Granted Patent B2

US 12,562,161 · App. 18/101,217 · Granted Feb 24, 2026

Inventors: Xiao Xia Mao (Shanghai, CN); Wei Jun Zheng (Shanghai, CN); Chun Hong Zheng (Beijing, CN); Fu Er Liu (Shanghai, CN); Danlei Zhang (Shanghai, CN); Xiao Feng Ji (Shanghai, CN)

Assignee: International Business Machines Corporation

G10L15/22G10L15/063G10L15/26

Abstract

Claims (65)

1 . A computer-implemented method, comprising:

training a machine learning computer model on a dual-modal input sequence comprising target content data and an input user query specifying attributes for locating a portion within the target content data, and wherein the machine learning computer model is trained to process one or more audio tokens corresponding to a sequence of spectrogram chunks associated with the target content data and one or more text tokens associated with the input user query and output an output text specifying start and end indices associated with the sequence of spectrograph chunks and corresponding to the portion within the target content data;

receiving runtime content data and a user spoken query comprising attributes directed to identifying a section of the runtime content data;

processing the runtime content data and user spoken query via the trained machine learning computer model to generate an output text specifying a predicted beginning index and predicted ending index within the runtime content data;

retrieving a portion of the runtime content data corresponding to a range of runtime content data bounded by the predicted beginning index and predicted ending index, inclusively; and

storing the retrieved portion in a data storage for later retrieval and processing.

2 . The computer-implemented method of claim 1 ,

wherein training the machine learning computer model comprises:

converting the target content data to the sequence of spectrogram chunks;

converting each spectrogram chunk to the one or more audio tokens;

converting the input user query to the one or more text tokens;

processing the sequence of one or more audio tokens corresponding to spectrogram chunks in the sequence of spectrogram chunks, along with the one or more text tokens of the input user query, to generate an output prediction of a first index and a second index, of spectrogram chunks in the sequence of spectrogram chunks;

comparing the first index and second index to a ground truth set of indices; and

modifying operational parameters of the machine learning computer model to reduce an error between the first index and second index and the ground truth, to thereby train the machine learning computer model to predict a starting index and an ending index for a portion of content based on a dual modal input of text tokens from an input user query and audio tokens from a spectrogram of input target content data.

3 . The computer-implemented method of claim 1 ,

wherein the input user query is a templated input query from a plurality of predetermined templated input queries, and wherein the machine learning computer model learns, for each templated input query in the plurality of predetermined templated input queries, first spectrogram chunks and second spectrogram chunks of a plurality of spectrogram chunks corresponding to the templated input query.

4 . The computer-implemented method of claim 3 ,

wherein the plurality of predetermined templated input queries comprise a first input query to identify a last or current content type relative to a current listening time point, a second query to identify a number of sentences centered on a specified keyword in the target content data, or a third query to retrieve a specified number of language elements prior to the current listening time point.

5 . The computer-implemented method of claim 1 ,

wherein retrieving a portion of the runtime content data corresponding to the range of runtime content data bounded by the predicted beginning index and the predicted ending index comprises mapping the predicted beginning index to a first timestamp of the runtime content data and mapping the predicted ending index to a second timestamp of the runtime content data.

6 . The computer-implemented method of claim 1 ,

wherein processing the user spoken query via the trained machine learning computer model comprises converting the user spoken query to a first textual representation data structure by executing speech-to-text conversion, and processing the first textual representation data structure via the trained machine learning computer model to generate the output text.

7 . The computer-implemented method of claim 1 ,

wherein storing the retrieved portion in the data storage for later retrieval and processing further comprises executing speech-to-text conversion to generate a second textual representation data structure corresponding to content of the retrieved portion and storing the second textual representation data structure in association with the retrieved portion in the data storage.

8 . The computer-implemented method of claim 1 ,

wherein generating the output text comprises generating a templated text sentence which includes the predicted beginning index and predicted ending index, each index corresponding to a spectrogram chunk in the sequence of spectrogram chunks.

9 . The computer-implemented method of claim 1 , wherein the target content data comprises pre-recorded digital data of at least one of a podcast or a digital audio book, and wherein the retrieved portion is a portion of the podcast or digital audio book that is to be highlighted.

10 . The computer-implemented method of claim 1 ,

wherein the target content data is audio data.

11 . A computer program product comprising a computer readable storage medium having a computer readable program stored therein, wherein the computer readable program, when executed on a computing device, causes the computing device to:

train a machine learning computer model on a dual-modal input sequence comprising target content data and an input user query specifying attributes for locating a portion within the target content data, and wherein the machine learning computer model is trained to process one or more audio tokens corresponding to a sequence of spectrogram chunks associated with the target content data and one or more text tokens associated with the input user query and output an output text specifying start and end indices associated with the sequence of spectrograph chunks and corresponding to the portion within the target content data;

receive runtime content data and a user spoken query comprising attributes directed to identifying a section of the runtime content data;

process the runtime content data and user spoken query via the trained machine learning computer model to generate an output text specifying a predicted beginning index and predicted ending index within the runtime content data;

retrieve a portion of the runtime content data corresponding to a range of runtime content data bounded by the predicted beginning index and predicted ending index, inclusively; and

store the retrieved portion in a data storage for later retrieval and processing.

12 . The computer program product of claim 11 ,

wherein training the machine learning computer model comprises:

converting the target content data to the sequence of spectrogram chunks;

converting each spectrogram chunk to the one or more audio tokens;

converting the input user query to the one or more text tokens;

processing the sequence of one or more audio tokens corresponding to spectrogram chunks in the sequence of spectrogram chunks, along with the one or more text tokens of the input user query, to generate an output prediction of a first index and a second index, of spectrogram chunks in the sequence of spectrogram chunks;

comparing the first index and second index to a ground truth set of indices; and

modifying operational parameters of the machine learning computer model to reduce an error between the first index and second index and the ground truth, to thereby train the machine learning computer model to predict a starting index and an ending index for a portion of content based on a dual modal input of text tokens from an input user query and audio tokens from a spectrogram of input target content data.

13 . The computer program product of claim 11 ,

wherein the input user query is a templated input query from a plurality of predetermined templated input queries, and wherein the machine learning computer model learns, for each templated input query in the plurality of predetermined templated input queries, first spectrogram chunks and second spectrogram chunks of a plurality of spectrogram chunks corresponding to the templated input query.

14 . The computer program product of claim 13 ,

wherein the plurality of predetermined templated input queries comprise a first input query to identify a last or current content type relative to a current listening time point, a second query to identify a number of sentences centered on a specified keyword in the target content data, or a third query to retrieve a specified number of language elements prior to the current listening time point.

15 . The computer program product of claim 11 ,

wherein retrieving a portion of the runtime content data corresponding to the range of runtime content data bounded by the predicted beginning index and predicted ending index comprises mapping the predicted beginning index to a first timestamp of the runtime content data and mapping the predicted ending index to a second timestamp of the runtime content data.

16 . The computer program product of claim 11 ,

wherein processing the user spoken query via the trained machine learning computer model comprises converting the user spoken query to a first textual representation data structure by executing speech-to-text conversion, and processing the first textual representation data structure via the trained machine learning computer model to generate the output text.

17 . The computer program product of claim 11 ,

wherein storing the retrieved portion in the data storage for later retrieval and processing further comprises executing speech-to-text conversion to generate a second textual representation data structure corresponding to content of the retrieved portion and storing the second textual representation data structure in association with the retrieved portion in the data storage.

18 . The computer program product of claim 11 ,

wherein generating the output text comprises generating a templated text sentence which includes the predicted beginning index and predicted ending index, each index corresponding to a spectrogram chunk in the sequence of spectrogram chunks.

19 . The computer program product of claim 11 ,

wherein the target content data comprises pre-recorded digital data of at least one of a podcast or a digital audio book, and wherein the retrieved portion is a portion of the podcast or digital audio book that is to be highlighted.

20 . An apparatus comprising:

at least one processor; and

at least one memory coupled to the at least one processor, wherein the at least one memory comprises instructions which, when executed by the at least one processor, cause the at least one processor to:

train a machine learning computer model on a dual-modal input sequence comprising target content data and an input user query specifying attributes for locating a portion within the target content data, and wherein the machine learning computer model is trained to process one or more audio tokens corresponding to a sequence of spectrogram chunks associated with the target content data and one or more text tokens associated with the input user query and output an output text specifying start and end indices associated with the sequence of spectrograph chunks and corresponding to the portion within the target content data;

receive runtime content data and a user spoken query comprising attributes directed to identifying a section of the runtime content data;

process the runtime content data and user spoken query via the trained machine learning computer model to generate an output text specifying a predicted beginning index and predicted ending index within the runtime content data;

retrieve a portion of the runtime content data corresponding to a range of runtime content data bounded by the predicted beginning index and predicted ending index, inclusively; and

store the retrieved portion in a data storage for later retrieval and processing.

Assignments (1)

Continuity (1)

References Cited (27)