Abstract: A method of enabling a language model trained in a first language to support a second language includes extending an existing vocabulary of the language model to include additional tokens for text in the second language. The method also includes initializing the additional tokens for text in the second language based on subtokens of tokens for text in the first language from the existing vocabulary. The method further includes training the language model using a mixed language dataset that includes a first language corpus and a second language corpus. In addition, the method includes performing instruction tuning using a dataset that includes (i) instruction and response pairs involving the first language and (ii) instruction and response pairs involving the second language.

Abstract: A method of enabling a language model trained in a first language to support a second language includes extending an existing vocabulary of the language model to include additional tokens for text in the second language. The method also includes initializing the additional tokens for text in the second language based on subtokens of tokens for text in the first language from the existing vocabulary. The method further includes training the language model using a mixed language dataset that includes a first language corpus and a second language corpus. In addition, the method includes performing instruction tuning using a dataset that includes (i) instruction and response pairs involving the first language and (ii) instruction and response pairs involving the second language.

Abstract: A method includes obtaining a parameter matrix associated with a linear layer of a first machine learning model and containing parameter values for parameters of the linear layer. The method also includes determining importance values corresponding to the parameter values. The method further includes generating factorized matrices such that a product of the importance values and factorized matrices contains approximated parameter values for the parameters of the linear layer. In addition, the method includes generating a second machine learning model representing a compressed version of the first machine learning model. The second machine learning model has first and second linear layers containing parameter values based on the importance values and the factorized matrices. The factorized matrices are generated based on weighted errors between the parameter values for the parameters of the linear layer and the approximated parameter values.

Abstract: A computer system that generates output speech is described. During operation, the computer system may receive an input associated with a type of interaction. Then, the computer system may generate, using a voice synthesis engine, the output speech corresponding to an individual based at least in part on the input, where the voice synthesis engine predicts positions and duration of a prosodic characteristic of speech by the individual, and selectively adds the prosodic characteristic of the speech by the individual in the output speech based at least in part on the prediction. Note that the prosodic characteristic may include: pauses in the speech by the individual, and/or disfluences in the speech by the individual.

Abstract: A computer system that generates output speech is described. During operation, the computer system may receive an input associated with a type of interaction. Then, the computer system may generate, using a voice synthesis engine, the output speech corresponding to an individual based at least in part on the input, where the voice synthesis engine predicts positions and duration of a prosodic characteristic of speech by the individual, and selectively adds the prosodic characteristic of the speech by the individual in the output speech based at least in part on the prediction. Note that the prosodic characteristic may include: pauses in the speech by the individual, and/or disfluences in the speech by the individual.