitch interval components and noise components of the excitation signal, are extracted and quantized. Here, data compression is accomplished by sending out the quantized data, i.e., quantization indices.
FIG. 18
depicts the CELP coding algorithm. Suppose that a source voice signal Sn is entered to an LPC analyzing means
21
. Using an all-pole filter characterized by equation (1) representing the human vocal tract model, the LPC analyzing means
21
calculates the coefficient &agr;
i
(i=1 to p) of that all-pole filter, where p represents its order. Typically, the filter's order p ranges from 10 to 12 for phone-quality speech signals, while 16 to 20 for wideband speech signals.
The LPC filter coefficients are then quantized with scalar quantization or vector quantization techniques (quantizer not shown in FIG.
18
). The resultant quantization indices are transmitted to the decoding end. The excitation signal is also quantized. For the quantization of its pitch interval components, the CELP algorithm employs an adaptive codebook Ba recording the past sound source signal series. For the quantization of noise components, the algorithm provides a noise codebook Bn storing various noise signal patterns.
The codebooks Ba and Bn are used in the A-b-s vector quantization process as follows. The process begins with a variable-gain multiplication of code vectors read out of the codebooks Ba and Bn. This operation is executed by multipliers
22
a
and
22
b.
The sum of the outputs of the multipliers
22
a
and
22
b
are calculated by an adder
23
and supplied to an LPC synthesis filter
24
, whose response is defined by the LPC filter coefficients. With its filtering algorithm, the LPC synthesis filter
24
reproduces a signal Sn*. This reproduced speech signal Sn* is then subjected to an arithmetic operator
26
for the calculation of its error en with respect to the source speech signal Sn.
An error power evaluation means
25
evaluates the error en for every possible combination of code vectors read out of the two codebooks Ba and Bn, changing the positions of selection switches SWa and SWb from one to another. Through the error evaluation, the error power evaluation means
25
obtains one particular combination of code vectors which exhibit the smallest error value among others. This combination is referred to herein as the “optimal code vector pair,” and the gain corresponding to the pair is referred to as the “optimal gain.” Finally, a quantizer (not shown) quantizes the obtained optimal code vector pair and optimal gain, thereby yielding quantization indices.
That is, the coder produces quantization indices of LPC filter coefficients, of optimal code vectors, and of optimal gains. The quantization indices of the optimal code vectors actually include: those of the code vectors selected from the noise codebook Bn, and those of what will be explained later as “lag,” i.e., a parameter used in extracting optimal vectors from the adaptive codebook Ba. Those quantization indices are transmitted to the decoding end.
The decoder obtains LPC filter coefficients, optimal code vector, and optimal gain by decoding the data received from the encoder. Employing the same codebooks Ba and Bn and LPC synthesis filter as those used at the encoding end, the decoder reproduces the original speech signal.
As described above, the CELP algorithm accomplishes speech compression by establishing a speech production process model and transmitting quantized characteristic parameters of that model. Since the characteristics of human voices exhibit little variation within a short time, e.g., 5 to 10 msec, the CELP algorithm updates vocal tract parameters and excitation parameters only at as short intervals as 5 to 10 msec. Such short time segments are referred to as “frames.” This method permits the CELP coders to provide coded speech signals without quality deterioration at reduced bitrates as low as 5 to 6 kb/s.
The above-described conventional speech coding algorithm, however, cannot reduce the bitrate further for the following reason. For bitrates of 4 kb/s or lower, the conventional algorithm requires that the frame length be elongated to more than 10 ms. This means that a single frame of a source speech signal is likely to contain two or more different pitch components, introducing quality deterioration to the resultant coded speech signal.
In other words, the conventional CELP algorithm is weak in modeling the periodicity of a speech signal within a single frame because the periodicity of output signals contained in the adaptive codebook Ba is strictly confined to one component per frame. For this reason, the conventional algorithm is unable to capture the periodicity of a source speech signal precisely enough to provide high coding efficiency in the cases where one frame contains a plurality of periodic pitch components, which limits its coding efficiency.
SUMMARY OF THE INVENTION
In view of the foregoing, it is an object of the present invention to provide a speech coding apparatus which encodes given speech signals in an optimal way.
It is another object of the present invention to provide a speech processing apparatus which performs optimal speech processing according to a given source speech signal, so that the signal will be reproduced with high quality at the receiving end.
It is still another object of the present invention to provide a speech processing method which performs speech processing in an optimal way for a given source speech signal, so that the signal will be reproduced with high quality at the receiving end.
To accomplish the above first object, according to the present invention, there is provided a speech coding apparatus. This speech coding apparatus performs speech coding based on a speech production model, in which a given speech signal Sn is divided into fixed-length segments. This speech coding apparatus comprises the following elements: an adaptive codebook Ba which stores a series of signal vectors of a past speech signal; a vector extraction means for extracting a signal vector and neighboring vectors adjacent thereto from the adaptive codebook, the signal vector being stored at a distance
Ota Yasuji
Suzuki Masanao
Tsuchinaga Yoshiteru
Dorvil Richemond
Fujitsu Limited
Katten Muchin Zavis & Rosenman
LandOfFree
Speech coding apparatus, speech processing apparatus, and... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Speech coding apparatus, speech processing apparatus, and..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Speech coding apparatus, speech processing apparatus, and... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2957229