Multi-task training architecture and strategy for attention-based speech recognition system
Methods and apparatuses are provided for performing sequence to sequence (Seq2Seq) speech recognition training performed by at least one processor. The method includes acquiring a training set comprising a plurality of pairs of input data and target data corresponding to the input data, encoding the input data into a sequence of hidden states, performing a connectionist temporal classification (CTC) model training based on the sequence of hidden states, performing an attention model training based on the sequence of hidden states, and decoding the sequence of hidden states to generate target labels by independently performing the CTC model training and the attention model training.
1. A method of performing sequence to sequence (Seq2Seq) speech recognition training by at least one processor, the Seq2Seq speech recognition training method comprising:
acquiring, by the at least one processor, a training set comprising a plurality of pairs of input data and target data corresponding to the input data;
encoding, by an encoder implemented by the at least one processor, the input data into a sequence of hidden states;
performing, by the at least one processor, a connectionist temporal classification (CTC) model training based on the sequence of hidden states;
performing, by the at least one processor, an attention model training based on a new sequence of hidden states generated based on the sequence of hidden states, the new sequence of hidden states being different from the sequence of hidden states used by the CTC model training; and
decoding, by a decoder implemented by the at least one processor, the sequence of hidden states to generate target labels by independently performing the CTC model training and the attention model training based on the same input data.
2. The Seq2Seq speech recognition training method of claim 1 , further comprising:
additionally transforming the sequence of hidden states using additional layers to generate the new sequence of hidden states to enable content match between query and context.
3. The Seq2Seq speech recognition training method of claim 2 , wherein the decoding by the decoder comprises:
generating a query information based on a previous target label and a previous prediction;
generating a context information by calculating a soft alignment over all steps of the additionally transformed sequence of hidden states based on the query; and
generating a target label based on the query information and the context information.
4. The Seq2Seq speech recognition training method of claim 3 , wherein the context information is a summary of speech signals encoded in hidden layers of the encoder.
5. The Seq2Seq speech recognition training method of claim 3 , wherein the context information is generated using scalar energy computed based on content similarity between the additionally transformed sequence of hidden states at each time step and the query information.
6. The Seq2Seq speech recognition training method of claim 1 , further comprising:
performing the CTC model training based on a CTC loss function.
7. The Seq2Seq speech recognition training method of claim 6 , wherein the CTC loss function is defined as a mean of normalized edit distance between hypothesis H(x) and the corresponding targets,
Loss
(
H
,
S
)
=
1
S
∑
x
,
t
∈
S
editDistance
(
H
(
x
)
,
t
)
t
where S=(x, t) is the training set containing all pairs of input x and its corresponding target t.
8. The Seq2Seq speech recognition training method of claim 1 , further comprising:
performing the attention model training based on a cross entropy loss function.
9. The Seq2Seq speech recognition training method of claim 1 , wherein the independently performing the CTC model training and the attention model training comprises:
performing the CTC model training to minimize CTC loss during a first time period; and
performing the attention model training to minimize cross entropy loss during a second period different from the first period.
10. The Seq2Seq speech recognition training method of claim 1 , wherein the independently performing the CTC model training and the attention model training comprises:
performing a mini-batch based alternate training in which one of the CTC model training and the attention model training is randomly selected for optimization in each mini-batch.
11. A sequence to sequence (Seq2Seq) speech recognition training apparatus comprising:
at least one memory operable to store program code; and
at least one processor operable to read said program code and operate as instructed by said program code to:
acquire a training set comprising a plurality of pairs of input data and target data corresponding to the input data;
encode the input data into a sequence of hidden states;
perform a connectionist temporal classification (CTC) model training based on the sequence of hidden states;
perform an attention model training based on a new sequence of hidden states generated based on the sequence of hidden states, the new sequence of hidden states being different from the sequence of hidden states used by the CTC model training; and
decode the sequence of hidden states to generate target labels by independently performing the CTC model training and the attention model training based on the same input data.
12. The Seq2Seq speech recognition training apparatus of claim 11 , wherein the at least one processor is further configured to:
additionally transform the sequence of hidden states using additional layers to generate the new sequence of hidden states to enable content match between query and context.
13. The Seq2Seq speech recognition training apparatus of claim 12 , wherein the decoding the sequence of hidden states to generate the target label comprises:
generating a query information based on a previous target label and a previous prediction;
generating a context information by calculating a soft alignment over all steps of the additionally transformed sequence of hidden states based on the query; and
generating a target label based on the query information and the context information.
14. The Seq2Seq speech recognition training apparatus of claim 13 , wherein the context information is a summary of speech signals encoded in hidden layers of the encoder.
15. The Seq2Seq speech recognition training apparatus of claim 13 , wherein the context information is generated using scalar energy computed based on content similarity between the additionally transformed sequence of hidden states at each time step and the query information.
16. The Seq2Seq speech recognition training apparatus of claim 11 , wherein the at least one processor is further configured to:
perform the CTC model training based on a CTC loss function.
17. The Seq2Seq speech recognition training apparatus of claim 11 , wherein the at least one processor is further configured to:
perform the attention model training based on a cross entropy loss function.
18. The Seq2Seq speech recognition training apparatus of claim 11 , wherein the independently performing the CTC model training and the attention model training comprises:
performing the CTC model training to minimize CTC loss during a first time period; and
performing the attention model training to minimize cross entropy loss during a second period different from the first period.
19. The Seq2Seq speech recognition training apparatus of claim 11 , wherein the independently performing the CTC model training and the attention model training comprises:
performing a mini-batch based alternate training in which one of the CTC model training and the attention model training is randomly selected for optimization in each mini-batch.
20. A non-transitory computer readable medium having stored thereon a program for performing sequence to sequence (seq2seq) speech recognition training method, the Seq2Seq speech recognition training method comprising:
acquiring, by at least one processor, a training set comprising a plurality of pairs of input data and target data corresponding to the input data;
encoding, by an encoder implemented by the at least one processor, the input data into a sequence of hidden states;
performing, by the at least one processor, a connectionist temporal classification (CTC) model training based on the sequence of hidden states;
performing, by the at least one processor, an attention model training based on a new sequence of hidden states generated based on the sequence of hidden states, the new sequence of hidden states being different from the sequence of hidden states used by the CTC model training; and
decoding, by a decoder implemented by the at least one processor, the sequence of hidden states to generate target labels by independently performing the CTC model training and the attention model training on the same input data.