Granted Patent B2

US 12,431,150 · App. 18/118,682 · Granted Sep 30, 2025

Inventors: Myeongjin Hwang (Seoul, KR); Suntae Kim (Seoul, KR); Changjin Ji (Seoul, KR)

Assignee: LLSOLLU CO., LTD.

G10L19/022G10L15/22G10L25/87

Abstract

Claims (44)

1. A voice conversation reconstruction method performed by a voice conversation reconstruction apparatus, the method comprising:

acquiring a plurality of speaker-specific voice recognition data corresponding to a plurality of speakers about voice conversation;

dividing each of the plurality of the speaker-specific voice recognition data into a plurality of blocks using a boundary between tokens such that each of the divided plurality of the speaker-specific voice recognition data includes voice data only by a single speaker, wherein the divided plurality of the speaker-specific voice recognition data are not in chronological order;

arranging the plurality of blocks of all the speaker-specific voice recognition data in chronological order without distinction of speaker;

among the arranged plurality of blocks, merging blocks when the blocks are neighbor and the speaker of the blocks are the same such that the speaker-specific voice recognition data in each of the merged blocks are in chronological order and include voice data only by the same speaker; and

reconstructing the plurality of blocks subjected to the merging in a conversation format in chronological order and based on a speaker step by step such that the speaker-specific voice recognition data in each of the reconstructed blocks are in chronological order and include voice data only by the same speaker,

wherein the steps are performed in order,

wherein acquiring the plurality of speaker-specific voice recognition data includes:

acquiring a first speaker-specific recognition result generated on an End Point Detection (EPD), and a second speaker-specific recognition result generated every preset time, and

collecting the first speaker-specific recognition result and the second speaker-specific recognition result without overlap and redundance therebetween to generate the speaker-specific voice recognition data, and

wherein the second speaker-specific recognition result is generated after a last EPD at which the first speaker-specific recognition result is generated occurs.

2. The method of claim 1 , wherein acquiring the speaker-specific voice recognition data includes:

acquiring a speaker-specific recognition result A generated on an EPD (End Point Detection) basis from the voice conversation and a speaker-specific recognition result B which is a partial result generated from the voice conversation; and

when the speaker of the A and the speaker of the B are the same, collecting the recognition result A and the recognition result B without overlap therebetween to generate the speaker-specific voice recognition data.

3. The method of claim 2 , wherein the second speaker-specific recognition result B is generated after the same speaker's last EPD occurs.

4. The method of claim 1 , wherein the merging is not performed when a silence period between neighboring tokens is longer than a predetermined time, or is not grammatically connected.

5. The method of claim 2 , wherein the method further comprises outputting the voice recognition data reconstructed in the conversation format on a screen, and wherein when the screen is updated, the speaker-specific voice recognition data is collectively updated or is updated based on the speaker-specific recognition result B.

6. A voice conversation reconstruction apparatus comprising:

an input unit configured to receive voice conversation input; and

a processor configured to process voice recognition of the voice conversation received through the input unit,

wherein the processor is configured to:

acquire a plurality of speaker-specific voice recognition data corresponding to a plurality of speakers about voice conversation;

divide each of the plurality of the speaker-specific voice recognition data into a plurality of blocks using a boundary between tokens such that each of the divided plurality of the speaker-specific voice recognition data includes voice data only by a single speaker, wherein the divided plurality of the speaker-specific voice recognition data are not in chronological order;

arrange the plurality of blocks of all the speaker-specific voice recognition data in chronological order without distinction of speaker;

merge blocks when the blocks are neighbor and the speaker of the blocks are the same such that the speaker-specific voice recognition data in each of the merged blocks are in chronological order and include voice data only by the same speaker; and

reconstruct the plurality of blocks subjected to the merging in a conversation format in chronological order and based on a speaker such that the speaker-specific voice recognition data in each of the reconstructed blocks are in chronological order and include voice data only by the same speaker,

wherein the processor is further configured to:

acquire a first speaker-specific recognition result generated on an End Point Detection (EPD), and a second speaker-specific recognition result generated every preset time, and

collect the first speaker-specific recognition result and the second speaker-specific recognition result without overlap and redundance therebetween to generate the speaker-specific voice recognition data, and

wherein the second speaker-specific recognition result is configured to be generated after a last EPD at which the first speaker-specific recognition result is generated occurs.

7. The apparatus of claim 6 , wherein the processor is further configured to:

acquire a speaker-specific recognition result A generated on an EPD (End Point Detection) basis from the voice conversation and a speaker-specific recognition result B which is a partial result generated from the voice conversation; and

collect the speaker-specific recognition result A and the speaker-specific recognition result B without overlap and redundance therebetween to generate the speaker-specific voice recognition data.

8. A non-transitory computer-readable recording medium storing instructions, when executed by one or more processors, that cause the one or more processors to perform a method comprising:

acquiring a plurality of speaker-specific voice recognition data corresponding to a plurality of speakers about voice conversation;

dividing each of the plurality of the speaker-specific voice recognition data into a plurality of blocks using a boundary between tokens such that each of the divided plurality of the speaker-specific voice recognition data includes voice data only by a single speaker, wherein the divided plurality of the speaker-specific voice recognition data are not in chronological order;

arranging the plurality of blocks of all the speaker-specific voice recognition data in chronological order without distinction of speaker;

among the arranged plurality of blocks, merging blocks when the blocks are neighbor and the speaker of the blocks are the same such that the speaker-specific voice recognition data in each of the merged blocks are in chronological order and include voice data only by the same speaker; and

reconstructing the plurality of blocks subjected to the merging in a conversation format in chronological order and based on a speaker step by step such that the speaker-specific voice recognition data in each of the reconstructed blocks are in chronological order and include voice data only by the same speaker,

wherein the steps are performed in order,

wherein acquiring the plurality of speaker-specific voice recognition data includes:

acquiring a first speaker-specific recognition result generated on an End Point Detection (EPD), and a second speaker-specific recognition result generated every preset time, and

collecting the first speaker-specific recognition result and the second speaker-specific recognition result without overlap and redundance therebetween to generate the speaker-specific voice recognition data, and

wherein the second speaker-specific recognition result is generated after a last EPD at which the first speaker-specific recognition result is generated occurs.

Assignments (1)

Priority Claims (1)

Continuity (2)

References Cited (41)