Conditional factorization for jointly modeling code-switched and monolingual ASR
A method, apparatus, and non-transitory computer-readable medium for automatic speech recognition using conditional factorization for bilingual code-switched and monolingual speech may include receiving an audio observation sequence comprising a plurality of frames, the audio observation sequence including audio in a first language or a second language. The approach may further include mapping the audio observation sequence into a first sequence of hidden representations, the mapping being generated by a first encoder corresponding to the first language and mapping the audio observation sequence into a second sequence of hidden representations, the mapping being generated by a second encoder corresponding to the second language. The approach may further include generating a label-to-frame sequence based on the first sequence of hidden representations and the second sequence of hidden representations, using a joint neural network based model.
1 . A method of automatic speech recognition using conditional factorization for bilingual code-switched and monolingual speech, the method executed by one or more processors, the method comprising:
receiving an audio observation sequence comprising a plurality of frames, the audio observation sequence including audio in a first language or a second language;
mapping the audio observation sequence into a first sequence of hidden representations, the mapping being generated by a first encoder corresponding to the first language, wherein the first encoder is configured to generate respective first labels associated with respective frames of the plurality of frames of the audio observation sequence based on determining that the respective frames of the plurality of frames of the audio observation sequence are associated with the first language, and wherein the first encoder is configured to generate blank labels associated with the respective frames of the plurality of frames of the audio observation sequence based on determining that the respective frames of the plurality of frames of the audio observation sequence are not associated with the first language;
mapping the audio observation sequence into a second sequence of hidden representations, the mapping being generated by a second encoder corresponding to the second language; and
generating a label-to-frame sequence based on the first sequence of hidden representations and the second sequence of hidden representations, using a joint neural network based model.
2 . The method of claim 1 , wherein the generating the label-to-frame sequence comprises:
fusing the first sequence of hidden representations and the second sequence of hidden representations to generate a fused sequence of hidden representations; and
generating the label-to-frame sequence based on the fused sequence of hidden representations using the joint neural network based model.
3 . The method of claim 1 , wherein a label at any position in the label-to-frame sequence is associated with the first language or the second language, and wherein the label is not associated with both the first language and the second language.
4 . The method of claim 1 , wherein the first encoder is a first pre-trained neural network model trained using a first connectionist temporal classification framework, and wherein the second encoder is a second pre-trained neural network model trained using a second connectionist temporal classification framework.
5 . The method of claim 4 , wherein the second encoder is configured to generate respective second labels associated with respective frames of the plurality of frames of the audio observation sequence, based on determining that the respective frames of the plurality of frames of the audio observation sequence are associated with the second language, and wherein the second encoder is configured to generate blank labels associated with the respective frames of the plurality of frames of the audio observation sequence based on determining that the respective frames of the plurality of frames of the audio observation sequence are not associated with the second language.
6 . The method of claim 1 , wherein the joint neural network based model is trained using a loss function based on a joint loss associated with the joint neural network based model, a first loss associated with the first encoder, and a second loss associated with the second encoder.
7 . An apparatus for automatic speech recognition using conditional factorization for bilingual code-switched and monolingual speech, the apparatus comprising:
at least one memory configured to store program code; and
at least one processor configured to read the program code and operate a instructed by the program code, the program code including:
first receiving code configured to cause the at least one processor to receive an audio observation sequence comprising a plurality of frames, the audio observation sequence including audio in a first language or a second language;
first mapping code configured to cause the at least one processor to map the audio observation sequence into a first sequence of hidden representations, the mapping being generated by a first encoder corresponding to the first language, wherein the first encoder is configured to generate respective first labels associated with respective frames of the plurality of frames of the audio observation sequence based on determining that the respective frames of the plurality of frames of the audio observation sequence are associated with the first language, and wherein the first encoder is configured to generate blank labels associated with the respective frames of the plurality of frames of the audio observation sequence based on determining that the respective frames of the plurality of frames of the audio observation sequence are not associated with the first language;
second mapping code configured to cause the at least one processor to map the audio observation sequence into a second sequence of hidden representations, the mapping being generated by a second encoder corresponding to the second language; and
first generating code configured to cause the at least one processor to generate a label-to-frame sequence based on the first sequence of hidden representations and the second sequence of hidden representations, using a joint neural network based model.
8 . The apparatus of claim 7 , wherein the first generating code comprises:
fusing code configured to cause the at least one processor to fuse the first sequence of hidden representations and the second sequence of hidden representations to generate a fused sequence of hidden representations; and
second generating code configured to cause the at least one processor to generate the label-to-frame sequence based on the fused sequence of hidden representations using the joint neural network based model.
9 . The apparatus of claim 7 , wherein a label at any position in the label-to-frame sequence is associated with the first language or the second language, and wherein the label is not associated with both the first language and the second language.
10 . The apparatus of claim 7 , wherein the first encoder is a first pre-trained neural network model trained using a first connectionist temporal classification framework, and wherein the second encoder is a second pre-trained neural network model trained using a second connectionist temporal classification framework.
11 . The apparatus of claim 10 , wherein the second encoder is configured to generate respective second labels associated with respective frames of the plurality of frames of the audio observation sequence, based on determining that the respective frames of the plurality of frames of the audio observation sequence are associated with the second language, and wherein the second encoder is configured to generate blank labels associated with the respective frames of the plurality of frames of the audio observation sequence based on determining that the respective frames of the plurality of frames of the audio observation sequence are not associated with the second language.
12 . The apparatus of claim 7 , wherein the joint neural network based model is trained using a loss function based on a joint loss associated with the joint neural network based model, a first loss associated with the first encoder, and a second loss associated with the second encoder.
13 . A non-transitory computer-readable medium storing instructions, the instructions comprising: one or more instructions that, when executed by one or more processors of a device for automatic speech recognition using conditional factorization for bilingual code-switched and monolingual speech, cause the one or more processors to:
receive an audio observation sequence comprising a plurality of frames, the audio observation sequence including audio in a first language or a second language;
map the audio observation sequence into a first sequence of hidden representations, the mapping being generated by a first encoder corresponding to the first language, wherein the first encoder is configured to generate respective first labels associated with respective frames of the plurality of frames of the audio observation sequence based on determining that the respective frames of the plurality of frames of the audio observation sequence are associated with the first language, and wherein the first encoder is configured to generate blank labels associated with the respective frames of the plurality of frames of the audio observation sequence based on determining that the respective frames of the plurality of frames of the audio observation sequence are not associated with the first language;
map the audio observation sequence into a second sequence of hidden representations, the mapping being generated by a second encoder corresponding to the second language; and
generate a label-to-frame sequence based on the first sequence of hidden representations and the second sequence of hidden representations, using a joint neural network based model.
14 . The non-transitory computer-readable medium of claim 13 , wherein the generating the label-to-frame sequence comprises:
fusing the first sequence of hidden representations and the second sequence of hidden representations to generate a fused sequence of hidden representations; and
generating the label-to-frame sequence based on the fused sequence of hidden representations using the joint neural network based model.
15 . The non-transitory computer-readable medium of claim 13 , wherein a label at any position in the label-to-frame sequence is associated with the first language or the second language, and wherein the label is not associated with both the first language and the second language.
16 . The non-transitory computer-readable medium of claim 13 , wherein the first encoder is a first pre-trained neural network model trained using a first connectionist temporal classification framework, and wherein the second encoder is a second pre-trained neural network model trained using a second connectionist temporal classification framework.
17 . The non-transitory computer-readable medium of claim 16 , wherein the second encoder is configured to generate respective second labels associated with respective frames of the plurality of frames of the audio observation sequence, based on determining that the respective frames of the plurality of frames of the audio observation sequence are associated with the second language, and wherein the second encoder is configured to generate blank labels associated with the respective frames of the plurality of frames of the audio observation sequence based on determining that the respective frames of the plurality of frames of the audio observation sequence are not associated with the second language.