Speech processing system and method
A speech processing method, comprising: receiving a speech input which comprises a sequence of feature vectors; determining the likelihood of a sequence of words arising from the sequence of feature vectors using an acoustic model and a language model, comprising: providing an acoustic model for performing speech recognition on an input signal which comprises a sequence of feature vectors, said model having a plurality of model parameters relating to the probability distribution of a word or part thereof being related to a feature vector, wherein said speech input is a mismatched speech input which is received from a speaker in an environment which is not matched to the speaker or environment under which the acoustic model was trained; and adapting the acoustic model to the mismatched speech input, the speech processing method further comprising determining the likelihood of a sequence of features occurring in a given language using a language model; and combining the likelihoods determined by the acoustic model and the language model and outputting a sequence of words identified from said speech input signal, wherein adapting the acoustic model to the mismatched speaker input comprises: relating speech from the mismatched speaker input to the speech used to train the acoustic model using: a mismatch function f for primarily modelling differences between the environment of the speaker and the environment under which the acoustic model was trained; and a speaker transform F for primarily modelling differences between the speaker of the mismatched speaker input, such that: y=f ( F ( x,v ), u ) where y represents the speech from the mismatched speaker input, x is the speech used to train the acoustic model, u represents at least one parameter for modelling changes in the environment and v represents at least one parameter used for mapping differences between speakers; and jointly estimating u and v.
1. A speech processing method, comprising: receiving a speech input which comprises a sequence of feature vectors;
determining a likelihood of a sequence of words arising from the sequence of feature vectors using an acoustic model and a language model, comprising:
providing an acoustic model for performing speech recognition on an input signal which comprises a sequence of feature vectors, said model having a plurality of model parameters relating to the probability distribution of a word or part thereof being related to a feature vector, wherein said speech input is a mismatched speech input which is received from a speaker in an environment which is not matched to the speaker or environment under which the acoustic model was trained; and
adapting the acoustic model to the mismatched speech input,
the speech processing method further comprising determining a likelihood of a sequence of features occurring in a given language using a language model; and
combining the likelihoods determined by the acoustic model and the language model and outputting a sequence of words identified from said speech input signal,
wherein adapting the acoustic model to the mismatched speaker input comprises:
relating speech from the mismatched speaker input to the speech used to train the acoustic model using: a mismatch function f for primarily modeling differences between the environment of the speaker and the environment under which the acoustic model was trained;
and a speaker transform F for primarily modeling differences between the speaker of the mismatched speaker input, such that:
y=f ( F ( x,v ), u )
where y represents the speech from the mismatched speaker input, x is the speech used to train the acoustic model, u represents at least one parameter for modeling changes in the environment and v represents at least one parameter used for mapping differences between speakers; and
jointly estimating u and v,
wherein said joint estimation of u and v is performed using the expectation maximization algorithm and comprises optimizing u and v in a single maximization step of said algorithm,
wherein said at least one parameter u comprises parameters n and h, where n is used to model additive noise and h is used to model convolutional noise, and
wherein said mismatch function f is of the form:
y
s
=
q
(
x
s
,
n
s
,
h
s
)
=
F
(
x
s
)
+
h
s
+
Cln
(
1
+
ⅇ
C
-
1
(
n
s
-
F
(
x
s
)
-
h
s
)
)
where C the discrete cosine transformation matrix and the subscript s denotes static part.
2. The method according to claim 1 , wherein said speaker transform is a linear transform.
3. The method according to claim 2 , wherein said speaker transform is a vocal tract length normalisation transform.
4. The method according to claim 3 , wherein at least one speaker transform parameter is a discrete parameter.
5. The method according to claim 1 , wherein adapting the acoustic model further comprising using PCMLLR with the CMLLR form:
p ( y|m )=| A c (r) |N ( A c (r) y+b c (r) ;μ x (m) ,Σ x (m) )
where m denotes the m th mixture in a Gaussian distribution of a Hidden Markov Model with mean μ x (m) and variance Σ x (m) , A c (r) and b c (r) are CMLLR transforms to be estimated using PCMLLR techniques which minimises the divergence between the CMLLR form and a target distribution, said target distribution being derived from y=f(F(x, v), u).
6. The method according to claim 1 , wherein the adaptation is provided in an adaptive training framework.
7. The method according to claim 1 , wherein adapting the acoustic model to the mismatched speaker input comprises receiving speech from the mismatched speaker input corresponding to known text.
8. The method according to claim 1 , wherein adapting the acoustic model to the mismatched speaker input comprises receiving speech from said new speaker and making a first estimate of the text corresponding to said speech.
9. A method of adapting an acoustic model for speech processing to a mismatched speech input, said mismatched speech input being received from a speaker in an environment which is not matched to the speaker or environment under which the acoustic model was trained, the method comprising:
receiving a mismatched speech input which comprises a sequence of feature vectors; and
providing an acoustic model for performing speech processing on an input signal which comprises a sequence of feature vectors, said model having a plurality of model parameters relating to the probability distribution of a word or part thereof being related to a feature vector,
wherein adapting the acoustic model to the mismatched speaker input comprises: relating speech from mismatched speaker input to the speech used to train the acoustic model using: a mismatch function f for primarily modeling differences between the environment of the speaker and the environment under which the acoustic model was trained; and a speaker transform F for primarily modeling differences between the speaker of the mismatched speaker input, such that:
y=f ( F ( x,v ), u )
where y represents the speech from the mismatched speaker input, x is the speech used to train the acoustic model, u represents at least one parameter for modeling changes in the environment and v represents at least one parameter used for mapping differences between speakers; and
jointly estimating u and v,
wherein said joint estimation of u and v is performed using the expectation maximization algorithm and comprises optimizing u and v in a single maximization step of said algorithm,
wherein said at least one parameter u comprises parameters n and h, where n is used to model additive noise and h is used to model convolutional noise, and
wherein said mismatch function f is of the form:
y
s
=
q
(
x
s
,
n
s
,
h
s
)
=
F
(
x
s
)
+
h
s
+
Cln
(
1
+
ⅇ
C
-
1
(
n
s
-
F
(
x
s
)
-
h
s
)
)
where C the discrete cosine transformation matrix and the subscript s denotes static part.
10. A non-transitory carrier medium carrying computer readable instructions for controlling the computer to carry out the method of claim 1 .
11. A speech processing system, comprising:
a receiver for receiving a speech input which comprises a sequence of feature vectors;
and
a processor configured to:
determine a likelihood of a sequence of words arising from the sequence of feature vectors using an acoustic model and a language model, comprising:
provide an acoustic model for performing speech recognition on an input signal which comprises a sequence of feature vectors, said model having a plurality of model parameters relating to the probability distribution of a word or part thereof being related to a feature vector, wherein said speech input is a mismatched speech input which is received from a speaker in an environment which is not matched to the speaker or environment under which the acoustic model was trained; and
adapt the acoustic model to the mismatched speech input,
the processor being further configured to determine a likelihood of a sequence of features occurring in a given language using a language model; and
combine the likelihoods determined by the acoustic model and the language model, the system further comprising an output configured to output a sequence of words identified from said speech input signal,
wherein adapting the acoustic model to the mismatched speaker input comprises:
relating speech from mismatched speaker input to the speech used to train the acoustic model using: a mismatch function f for primarily modeling differences between the environment of the speaker and the environment under which the acoustic model was trained; and a speaker transform F for primarily modeling differences between the speaker of the mismatched speaker input, such that:
y=f ( F ( x,v ), u )
where y represents the speech from the mismatched speaker input, x is the speech used to train the acoustic model, u represents at least one parameter for modeling changes in the environment and v represents at least one parameter used for mapping differences between speakers; and
jointly estimating u and v,
wherein said joint estimation of u and v is performed using the expectation maximization algorithm and comprises optimizing u and v in a single maximization step of said algorithm,
wherein said at least one parameter u comprises parameters n and h, where n is used to model additive noise and h is used to model convolutional noise, and
wherein said mismatch function f is of the form:
y
s
=
q
(
x
s
,
n
s
,
h
s
)
=
F
(
x
s
)
+
h
s
+
Cln
(
1
+
ⅇ
C
-
1
(
n
s
-
F
(
x
s
)
-
h
s
)
)
where C the discrete cosine transformation matrix and the subscript s denotes static part.