3 and T5”, “T3 and T6”, “T3 and T7”, “T4 and T5”, “T4 and T6”, and “T4 and T7”, as comparison between the candidate text group 30 of the piece of audio data B, and the candidate text group 30 of the piece of audio data C.

As a relevance index value, any index value capable of quantifying relevance between sentences can be used. For example, an inter-sentence distance is used as the relevance index value. An inter-sentence distance between two sentences can be, for example, computed as a distance between vectors (hereinafter, sentence vectors), each of which represents a sentence. As the sentence vector, various vectors such as Bag of Words or TF-IDF can be used. Further, as the distance between vectors, a Euclidean distance, a cosine distance, Word Move's Distance, or the like can be used. Note that, a Euclidean distance and Word Move's Distance represent that, as a value decreases, a similarity between sentences increases. On the other hand, a cosine distance represents that, as a value increases, a similarity between sentences increases.

For example, in a case where a candidate text 32 is selected one by one from all candidate text groups 30, the selecting unit 2040 determines a candidate text 32 to be selected from each candidate text group 30, based on a sum of relevance index values to be computed for each pair of candidate texts 32 to be acquired from the selected candidate texts 32. In a case of using a relevance index value representing that, as a value decreases, a similarity between sentences increases, it is configured in such a way that a sum of relevance index values is minimized. On the other hand, in a case of using a relevance index value representing that, as a value increases, a similarity between sentences increases, it is configured in such a way that a sum of relevance index values is maximized. The processing can be formulated as follows, for example.

First, an index vector is defined as expressed by the following formula (1).

[

Mathematical
⁢

1

]



A
∈
C

=

{

(

c
1

c
2

⋮

c
M

)

❘

c
m

∈

{

1
,
2
,
…

,

N
m

}

,

m
∈

{

1
,
2
,
…

,
M

}

}

(
1
)

Herein, M denotes a total number of pieces of audio data 20. Further, m denotes an index of a piece of audio data 20, and represents at which position from a leading position, the piece of audio data 20 is located in a time-series order. Nm denotes a total number of candidate texts 32 included in a candidate text group 30 of the m-th piece of audio data 20. cm denotes an index of a candidate text 32 to be selected from a candidate text group 30 of the m-th piece of audio data 20 from a leading position.

An index vector A represents an index of a candidate text 32 to be selected from each candidate text group 30. For example, A=(2, 3, 1) means that a second candidate text 32 is selected from a candidate text group 30 of a first piece of audio data 20, a third candidate text 32 is selected from a candidate text group 30 of a second piece of audio data 20, and a first candidate text 32 is selected from a candidate text group 30 of a third piece of audio data 20. Therefore, processing of determining a candidate text 32 to be selected from each candidate text group 30 can be formulated as processing of determining an index vector. The selecting unit 2040 determines a candidate text 32 to be selected from a candidate text group 30 of each piece of audio data 20 by determining an index vector, based on the above-described relevance index value.

In a case of using an inter-sentence distance such as a Euclidean distance, which represents that, as a value decreases, a similarity increases, for example, the selecting unit 2040 determines an index vector, as expressed by the following expression (2).

[

Mathematical
⁢

2

]



arg
⁢
min

A

⁢

∑

i
=
1

M

∑

j
>
i

M

d
⁡
(

V

i
,

G
i

(
A
)

,

V

j
,

G
j

(
A
)

)

(
2
)

V_i,j: a sentence vector of the j-th candidate text in the i-th candidate text group

- d(v_i,j, v_k,l): an inter-sentence distance between v_i,jand v_k,l
- G_i(A)=a function from which the i-th element of A is extracted

On the other hand, in a case of using an inter-sentence distance such as a cosine distance, which represents that, as a value increases, a similarity increases, for example, the selecting unit 2040 determines an index vector, as expressed by the following expression (3).

[

Mathematical
⁢

3

]



arg
⁢
max

A

⁢

∑

i
=
1

M

∑

j
>
i

M

d
⁡
(

V

i
,

G
i

(
A
)

,

V

j
,

G
j

(
A
)

)

(
3
)

Note that, an index vector may be determined by applying a weight with respect to each piece of audio data 20, and taking into consideration the weight. In a case of considering a weight, for example, the above-described expressions (2) and (3) can be expressed as the following expressions (4) and (5), respectively.

[

Mathematical
⁢

4

]



arg
⁢
min

A

⁢

∑

i
=
1

M

∑

j
>
i

M

w
i

⁢

w
j

⁢

d
⁡
(

V

i
,

G
i

(
A
)

,

V

j
,

G
j

(
A
)

)

(
4
)

[

Mathematical
⁢

5

]



arg
⁢
max

A

⁢

∑

i
=
1

M

∑

j
>
i

M

w
i

⁢

w
j

⁢

d
⁡
(

V

i
,

G
i

(
A
)

,

V

j
,

G
j

(
A
)

)

(
5
)

Herein, wi represents a weight to be applied to the i-th piece of audio data 20.

A method of applying a weight with respect to a piece of audio data 20 is optional. For example, a weight of a piece of audio data 20 including only a filler word such as “uh” or “hmm” is set to be smaller than a weight of another piece of audio data 20 (e.g., the former is set to be 0, and the latter is set to be 1). In addition to the above, for example, a weight of a piece of audio data 20 including a predetermined word may be set larger than a weight of another piece of audio data 20 (e.g., the former is set to be 2, and the latter is set to be 1). Further, an optimum weight may be determined by machine learning.

In the above-described expressions (2) to (5), an inter-sentence distance between selected candidate texts 32 is computed for each pair of any candidate text groups 30. Therefore, it is possible to determine a content of a remark represented by source data 10, also taking into consideration relevance between sentences not being adjacent to each other in time-series.

However, the selecting unit 2040 may compute an inter-sentence distance of a selected candidate text 32 only for a pair of candidate text groups 30 being adjacent to each other in time-series. In this case, for example, the following expressions (6) and (7) can be used instead of the expressions (2) and (3).

[

Mathematical
⁢

6

]



arg
⁢
min

A

⁢

∑

i
=
1

M
-
1

d
⁡
(

V

i
,

G
i

(
A
)

,

V

i
+
1

,

G

i
+
1

(
A
)

)

(
6
)

[

Mathematical
⁢

7

]



arg
⁢
max

A

⁢

∑

i
=
1

M
-
1

d
⁡
(

V

i
,

G
i

(
A
)

,

V

i
+
1

,

G

i
+
1

(
A
)

)

(
7
)

Note that, a weight may be applied to these expressions (6) and (7), as well as the expressions (4) and (5).

According to the speech recognition apparatus 2000, a sentence (text) representing a content of a piece of audio data 20 acquired from source data 10 is determined for each piece of the audio data 20. Specifically, a sentence string representing a content of an utterance represented by source data 10 is acquired. A use method of a sentence string acquired as described above is optional. For example, the speech recognition apparatus 2000 outputs a generated sentence string. An output destination of the sentence string is optional. For example, the speech recognition apparatus 2000 stores the sentence string in a storage apparatus, causes a display apparatus to display the sentence string, or transmits the sentence string to any another apparatus (e.g., a transmission source of source data 10).

Since a sentence string is a text representing a content of an utterance represented by source data 10, by browsing the sentence string output as described above, a user of the speech recognition apparatus 2000 can visually recognize a content of an utterance represented by the source data 10. For example, in a case where source data 10 are recorded data of a conference, a sentence string to be acquired from the speech recognition apparatus 2000 can be handled as minutes data in which an utterance in the conference is converted into a text.

A part or all of the above-described example embodiment may also be described as the following supplementary notes, but is not limited to the following.

- 1. A speech recognition apparatus including:
  - a candidate generating unit that generates, for each of a plurality of pieces of audio data acquired from a source audio signal including an utterance, a candidate text group including a plurality of candidate texts each being a candidate of a text representing a content of an utterance corresponding to the piece of audio data; and
  - a selecting unit that selects, based on a comparison result between the plurality of candidate text groups, for each of pieces of the audio data, a candidate text representing a content of an utterance represented by the piece of audio data from the candidate text group generated for the piece of audio data, wherein
  - the candidate text represents a sentence.
- 2. The speech recognition apparatus according to supplementary note 1, wherein
  - the selecting unit performs comparison between the candidate text groups generated for at least each of two pieces of audio data not being adjacent to each other in time-series.