System and method to correct for packet loss using hidden markov models in ASR systems
A system and method are presented for the correction of packet loss in audio in automatic speech recognition (ASR) systems. Packet loss correction, as presented herein, occurs at the recognition stage without modifying any of the acoustic models generated during training. The behavior of the ASR engine in the absence of packet loss is thus not altered. To accomplish this, the actual input signal may be rectified, the recognition scores may be normalized to account for signal errors, and a best-estimate method using information from previous frames and acoustic models may be used to replace the noisy signal.
1. A method to correct for packet loss in an audio signal in automatic speech recognition systems comprising maximum likelihood prediction wherein the maximum likelihood prediction further comprises the steps of:
a. determining Hidden Markov Model state occupation probabilities in an acoustic model within the automatic speech recognition system prior to a packet loss event, wherein the Hidden Markov Model state occupation probabilities are determined using the mathematical equation
P
(
x
(
t
+
1
)
❘
x
(
1
)
,
x
(
2
)
…
x
(
t
)
)
=
∑
ik
alpha
(
i
)
t
(
ik
)
b
(
k
)
(
x
(
t
+
1
)
)
where P represents a probability distribution of a feature vector at x(t+1), alpha(i) represents a state occupation probability at time t for state (i), t(ik) represents a state transition probability from state i to k, and where b(k)(x(t+1)) represents a Gaussian mixture distribution for a feature x(t+1) from a state k belonging to a phoneme model; and
b. predicting feature values of packets in the packet loss section of the audio signal based on the determined Hidden Markov Model state occupation probabilities.
2. The method of claim 1 , wherein the predicting of the feature values further comprises determining if a feature value is at a first state of the acoustic model and if a next state is at the first state or at a second state.
3. The method of claim 1 , wherein the predicting of the feature values further comprises determining if a feature value is at a second state and a next state is in a same state or an other state of the acoustic model.
4. The method of claim 3 , wherein the other state of the acoustic model comprises a last state.
5. The method of claim 4 , wherein a next state after the last state must be the same state or a first state of a next node in a lexical network.
6. The method of claim 1 wherein audio frames of the audio signal are decomposed into overlapping frames of 20 ms with an overlap factor.