コールセンターにおける音声技術の利用～音声のパラ言語情報・非言語情報を含めた活用～

全文

(1)Vol.2013-SLP-98 No.2 2013/10/25. ᝟ሗฎ⌮Ꮫ఍◊✲ሗ࿌ IPSJ SIG Technical Report. ࢥ࣮ࣝࢭࣥࢱ࣮࡟࠾ࡅࡿ㡢ኌᢏ⾡ࡢ฼⏝ . 㹼㡢ኌࡢࣃࣛゝㄒ᝟ሗ࣭㠀ゝㄒ᝟ሗࢆྵࡵࡓά⏝㹼㹼. ᱜ஭῟ᏹ† ᮌᮧ᫴ኴ† ࢥ࣮ࣝࢭࣥࢱ࣮ࡣ㡢ኌ࡟ࡼࡿࢥ࣑ࣗࢽࢣ࣮ࢩࣙࣥࡢ㔜せ࡞άື㡿ᇦ࡛࠶ࡿࡓࡵࠊ㡢ኌ᝟ሗฎ⌮࡟ᑐࡍࡿᮇᚅࡣ≉࡟኱ ࡁ࠸㸬ࡲࡓࠊࢥࣥࣉࣛ࢖࢔ࣥࢫࡢ⥔ᣢ࠾ࡼࡧࢧ࣮ࣅࢫရ㉁ࡢྥୖࡢࡓࡵ࡟㏻ヰ㘓㡢ࡣᚲ㡲࡟࡞ࡗ࡚ࡁ࡚࠸ࡿ㸬ࡑࡢ⤖ ᯝࠊ⭾኱࡞㡢ኌࢹ࣮ࢱࡀ⵳✚ࡉࢀࡿࡼ࠺࡟࡞ࡾࠊࡉࡽ࡟ࡑࢀࢆά⏝ࡍࡿືࡁࡀάⓎ໬ࡋ࡚࠸ࡿ㸬୍᪉ࠊᐇᴗົ࡟࠾࠸ ࡚㡢ኌᢏ⾡ࢆά⏝ࡍࡿ࡟ࡣࠊ⌧ሙࡢ⤒㦂࡟ᇶ࡙࠸ࡓࡁࡵ⣽࠿࠸㓄៖ࡸᕤኵࡀᚲせ࡛࠶ࡿ㸬ᙜ♫ࡣ㏻ヰ㘓㡢࣭㡢ኌㄆ㆑࣭ ᑐヰศᯒ࣭㠀ゝㄒࡸࣃࣛゝㄒ᝟ሗࡢศᯒ࡜࠸ࡗࡓᢏ⾡ࢆࢥ࣮ࣝࢭࣥࢱ࣮ࡢᴗົᨵၿ࡟ᛂ⏝ࡋࠊ㡢ኌᢏ⾡ࡢά⏝ࡢሙࢆ ᗈࡆࡼ࠺࡜ࡋ࡚࠸ࡿ㸬ᮏⓎ⾲࡛ࡣࠊ஦౛ࢆ⤂௓ࡋ࡞ࡀࡽࠊࢥ࣮ࣝࢭࣥࢱ࣮ᴗ⏺࡟ᚲせ࡜ࡉࢀࡿ㡢ኌᢏ⾡ࡢලయീࢆ᥈ ࡾࡓ࠸㸬. The use of speech technologies in call centers - Including para- and non-linguistic information. ATSUHIRO SAKURAI†1 SHINTA KIMURA†. Call centers constitute an important arena for the use of speech processing technology. Moreover, an increasing number of companies perform voice logging, resulting in the accumulation of a large amount of speech data, paving the way for new applications of such resources. On the other hand, the use of speech technology in practice is not an easy task, requiring careful attention and know-how. Our company specializes in the use of speech technologies ranging from voice logging, speech recognition, dialog analysis, and para- or non-linguistic analysis to create new solutions for the call center industry. In this report, we show some examples and try to grasp what is really expected from speech technologies in the call center industry.. 1. ࡣࡌࡵ࡟. 㝈࡟ά࠿ࡍࡇ࡜ࡀ㘽࡜࡞ࡿ㸬ࡲࡓࠊ࡝ࡢࡼ࠺࡞ࢯ࣮ࣜࣗࢩ. ࢥ࣮ࣝࢭࣥࢱ࣮࡛ࡣ㡢ኌࡀ᭱ࡶ㔜せ࡞ࢥ࣑ࣗࢽࢣ࣮ࢩࣙ. ᴗ⏺ࡢ⌧≧ࡸㄢ㢟ࢆᢕᥱࡍࡿᚲせࡀ࠶ࡿ㸬ᮏⓎ⾲࡛ࡣࠊࢥ. ࣥᡭẁ࡛࠶ࡿࡓࡵࠊ㡢ኌ᝟ሗฎ⌮ᢏ⾡㸦௨ୗ㸸㡢ኌᢏ⾡㸧. ࣮ࣝࢭࣥࢱ࣮ࡢ⌧≧ࡸㄢ㢟ࢆᢕᥱࡋࡓୖࠊ㡢ኌᢏ⾡ࡢά⏝. ࢆᑟධࡍࡿ࣓ࣜࢵࢺࡣ኱ࡁ࠸࡜ࡉࢀ࡚࠸ࡿ㸬ࢥ࣮ࣝࢭࣥࢱ. ࡢලయ౛ࢆ⤂௓ࡋ࡞ࡀࡽ௒ᚋࡢᒎ㛤࡟ࡘ࠸࡚⪃࠼ࡿ㸬. ࣮࡟࠾ࡅࡿ㡢ኌᢏ⾡࡜࠸࠼ࡤࠊ㡢ኌㄆ㆑࡜㡢ኌྜᡂࢆ⤌ࡳ. ࣙࣥࡀ㐺ษ࡛࠶ࡿ࠿ぢᴟࡵࡿࡓࡵࠊࢥ࣮ࣝࢭࣥࢱ࣮࡜࠸࠺. ྜ ࢃ ࡏ ࡓ 㡢 ኌ ⮬ ື ᛂ ⟅ ࢩ ࢫ ࢸ ࣒ 㸦 IVR: interactive voice. 2. ࢥ࣮ࣝࢭࣥࢱ࣮ࡢ⌧≧ࡸㄢ㢟. response system㸧ࢆ㐃᝿ࡋࡸࡍ࠸ࡀࠊ᪤Ꮡࡢ IVR ࢩࢫࢸ࣒. ࢥ࣮ࣝࢭࣥࢱ࣮ࡣ㢳ᐈ࡬ࡢ㟁ヰᑐᛂᴗົࢆ⾜࠺⤌⧊࡛࠶. ࡟࠾ࡅࡿ㡢ኌㄆ㆑ࡢ฼⏝ࡣ㧗ᗘ࡞ᛂᑐࢆᚲせ࡜ࡋ࡞࠸ᴗົ. ࡿ㸬ࡑࡢᙺ๭ࡣᴗ✀࡟ࡼࡗ࡚␗࡞ࡿࡀࠊᇶᮏⓗ࡟௨ୗࡢᶵ. ࡟㝈ᐃࡉࢀࡿ㸬ࡲࡓࠊ㡢ኌྜᡂࡣᅛᐃᩥࡢ࢜ࣇࣛ࢖ࣥ⏕ᡂ. ⬟ࢆᣢࡘࡇ࡜ࡀከ࠸㸬. ࡜࠸࠺฼⏝᪉ἲࡀ኱ໃࢆࡋࡵࡿ㸬ᐇ㝿ࠊ㡢ኌᢏ⾡࡜⮬↛ゝ. z. ஦ົᡭ⥆ࡁ࡞࡝ࡢཷ௜ࡸฎ⌮. ㄒฎ⌮ࢆ⤌ࡳྜࢃࡏࡓᑐヰࢩࢫࢸ࣒ࡀࢥ࣮ࣝࢭࣥࢱ࣮ᴗ⏺. z. 㢳ᐈ࠿ࡽࡢࢡ࣮࣒࣭ࣞせᮃ࣭ၥ࠸ྜࢃࡏ࡞࡝ࡢཷ௜ࡸ. ࡛ᬑཬࡍࡿࡢ࡟ࡲࡔ᫬㛫ࡀ࠿࠿ࡿ࡜ᛮࢃࢀࡿ㸬 ࡋ࠿ࡋ࡞ࡀࡽࠊ㡢ኌᢏ⾡ࡢά⏝ࡢሙࡣ IVR ௨እ࡟ᵝࠎ࡞. ࡑࢀ࡟ᑐࡍࡿㄝ᫂࡞࡝ z. ศ㔝࡟ཬࡪ㸬IVR ࡣே㛫ᑐ࣐ࢩࣥࡢࣜ࢔ࣝࢱ࢖࣒࢖ࣥࢱࣛ. ࢡࣟࢫࢭࣝ࡞࡝⬟ືⓗ࡞ࢭ࣮ࣝࢫࡸ࣐࣮ࢣࢸ࢕ࣥࢢ άື. ࢡࢩࣙࣥࢆᐇ⌧ࡋࡼ࠺࡜ࡋ࡚࠸ࡿࡀࠊே㛫ᑐே㛫ࠊࡶࡋࡃ. ௻ᴗάື࡟࠾ࡅࡿࢥ࣮ࣝࢭࣥࢱ࣮ࡢᙺ๭ࡣᴟࡵ࡚኱ࡁ. ࡣ㠀ࣜ࢔ࣝࢱ࢖࣒㸦࢜ࣇࣛ࢖ࣥ㸧ศᯒ࡜࠸ࡗࡓ㡿ᇦ࡟ࡶᙜ. ࠸ࡇ࡜ࡣ࠸࠺ࡲ࡛ࡶ࡞࠸㸬ࢥ࣮ࣝࢭࣥࢱ࣮ࡣ௻ᴗ࡟࡜ࡗ࡚. ↛࡞ࡀࡽ㡢ኌᢏ⾡ࢆ฼⏝࡛ࡁࡿ㸬ࡲࡓࠊ㡢ኌྜᡂࡸ㡢ኌㄆ. 㢳ᐈ࡜᥋ࡍࡿ㈗㔜࡞ࢳࣕࣥࢿ࡛ࣝ࠶ࡾࠊ௻ᴗࡢࠕ㢦ࠖࡑࡢ. ㆑௨እ࡟ࡶࣃࣛゝㄒ࣭㠀ゝㄒ᝟ሗࡢ฼⏝ࡶྍ⬟࡛࠶ࡿ㸬. ࡶࡢ࡜࠸࠼ࡿ[a]㸬ࡲࡓࠊ㏆ᖺࠊ௨ୗࡢ஦᝟࡟ࡼࡾࢥ࣮ࣝࢭ. ࡓࡔࡋࠊᐇ⏝ⓗ࡞ࢯ࣮ࣜࣗࢩࣙࣥࢆᵓ⠏ࡍࡿ࡟࠶ࡓࡾࠊ ⢭ᗘࡀ୙᏶඲࡛࠶ࡿ࡜࠸࠺㡢ኌᢏ⾡ࡢᮏ㉁ⓗ࡜ࡶ࠸࠺࡭ࡁ 㝈⏺ࢆ⌮ゎࡋࠊᶵᲔࡢᣢࡘ≉㛗㸦㏿࠸ࠊ⑂ࢀ࡞࠸㸧ࢆ᭱኱ † ᰴᘧ఍♫࢔ࢽࣔ㸸http://www.animo.co.jp Animo Limited: http://www.animo.co.jp/EN/. ‫ף‬2013 Information Processing Society of Japan. ࣥࢱ࣮ࡢ㔜せᛶࡣࡲࡍࡲࡍ㧗ࡲࡾࡘࡘ࠶ࡿ㸬 z. ἲⓗせᅉ㸦㔠⼥⏺ࡢሙྜࡣ㔠⼥ၟရྲྀᘬἲ࡛ᐃࡵࡽࢀ ࡿ㔠⼥ၟရ࡟㛵ࡍࡿㄝ᫂⩏ົ㸧࡟ࡼࡾࠊ㏻ヰ㘓㡢ࢆᑟ. a) ᭱㏆ࡣ e-mail ࡸ SNS ࡞࡝᪂ࡋ࠸ࢱ࢖ࣉࡢ࢖ࣥࢱࣛࢡࢩࣙࣥࡀᬑཬࡋ࡚ ࠸ࡿ㸬ࡑࡇ࡛ࠊࢸ࢟ࢫࢺ࡟ࡼࡿࢥ࣑ࣗࢽࢣ࣮ࢩࣙࣥ࡟ࡶᑐᛂࡍࡿ⤌⧊࡜ࡋ ࡚ࠕࢥࣥࢱࢡࢺࢭࣥࢱ࣮ࠖ࡜࠸࠺࿧⛠ࡀᐃ╔ࡋࡘࡘ࠶ࡿ㸬. 1.

(2) Vol.2013-SLP-98 No.2 2013/10/25. ᝟ሗฎ⌮Ꮫ఍◊✲ሗ࿌ IPSJ SIG Technical Report ධࡍࡿ௻ᴗࡀከࡃ࡞ࡗ࡚ࡁ࡚࠸ࡿ㸬ࢫࢺ࣮ࣞࢪࡢ౯᱁. ⾲㸯 ࣮࢜࣌ࣞࢱᨭ᥼ࢩࢫࢸ࣒ࠕKnowledgeNaviࠖࡢࢥࣥ. ࡀୗⴠഴྥ࡟࠶ࡿࡇ࡜ࡶࡑࢀࢆᚋᢲࡋࡍࡿ㸬 z. ࢭࣉࢺ㸬. ㏻ヰ㘓㡢ࡢᬑཬ࡟క࠸ࠊ኱㔞ࡢ㏻ヰ㘓㡢ࢹ࣮ࢱࡀ⵳✚. ࠕࢹࢪࢱ࣓ࣝ࢝ࣛࠖ࡜ࠕᐮ࠸ࠖ࡜࠸࠺࣮࣮࢟࣡ࢻࡢ⤌ࡳྜ. ࡉࢀࡿࡼ࠺࡟࡞ࡗࡓ㸬ࡑࡢࡓࡵࠊ࣐࣮ࢣࢸ࢕ࣥࢢ᝟ሗ. ࢃࡏ࡛ࠕࢹࢪࢱ࣓ࣝ࢝ࣛࡢᐮ෭ᆅ࡛ࡢ౑⏝ࠖ࡜࠸࠺࣐ࢽࣗ ࢔ࣝࡀ࣏ࢵࣉ࢔ࢵࣉࡍࡿ㸬. ࡜ࡋ࡚㘓㡢ࢹ࣮ࢱࢆά⏝ࡋࡼ࠺࡜࠸࠺ືࡁࡀ࠶ࡿ㸬. Table 1 KnowledgeNavi agent support system concept. ࢥ࣮ࣝࢭࣥࢱ࣮࡟ࡣ㸯㸧ࢥࢫࢺ๐ῶࠊ㸰㸧ࢧ࣮ࣅࢫྥ ୖࠊ㸱㸧᭦࡞ࡿ௜ຍ౯್ࡢ㏣ồ࡜࠸࠺ᇶᮏⓗ࡞┠ᶆࡀ࠿. Keywords “digital camera” and “cold” within the conversation lead to the manual “Using the digital camera in cold areas”.. ࠿ࡆࡽࢀࡿࡀࠊୖグࡢᶵ⬟ࡸ஦᝟ࢆ⫼ᬒ࡟ࠊ┠ᶆࢆ㐩ᡂ ࡍࡿࡓࡵࡢㄢ㢟࡜ࡋ࡚௨ୗࡢ㡯┠ࡀ࠶ࡆࡽࢀࡿ㸬 ձ ࣮࢜࣌ࣞࢱసᴗࡢຠ⋡໬ ղ ࣮࢜࣌ࣞࢱࡢホ౯࣭ᩍ⫱. ᶵ⬟䠍. ᶵ⬟䠎. ᐮ෭ᆅ. ᶵ⬟䠏. 〇ရ䠝〇ရ䠞䝕䝆䝍䝹䜹䝯䝷〇ရ䠟䈈. ճ ⫈ྲྀ࣭ࣔࢽࢱࣜࣥࢢసᴗࡢຠ⋡໬ ࣮࢜࣌ࣞࢱࡢᛂᑐࡢᵝᏊࢆ」ᩘࡢᩘ್࡛⾲⌧࡛ࡁࢀࡤࠊ. մ ㏻ヰࢹ࣮ࢱࡢᒓᛶ௜୚࣭ศ㢮 յ ㏻ヰෆᐜࡢศᯒ࣭ά⏝࡞࡝. ලయⓗ࡞ᇶ‽್࡟ᇶ࡙ࡃホ౯ᑻᗘࡢタᐃࡀྍ⬟࡟࡞ࡿ㸬㐠. ḟ⠇௨㝆ࠊୖグࡢྛ┠ⓗ࡟࠾࠸࡚㡢ኌᢏ⾡ࢆ࡝ࡢࡼ࠺࡟. Ⴀ᪉ἲ࡜ࡋ࡚ࠊᚑ᮶ࡢ୺ほホ౯ࢆ⿬௜ࡅࡿࡓࡵࡢᇶᮏࢹ࣮. ά⏝࡛ࡁࡿ࠿ࠊ஦౛ࢆ⤂௓ࡋ࡞ࡀࡽㄝ᫂ࡍࡿ㸬. ࢱ࡜ࡋ࡚⮬ື࡛⟬ฟࡋࡓᩘ್ࢆ⏝࠸ࡿࡇ࡜ࡀ࡛ࡁࡿ㸬. 3. 㡢ኌᢏ⾡ࡢ౑⏝౛ 3.1 ࣮࢜࣌ࣞࢱసᴗࡢຠ⋡໬ ᑓ㛛ⓗ࡞ㄝ᫂ࡀᚲせ࡞〇ရࢆᩘከࡃྲྀࡾᢅ࠺௻ᴗࡢࢥ. ᙜ♫ࡣ࠿ࡡ࡚࠿ࡽᑐヰࡢὶࢀࢆ⾲ࡍᩘ್࡟ᇶ࡙ࡃホ౯ ᑻᗘࢆᥦ౪ࡍࡿࢧ࣮ࣅࢫࢆᒎ㛤ࡋ࡚ࡁࡓࡀࠊ᭱㏆ࡣ⮬♫ࡢ ㏻ヰ㘓㡢ࢩࢫࢸ࣒࡟ホ౯ᣦᶆ⏝ࡢᩘ್ࡢ⟬ฟࢆ⾜࠺ࣉࣛࢢ ࢖ࣥࡢᥦ౪ࢆጞࡵࡓ[c]㸬. ࣮ࣝࢭࣥࢱ࣮࡛ࡣ࣮࢜࣌ࣞࢱࡀ⮬ࡽ࣐ࢽࣗ࢔ࣝ㢮ࢆ᳨⣴ࡋࠊ. ୍⯡ⓗ࡟ࠊ࣮࢜࣌ࣞࢱࡢホ౯࡟᭷ຠ࡞ᣦᶆ࡜ࡋ࡚௨ୗ࡟. ࡑࢀ࡟ᇶ࡙࠸࡚㢳ᐈ࡟ᑐࡍࡿㄝ᫂ࢆ⾜࠺ࡀࠊ᳨⣴࡟࠿࠿ࡿ. ♧ࡍ㏻ࡾࠊⓎኌෆᐜ㸦ゝㄒ᝟ሗ㸧࡟ᇶ࡙ࡃᣦᶆ࡜ࡑࢀ௨እ. ᫬㛫ࡣ࣮࢜ࣂ࣮࣊ࢵࢻ࡜࡞ࡿࡓࡵࠊ࡞ࡿ࡭ࡃ▷⦰ࡍࡿᚲせ. ࡢ᝟ሗ㸦ᑐヰཬࡧࣃࣛゝㄒ࣭㠀ゝㄒ᝟ሗ㸧࡟ᇶ࡙ࡃᣦᶆࡀ. ࡀ࠶ࡿ㸬ࡲࡓࠊࢱ࢖ࣆࣥࢢ㡢ࡀ㢳ᐈ࡟⪺ࡇ࠼࡚ࡋࡲ࠺࡜ᝏ. ࠶ࡿ㸬௨ୗࠊࡑࢀࡒࢀ࡟ࡘ࠸࡚㏙࡭ࡿ㸬. ࠸༳㇟ࢆ୚࠼ࡿࡔࡅ࡛࡞ࡃࠊ఍ヰࡢὶࢀࢆጉࡆࡿ࡜࠸࠺ၥ. (1) Ⓨኌෆᐜ࡟ᇶ࡙ࡃ࣮࢜࣌ࣞࢱホ౯ᣦᶆ. 㢟ࡀ࠶ࡿ㸬ࡑࡇ࡛ࠊ࣮࢜࣌ࣞࢱ࡟ࡼࡿ࣐ࢽࣗ࢔ࣝ㢮ࡢ᳨⣴. Ⓨኌෆᐜ࡟ᇶ࡙ࡃホ౯ᣦᶆ࡜ࡋ࡚ࠊ㏻ヰࡢྛẁ㝵㸦࣮࢜. సᴗࢆᨭ᥼ࡍࡿࡓࡵࠊ㡢ኌㄆ㆑ᢏ⾡ࢆ฼⏝࡛ࡁࡿ㸬ᙜ♫ࡣ. ࣉࢽࣥࢢࠊ࠾♩㸭࠾ルࡧࠊၥデࠊᮏே☜ㄆࠊࢡ࣮ࣟࢪࣥࢢ㸧. ≉ᐃ࣮࣮࢟࣡ࢻ࡟࣐ࢽࣗ࢔ࣝ㢮ࢆࣜࣥࢡࡉࡏࠊ࣮࣡ࢻࢫ࣏. ࡟ゝ࠺࡭ࡁゝⴥ㸦OK ࣮࣡ࢻ㸧ࢆⓎኌࡋࡓ࠿ࠊࡲࡓࡣ⚗ࡌ. ࢵࢸ࢕ࣥࢢᆺࡢ㡢ኌㄆ㆑࢚ࣥࢪࣥࢆ฼⏝ࡋ᳨࡚⣴సᴗࢆᨭ. ࡽࢀࡓゝⴥࡸ⾲⌧㸦NG ࣮࣡ࢻ㸧ࢆⓎኌࡋ࡚࠸࡞࠸࠿࡞࡝. ᥼ࡍࡿࢯࣇࢺ࢙࢘࢔ࢆ〇ရ໬ࡋ࡚࠸ࡿ[ b]㸬ࡇࢀࢆ⏝࠸࡚ࠊ. ࡜࠸ࡗࡓࢳ࢙ࢵࢡ㡯┠ࡀ࠶ࡿ㸬ࡇࡢࡼ࠺࡞ࢱࢫࢡ࡟ࡣ㡢ኌ. ࣮࢜࣌ࣞࢱࡣᑐᛂ᱌௳࡟㛵㐃ࡍࡿ࣮࣮࢟࣡ࢻࢆ᚟ၐࡍࡿࡇ. ㄆ㆑ᢏ⾡ࢆά⏝࡛ࡁࡿࡀࠊᙜ♫࡛ࡣ࣮࣮࢟࣡ࢻࡢ᳨⣴࡟≉. ࡜࡛ࠊ఍ヰࡢὶࢀࢆ㏵ษࢀࡎ࡟ࢻ࣓࢟ࣗࣥࢺࡢ᳨⣴ࢆ⾜࠺. ໬ࡋࡓ࣮࣡ࢻࢫ࣏ࢵࢸ࢕ࣥࢢᆺ㡢ኌㄆ㆑ࢯࣇࢺ࢙࢘࢔[ d ]. ࡇ࡜ࡀ࡛ࡁࡿ㸬ࡲࡓࠊᡭືධຊࡢᚲせᛶࡀῶࡾࠊࢱ࢖ࣆࣥ. ࡛ࡑࢀࢆᐇ⌧ࡍࡿ㸬௨ୗࡢ஦౛ࡣࠊ࠶ࡿఫᏯタഛ࣓࣮࣮࢝. ࢢ㡢ࡢၥ㢟ࡶ㍍ῶࡉࢀࡿ㸬⾲㸯࡟ࢩࢫࢸ࣒ࡢࢥࣥࢭࣉࢺࢆ. ࡢࢥ࣮ࣝࢭࣥࢱ࣮ࡢ࣮࢜࣌ࣞࢱࢆᑐ㇟࡟⾜ࡗࡓⓎゝ࣮࢟࣡. ♧ࡍ㸬୍ࡘࡢ㛵㐃࣮࣮࢟࣡ࢻࢆⓎኌࡍࡿ࡜」ᩘࡢࢻ࣓࢟ࣗ. ࣮ࢻࡢㄪᰝ࡛࠶ࡿ㸬࣮࢜࣌ࣞࢱࡢ໅⥆ᖺᩘ࡟ᛂࡌ࡚࣋ࢸࣛ. ࣥࢺࡀࣁ࢖ࣛ࢖ࢺࡉࢀࠊ㏣ຍࡢ࣮࣮࢟࣡ࢻ࡛ࢹ࣮ࢱࢆࡉࡽ. ࣭ࣥ୰ሀ࣭᪂ே࡟ศࡅ࡚ㄪᰝࢆ⾜ࡗࡓ㸬ᅗ㸯ࡣ OK ࣮࣡ࢻ. ࡟⤠ࡿ㸬. 㸭NG ࣮࣡ࢻࡢㄪᰝ⤖ᯝࡢ୍㒊࡛࠶ࡿ㸬⌧ሙ࡛ࡢࣄ࢔ࣜࣥ. 3.2 ࣮࢜࣌ࣞࢱࡢホ౯࣭ᩍ⫱. ࢢࡢᚋࠊୗグࡢࡼ࠺࡞⤖ㄽ࡟⮳ࡗࡓ㸬. ࣮࢜࣌ࣞࢱࡢホ౯ࡸᩍ⫱ࡣࢥ࣮ࣝࢭࣥࢱ࣮࡟࠾࠸࡚㔜. z. ᪂ே㸸ゝⴥ౑࠸ࡢᩍ⫱ࡣẚ㍑ⓗࡼࡃ࡛ࡁ࡚࠸ࡿ. せ࡞ㄢ㢟࡛࠶ࡿ㸬⌧ᅾࠊࢫ࣮ࣃ࣮ࣂ࢖ࢨ̿㸦ୖྖ㸧࡟ࡼࡿ. z. ࣋ࢸࣛࣥ㸸OK ࣮࣡ࢻⓎゝࡢᇶ‽ࢆ‶ࡓࡋ࡚࠸ࡿࡀࠊ. ୺ほホ౯ࡸᨃఝ㢳ᐈ㸦ホ౯⪅㸧࡟ࡼࡿࠕ࣑ࢫࢸ࣮ࣜࢥ࣮ࣝࠖ. NG ࣮࣡ࢻࡢⓎኌࡀከ࠸㸦័ࢀ࠿ࡽ᮶ࡿ␎ㄒࡢ฼⏝࡞. ࡜࠸࠺ホ౯ᡭἲࡀ୺ὶ࡛࠶ࡿࡀࠊࡼࡾᐈほⓗ࡞ィ㔞ᇶ‽⩌. ࡝㸧. (KPI: key performance indicator)[1]ࡶᚲせ࡜ࡉࢀࡿ㸬. b) ࣜ࢔ࣝࢱ࢖࣒ࢼࣞࢵࢪ᳨⣴ࢩࢫࢸ࣒ KnowledgeNavi. ‫ף‬2013 Information Processing Society of Japan. c) ࣮࢜࣌ࣞࢱホ౯ࢩࢫࢸ࣒ VoiceTracking Quality Manager d) ࣮࣮࢟࣡ࢻ᳨⣴ࢩࢫࢸ࣒ VoiceTracking KeywordFinder. 2.

(3) Vol.2013-SLP-98 No.2 2013/10/25. ᝟ሗฎ⌮Ꮫ఍◊✲ሗ࿌ IPSJ SIG Technical Report. ⾲ 2 ࣮࢜࣌ࣞࢱホ౯ᣦᶆ࡜ࡋ࡚฼⏝ࡉࢀࡿᑐヰศᯒࣃ. 0.8. ࣓࣮ࣛࢱ. 0.7 0.6. Table 2 Dialog analysis parameters used for evaluating. 0.5 䝧䝔䝷䞁 0.4. ୰ሀ. 0.3. ᪂ே. 0.2. operator’s performance ᑐヰศᯒࣃ࣓࣮ࣛࢱ. ព࿡. Ⓨゝ㔞. ࣮࢜࣌ࣞࢱ࣭㢳ᐈࡑࢀࡒࢀ. 0.1. ࡢⓎゝ㔞ཬࡧࡑࡢ๭ྜ࡞࡝. 0 䛣䜣䛻䛱䛿. 䛒䜚䛜䛸䛖䛤䛦䛔䜎䛩. 䛚䛶䛩䛖䛷䛩䛜. ࠿ࡪࡾࡢᅇᩘ࣭㢖ᗘ. Ⓨゝ୰࡟๭ࡾ㎸ࡴ⾜Ⅽ㸦࠿. (a) OK ࣮࣡ࢻ. ࡪࡾ㸧ࡢᩘ್໬ ࠶࠸࡙ࡕࡢᅇᩘ࣭㢖ᗘ. ࠕ⪺ࡁୖᡭࠖࡢᣦᶆ. ヰ㏿. 㢳ᐈ࡟ྜࢃࡏࡓⓎヰࢫࣆ࣮ ࢻ࡟࡞ࡗ࡚࠸ࡿ࠿. 0.45 0.4 0.35 0.3 8. 䝧䝔䝷䞁. 0.25. ୰ሀ. 0.2. ᪂ே. 4 3. 0.1. 25. 1. 20. 0.8. 15. 0.6. 10. 0.4. 5. 0.2. 6 5. 0.15. 1.2. 30. 7. 2 1. 0.05. 0. 0 䝧䝔䝷䞁. 0 䛭䜣䛺䜒䛾. ⌧ㄪ㈝. (b) NG ࣮࣡ࢻ. ୰ሀ. ᪂ே. (a)ᖹᆒ㏻ヰ᫬㛫䠄ศ䠅. 0 䝧䝔䝷䞁. ୰ሀ. ᪂ே. (b)ᖹᆒ↓㡢๭ྜ䠄䠂䠅. 䝧䝔䝷䞁. ୰ሀ. ᪂ே. (c)ᖹᆒ䠫䠬䛛䜆䜚ᅇᩘ 䠄䠍ศ䛒䛯䜚䠅. ᅗ 2 ᑐヰศᯒ࡟ࡼࡿ࣮࢜࣌ࣞࢱࡢホ౯ Figure 2 Agent evaluation based on dialog analysis.. ᅗ 1 ࣮࣮࢟࣡ࢻࡢⓎゝ࡟ࡼࡿ࣮࢜࣌ࣞࢱホ౯㸬 ⦪㍈ࡣ㏻ヰ࠶ࡓࡾࡢᖹᆒⓎヰᅇᩘ࡛࠶ࡿ㸬 Figure 1 Agent evaluation based on keyword utterance.. 3.3 ⫈ྲྀ࣭ࣔࢽࢱࣜࣥࢢసᴗࡢຠ⋡໬ 㘓㡢ࡉࢀࡓ㡢ኌࢆ⫈ྲྀࡋࠊ㏻ヰෆᐜࢆ☜ㄆࡍࡿࡇ࡜ࡣࢥ. The vertical axis represents the average number of utterances. ࣮ࣝࢭࣥࢱ࣮࡛᪥ᖖⓗ࡟⾜ࢃࢀࡿసᴗ࡛࠶ࡿ㸬సᴗࡢຠ⋡. per call.. ໬ࢆࡣ࠿ࡿࡓࡵ࡟ᵝࠎ࡞㡢ኌᢏ⾡ࡀ฼⏝ࡉࢀࡿࡀࠊࡇࡇ࡛ ௦⾲ⓗ࡞౛ࢆ࠶ࡆࡿ㸬. (2) ᑐヰཬࡧࣃࣛゝㄒ࣭㠀ゝㄒ᝟ሗ࡟ᇶ࡙ࡃ࣮࢜࣌ࣞࢱホ ౯ᣦᶆ. (1) ヰ㏿ኚ᥮ ேᡭ࡛㏻ヰࡢෆᐜࢆ⫈ྲྀࡍࡿሙྜࠊᮏ᮶ࡣᑡ࡞ࡃ࡜ࡶ㏻. ከࡃࡢ㏻ヰ㘓㡢ࢩࢫࢸ࣒ࡣ㏻ヰࡢୖࡾ࣭ୗࡾࢆูࠎࡢࢳ. ヰ᫬㛫࡜ྠ➼ࡢ᫬㛫ࡀᚲせ࡛࠶ࡿ㸬ࡋ࠿ࡋ࡞ࡀࡽࠊヰ㏿ኚ. ࣕࣥࢿ࡛ࣝ㘓㡢ࡍࡿࢫࢸࣞ࢜㘓㡢᪉ᘧࢆ᥇⏝ࡋ࡚࠸ࡿ㸬ࡇ. ᥮ᢏ⾡㸦ᇶᮏ࿘Ἴᩘࡸࢫ࣌ࢡࢺࣝໟ⤡ࢆಖࡗࡓࡲࡲࠊ᫬㛫. ࡢࡼ࠺࡞ࢩࢫࢸ࣒ࢆ฼⏝ࡍࢀࡤ࣮࢜࣌ࣞࢱ࡜㢳ᐈࡢⓎኌࢆ. ࢫࢣ࣮ࣝࢆኚ᭦ࡍࡿࡓࡵࡢᢏ⾡㸧ࢆ฼⏝ࡍࢀࡤసᴗ᫬㛫ࢆ. ศ㞳ࡋ࡚ศᯒ࣭ホ౯࡛ࡁࡿ㸬ࡑࢀࡒࢀࡢⓎኌࡢࢱ࢖࣑ࣥࢢ. ▷⦰࡛ࡁࡿ㸬ヰ㏿ኚ᥮ࡢ᪉ἲ࡜ࡋ࡚ᵝࠎ࡞ᡭἲࡀᥦ᱌ࡉࢀ. ࡀࢃ࠿ࢀࡤࠊ࣮࢜࣌ࣞࢱࡢホ౯࡟ᙺ❧ࡘᵝࠎ࡞ᣦᶆࡀᚓࡽ. ࡚࠸ࡿࡀࠊᙜ♫࡛ࡣ⊂⮬ࡢ࢔ࣝࢦࣜࢬ࣒ࢆ฼⏝ࡋ࡚࠸ࡿ㸬. ࢀࡿ㸬⾲ 2 ࡟ࡣ࣮࢜࣌ࣞࢱホ౯࡛฼⏝ࡉࢀ࡚࠸ࡿ୺࡞ศᯒ. (2) 㡢ኌㄆ㆑. ࣃ࣓࣮ࣛࢱࢆ♧ࡍ㸬. ㏻ヰෆᐜࢆணࡵ኱ࡲ࠿࡟ᢕᥱ࡛ࡁࢀࡤࠊ⫈ྲྀసᴗࢆ኱ᖜ. ᅗ 2 ࡟๓グࡢఫᏯタഛ࣓࣮࣮࢝ࡢࢥ࣮ࣝࢭࣥࢱ࣮ࡢศᯒ. ࡟ຠ⋡໬࡛ࡁࡿྍ⬟ᛶࡀ࠶ࡿࡀࠊࡑ࠺࠸ࡗࡓ⏝㏵࡟ࡣ㡢ኌ. ⤖ᯝࢆ♧ࡍ㸬ࡇࡇࡶྠࡌࡃ⤖ᯝࢆ࣋ࢸ࣭ࣛࣥ୰ሀ࣭᪂ே࢜. ㄆ㆑ᢏ⾡ࡣ᭷ຠ࡛࠶ࡿ㸬౛࠼ࡤ࣮࣡ࢻࢫ࣏ࢵࢸ࢕ࣥࢢᆺࡢ. ࣮࣌ࣞࢱ࡟ศࡅࡓ㸬ࡇࡢࢹ࣮ࢱࡣ᭷ពᕪ࡟ᇶ࡙ࡃཝᐦ࡞⤖. 㡢ኌㄆ㆑ฎ⌮ࢆ᪋ࡋࠊ㔜せ࡞࣮࣮࢟࣡ࢻ࡟ࡋ࠾ࡾ㸦ࢱࢢ㸧. ㄽࢆᑟࡃࡓࡵࡢࡶࡢ࡛ࡣ࡞࠸ࡀࠊ୺ほ࡟ࡼࡿ▱ぢࢆㄝ᫂ࡍ. ࢆ௜୚ࡍࢀࡤࠊ⫈ྲྀసᴗࡢຠ⋡໬ࢆࡣ࠿ࢀࡿ㸬ࡑࡢࡼ࠺࡞. ࡿࡓࡵࡢ⿬௜ࡅࢹ࣮ࢱ࡜ࡋ࡚฼⏝࡛ࡁࡿ㸬౛࠼ࡤࠊ᪂ே࢜. ᶵ⬟ࡣᙜ♫ࡢ㏻ヰ㘓㡢ࢯࣇࢺ࢙࢘࢔ࡢࣉࣛࢢ࢖ࣥ࡜ࡋ࡚ᥦ. ࣮࣌ࣞࢱࡣ▷᫬㛫࡛㏻ヰࢆ⤊ࢃࡽࡏࡼ࠺࡜ࡍࡿࡀࠊ⤒㦂ࢆ. ౪ࡋ࡚࠸ࡿ㸬. ✚ࡴ࡟ࡘࢀࠊ㢳ᐈ࡜఍ヰ࡛ࡁࡿࡼ࠺࡟࡞ࡿࡇ࡜ࡸ᪂ேࡢᖹ ᆒ↓㡢๭ྜཬࡧࠕ࠿ࡪࡾࠖࡢᅇᩘࡀ㧗࠸ࡇ࡜࡞࡝ࡣ⌧ሙࡢ ᖖ㆑࡜୍⮴ࡍࡿ⤖ᯝ࡛࠶ࡿ㸬. ‫ף‬2013 Information Processing Society of Japan. 3.

(4) Vol.2013-SLP-98 No.2 2013/10/25. ᝟ሗฎ⌮Ꮫ఍◊✲ሗ࿌ IPSJ SIG Technical Report (3) ⫈ྲྀᑐ㇟ࢹ࣮ࢱࡢ⤠ࡾ㎸ࡳ. 㸦⫈ྲྀసᴗ㸧ࡢຠ⋡໬ࢆࡣ࠿ࢀࡿ㸬. ⫈ྲྀ࣭ࣔࢽࢱࣜࣥࢢࡢᑐ㇟࡜࡞ࡿ㏻ヰࢹ࣮ࢱࡑࡢࡶࡢࢆ. ๓㏙ࡢ㏻ࡾࠊ㡢ኌ࡟ࡣゝㄒ᝟ሗࡢ࡯࠿࡟ࣃࣛゝㄒ࣭㠀ゝ. ஦๓࡟㑅ูࡍࡿࡇ࡜ࡀ࡛ࡁࢀࡤసᴗࡢຠ⋡໬࡟ࡘ࡞ࡀࡿ㸬. ㄒ᝟ሗࡀྵࡲࢀࡿࡀࠊ ࠕၥ㢟࿧ࠖ࡜࠸࠺⌧㇟ࡣឤ᝟࡜῝ࡃ㛵. ࡇࢀࡣ㏻ヰࢹ࣮ࢱ࡟ఱࡽ࠿ࡢᒓᛶࢆ௜୚ࡋࠊࡑࢀ࡟ᇶ࡙࠸. ࢃࡗ࡚࠸ࡿࡓࡵࠊࣃࣛゝㄒ࣭㠀ゝㄒ᝟ሗࡢᙳ㡪ࡣ኱ࡁ࠸࡜. ࡚ศ㢮ࡍࡿࡇ࡜࡟➼ࡋ࠸ࡀࠊᒓᛶ௜୚ࡸศ㢮ࡢ⏝㏵ࡣ⫈ྲྀ. ࠸࠼ࡿ㸬ࡑࡢᙳ㡪ࡣ㡩ᚊⓗ≉ᚩࡸኌ㉁࡞࡝㡢㡪ⓗ≉ᚩࡢኚ. సᴗࡢຠ⋡໬࡟㝈ࡽࡎࠊ࣐࣮ࢣࢸ࢕ࣥࢢㄪᰝࡸ௻ᴗࢥࣥࣉ. ໬࡜ࡋ࡚⾲ࢀࡿࡀࠊឤ᝟ࡣ࠶ࡃࡲ࡛ෆ࡞ࡿᴫᛕ࡛࠶ࡿࡓࡵࠊ. ࣛ࢖࢔ࣥࢫ⟶⌮࡞࡝ᵝࠎ࡞ศ㔝࡟㐺⏝࡛ࡁࡿࡓࡵࠊḟ⠇࡛. 㡢㡪ⓗ≉ᚩ㔞࡜ࡢ㛵ಀࢆ᥎ᐃ࣭ᐃᘧ໬ࡍࡿࡇ࡜ࡣ㞴ࡋ࠸㸬. ྲྀࡾୖࡆࡿ㸬. ࡑࡢࡓࡵࠊᙜ♫࡛ࡣឤ᝟ࡢ᥎ᐃࢆヨࡳࡿࡢ࡛ࡣ࡞ࡃࠊኌࡢ. 3.4 ㏻ヰࢹ࣮ࢱࡢᒓᛶ௜୚࣭ศ㢮. ≀⌮ⓗ≉ᚩ㔞㸦㡩ᚊⓗ≉ᚩ࠾ࡼࡧࢫ࣌ࢡࢺࣝ᝟ሗ㸧࠿ࡽᑟ. ኱つᶍࡢࢥ࣮ࣝࢭࣥࢱ࣮࡛ࡣࡍ࡭࡚ࡢ㘓㡢ࢹ࣮ࢱࢆ⫈. ฟࡋࡸࡍ࠸ࠕኌࡢࢸࣥࢩࣙࣥࠖ࡜࠸࠺ィ㔞್ࢆᐃ⩏ࡍࡿ㸬. ྲྀ࣭ࣔࢽࢱࣜࣥࢢࡍࡿࡇ࡜ࡣ㞴ࡋ࠸㸬⫈ྲྀࡸࣔࢽࢱࣜࣥࢢ. ࡇࡢ್ࡣ࠶ࡃࡲ࡛ኌࡢ≧ែࢆ⾲ࡍࡶࡢ࡛࠶ࡾࠊឤ᝟ࢆ┤᥋. ࡢᑐ㇟࡜࡞ࡿ㏻ヰࢆఱࡽ࠿ࡢ᮲௳࡛㑅ูࡋ࡞ࡅࢀࡤࠊ↓స. ᣦࡍࡶࡢ࡛ࡣ࡞࠸㸬. Ⅽ࡟ࢧࣥࣉࣜࣥࢢࡍࡿࡇ࡜࡟࡞ࡗ࡚ࡋࡲ࠺㸬ࡲࡓࠊ๓㏙ࡢ. ࢸࣥࢩࣙࣥࡢ್ࡣ㡢ኌ༊㛫༢఩࡛⟬ฟࡍࡿࡀࠊᑐヰ඲య. ㏻ࡾࠊ㏻ヰ࡟ᒓᛶࢆ௜୚ࡍࡿࡇ࡜࡛㡢ኌࢹ࣮ࢱࡢ஦ᚋά⏝. ࡟࠾ࡅࡿࢸࣥࢩࣙࣥࢆ᫬⣔ิ㸦௨ୗࠊࢸࣥࢩࣙࣥࣃࢱ࣮ࣥ㸧. ࡢྍ⬟ᛶࡀᗈࡀࡿ㸬. ࡜ࡋ࡚⾲ࡍ㸬ᑐヰࢆ≉ᚩ࡙ࡅࡿࡶࡢ࡜ࡋ࡚ࢸࣥࢩࣙࣥࣃࢱ. ㏻ヰࢹ࣮ࢱࡢᒓᛶ௜୚࣭ศ㢮ࡢࡦ࡜ࡘࡢ᪉ἲ࡜ࡋ࡚ࠊ㏻ヰ㡢ኌ࡟㡢ኌㄆ㆑ฎ⌮ࢆ᪋ࡋࠊࢱ࣮ࢤࢵࢺ࡜࡞ࡿ≉ᚩࢆ♧. ࣮ࣥࢆ═ࡵࡓ࡜ࡁࠊࢸ࢟ࢫࢺ࡟⾲ࢀ࡞࠸ᑐヰࡢືⓗ࡞≉ᚩ ࡀ᫂ࡽ࠿࡟࡞ࡿ㸬. ࡍෆᐜࡢ᭷↓࡛ᒓᛶࢆỴᐃࡍࡿࡇ࡜ࡀ⪃࠼ࡽࢀࡿ㸬ࡇࡢ࡜. ᅗ 3 ࡟ࡣ࣮࢜࣌ࣞࢱ࡜㢳ᐈࡢᑐヰࡢࢸࣥࢩࣙࣥࣃࢱ࣮ࣥ. ࡁࠊ኱ㄒᙡ㐃⥆㡢ኌㄆ㆑࡜㧗ᗘ࡞ព࿡⌮ゎࢆᐇ⌧ࡍࡿ⮬↛. ࡢ౛ࢆ♧ࡍ㸬ࡑࢀࡒࢀࡢⓎኌࢆ㡢ኌ༊㛫࡟ศ๭ࡋࡓᚋࠊ≀. ゝㄒฎ⌮࢚ࣥࢪࣥࡢ⤌ࡳྜࢃࡏࡀ㛗ᮇⓗ࡟᭱ࡶ」㞧࡞ࢱࢫ. ⌮ⓗ≉ᚩ㔞ࢆồࡵࡿ㸬ࡑࡢᚋࠊ⤫ィࣔࢹࣝࢆ⏝࠸࡚ࢸࣥࢩ. ࢡ࡟ᑐᛂ࡛ࡁࡿ࡜ᛮࢃࢀࡿ㸬ࡋ࠿ࡋ࡞ࡀࡽࠊ⌧ᅾࡢ࡜ࡇࢁࠊ. ࣙࣥࡢ್ࢆィ⟬ࡋࠊ᫬⣔ิ㸦ࢸࣥࢩࣙࣥࣃࢱ࣮ࣥ㸧ࢆồࡵ. ≉࡟ィ⟬ᶵ㈨※ࡀ᭷㝈ࡢሙྜࠊ≉ᐃ࣮࣮࢟࣡ࢻࡢࡳࢆ᳨ฟ. ࡿ㸦ᅗ୰ࠊ࣮࢜࣌ࣞࢱ࡜㢳ᐈࡢࣃࢱ࣮ࣥࢆ␗࡞ࡿⰍ࡛⾲♧. ࡍࡿ࣮࣡ࢻࢫ࣏ࢵࢸ࢕ࣥࢢᆺ㡢ኌㄆ㆑ࡶࡋࡃࡣ࣮࣡ࢻࢫ࣏. ࡋ࡚࠸ࡿ㸧㸬ࡇࡢࢸࣥࢩࣙࣥࣃࢱ࣮ࣥࡣ඾ᆺⓗ࡞ၥ㢟࿧㸦ࢡ. ࢵࢸ࢕ࣥࢢ࡜㐃⥆㡢ኌㄆ㆑ࡢ⤌ࡳྜࢃࡏࡀ⌧ᐇⓗ࡞ࢯࣜࣗ. ࣮࣒ࣞ㸧࡛࠶ࡿ㸬఍ヰࡢᚋ༙࠶ࡓࡾ࠿ࡽ㢳ᐈࡀᛣࡾࢆ♧ࡋࠊ. ࣮ࢩ࡛ࣙࣥ࠶ࡿ㸬. ࡑࢀ࡟ᑐࡋ࣮࢜࣌ࣞࢱࡣᑠࡉࡃ࠺࡞ࡎࡃࡀࠊࢸࣥࢩࣙࣥࣃ. ୍᪉ࠊ㡢ኌ࡟ࡣゝㄒ᝟ሗࡢ࡯࠿࡟ヰ⪅ᛶ࡞࡝㠀ゝㄒ᝟ሗ. ࢱ࣮ࣥ࡟ࡣࡑࡢࡼ࠺࡞ᵝᏊࡀ⾲ࢀࡿ㸬. ࡸࣃࣛゝㄒ᝟ሗࡶྵࡲࢀࡿࡀࠊࡑࢀ࡟ᇶ࡙ࡃᒓᛶ௜୚ࡸศ 㢮ࡶ᭷ຠ࡛࠶ࡿሙྜࡶ࠶ࡿ㸬௨ୗࠊࡑࡢࡼ࠺࡞ᢏ⾡࡟ࡘ࠸ ࡚ㄝ᫂ࡍࡿ㸬 (1) ヰ⪅ㄆ㆑࡟ࡼࡿ㏻ヰࢹ࣮ࢱࡢศ㢮ヰ⪅ᛶࡣ㏻ヰࢹ࣮ࢱࡢ㔜せ࡞ᒓᛶࡢࡦ࡜ࡘ࡛࠶ࡿ㸬ࢥ࣮ ࣝࢭࣥࢱ࣮࡟࠾࠸࡚ࡣࠊ≉ᐃヰ⪅ࡢ㏻ヰࢆ᳨⣴ࡍࡿ࡜࠸࠺ ฼⏝ࢩ࣮ࣥࡀ⪃࠼ࡽࢀࡿ㸬ࡑࡢ࡜ࡁࠊヰ⪅࡟㛵ࡍࡿ᝟ሗࡀ ᐜ᫆࡟ධᡭ࡛ࡁ࡞࠸ሙྜࠊヰ⪅ㄆ㆑ᢏ⾡ࡀ᭷ຠ࡛࠶ࡿ㸬ࡇ ࡢࢽ࣮ࢬࢆ‶ࡓࡍࡢࡣⓎኌෆᐜ࡟౫Ꮡࡋ࡞࠸ࢸ࢟ࢫࢺ㠀౫ Ꮡᆺࡢヰ⪅ㄆ㆑ᢏ⾡࡛࠶ࡿ[2]㸬ᙜ♫ࡣ⊂⮬ࡢヰ⪅ㄆ㆑ࢯࣇ ࢺ࢙࢘࢔[e]ࢆ࣑ࢻ࢙ࣝ࢘࢔࡜ࡋ࡚㈍኎ࡋ࡚࠸ࡿࡀࠊ௒ᚋࡣ ࢡ࣮࣐࣮ࣞࡢ≉ᐃࡸ࠸ࡓࡎࡽ㟁ヰࡢ㜵Ṇࡢ࡯࠿ࠊ࣐࣮ࢣࢸ ࢕ࣥࢢࢶ࣮ࣝ࡜ࡋ࡚ࡢ฼⏝ࡀᮇᚅࡉࢀࡿ㸬 (2) ኌࡢࢸࣥࢩࣙࣥࡢศᯒ ㏻ヰࢹ࣮ࢱࡢᒓᛶ௜୚࣭ศ㢮ࡀᚲせ࡜࡞ࡿࡶ࠺ࡦ࡜ࡘࡢ ලయ౛ࡣࠊࢺࣛࣈࣝ㏻ヰ㸦࠸ࢃࡺࡿࠕၥ㢟࿧ࠖ㸧ࡢ᳨ฟ࡜࠸ ࠺ࣔࢽࢱࣜࣥࢢᴗົ࡛࠶ࡿ㸬୙‶ࢆᣢࡘ㢳ᐈࡀࢥ࣮ࣝࢭࣥ ࢱ࣮ࡢ࣮࢜࣌ࣞࢱ࡟ᑐࡋ࡚ኌࢆⲨࡆࡿ࡜࠸ࡗࡓሙ㠃ࡀ඾ᆺ ⓗ࡞౛࡛࠶ࡿ㸬஦๓࡟ࡍ࡭࡚ࡢ㏻ヰࢆ⮬ື࡛ศᯒࡋࠊ᳨ฟ. ᅗ 3 ኌࡢࢸࣥࢩࣙࣥࡢศᯒ. ࡢ☜⋡ࡢ㧗࠸㡰࡟ࢹ࣮ࢱࢆ㓄⨨࡛ࡁࢀࡤᡭືࣔࢽࢱࣜࣥࢢ. Figure 3 Voice tension analysis. e) ヰ⪅ㄆ㆑ࢩࢫࢸ࣒ VoiceSyncII. ‫ף‬2013 Information Processing Society of Japan. 4.

(5) Vol.2013-SLP-98 No.2 2013/10/25. ᝟ሗฎ⌮Ꮫ఍◊✲ሗ࿌ IPSJ SIG Technical Report ࡇࡢࢱ࢖ࣉࡢ㏻ヰࡢ≉ᚩࢆᤊ࠼ࠊၥ㢟࿧ࡢ᳨ฟࢆྍ⬟࡟. ࡃࡢၥ㢟࿧ࡢ᳨ฟࡀྍ⬟࡛࠶ࡗࡓࡇ࡜࡟࡞ࡿ㸬ࡑࡢࡓࡵࠊ. ࡍࡿࡓࡵࠊ ࠕၥ㢟࿧ࡽࡋࡉࠖࢆ⾲ࡍࠕࢸࣥࢩࣙࣥࢫࢥ࢔ࠖ࡜. ணࡵࢸࣥࢩࣙࣥࣃࢱ࣮ࣥ࡟ᇶ࡙ࡃศᯒࢆ⾜࠼ࡤ᳨ฟసᴗࡢ. ࠸࠺⤫ィ㔞㸦ࡶࡋࡃࡣ ad-hoc ್㸧ࢆᐃ⩏ࡍࡿ㸬ࡇࢀ࡟ᇶ࡙. ຠ⋡໬࡟ࡘ࡞ࡀࡿࡇ࡜ࡀศ࠿ࡗࡓ㸬. ࠸࡚㏻ヰࢹ࣮ࢱࢆศ㢮ࡋࠊၥ㢟࿧ࡢ᳨ฟసᴗࡢຠ⋡໬ࢆࡣ 3.5 ㏻ヰෆᐜࡢศᯒ࣭ά⏝. ࠿ࡿ㸬 ௨ୗࠊࢸࣥࢩࣙࣥศᯒ࡟ᇶ࡙ࡃၥ㢟࿧᳨ฟࡢ஦౛ࢆ♧ࡍ㸬. ๓㏙ࡢ㏻ࡾࠊ㏻ヰ㡢ኌ࡟ࡣ㔜せ࡞᝟ሗࡀྵࡲࢀ࡚࠸ࡿࡓ. ᮏ஦౛ࡣ㔠⼥ᶵ㛵ࡢࢥ࣮ࣝࢭࣥࢱ࣮࡛⾜ࡗࡓᐇドᐇ㦂࡛࠶. ࡵࠊ࣐࣮ࢣࢸ࢕ࣥࢢࡸ௻ᴗࢥࣥࣉࣛ࢖࢔ࣥࢫ⟶⌮࡞࡝ᵝࠎ. ࡿ㸬஦ᴗࡢ୍⎔࡜ࡋ࡚ࠊമົ⪅࡬㟁ヰࢆ࠿ࡅࠊ≧ἣࢆ☜ㄆ. ࡞ศ㔝࡛฼⏝࡛ࡁࡿ㸬ࡓࡔࡋࠊ㏻ヰෆᐜࢆࢸ࢟ࢫࢺ࡟ኚ᥮. ࡍࡿ࡜࠸࠺ᴗົࡀ࠶ࡿࡀࠊ఍ヰࡀࢫ࣒࣮ࢬ࡟⾜ࢃࢀࡓ࠿ࡀ. ࡋࠊࡉࡽ࡟ព࿡ࢆ⌮ゎࡋ࡚せ⣙ࡍࡿ࡜࠸࠺୍㐃ࡢࣉࣟࢭࢫ. 㔜せ࡞࣏࢖ࣥࢺ࡛࠶ࡿ㸬ࡑࡢࡓࡵࠊࡍ࡭࡚ࡢ㏻ヰࢆ㘓㡢ࡋࠊ. ࢆࡍ࡭࡚⮬ື࡛⾜ࢃ࡞࠸㝈ࡾࠊ㡢ኌㄆ㆑ࡸ⮬↛ゝㄒฎ⌮ࡢ. ᡭື࡛ࣔࢽࢱࣜࣥࢢసᴗࢆ⾜ࡗ࡚࠸ࡿ㸬ࡋ࠿ࡋ࡞ࡀࡽࠊࣔ. ࣏ࢸࣥࢩࣕࣝࢆࣇࣝ࡟ά⏝࡛ࡁࡓ࡜ࡣ࠸࠼࡞࠸㸬ᐇ㝿ࠊୖ. ࢽࢱࣜࣥࢢసᴗࡣᩘ༑ྡࡢయไ࡛⾜ࡗ࡚࠸ࡿࡓࡵࠊẖ᪥Ⓨ. グࡢᢏ⾡ࡣⓎᒎ㏵ୖ࡟࠶ࡿࡓࡵࠊ᭷ຠ࡞ࢯ࣮ࣜࣗࢩࣙࣥ࡟. ⏕ࡍࡿᩘ༓༢఩ࡢ㏻ヰࢆࡍ࡭࡚ࢳ࢙ࢵࢡࡍࡿࡇ࡜ࡣ࡛ࡁ࡞. ࡍࡿࡓࡵ࡟ࡣே㛫ࡢ௓ᅾࡀᚲせ࡟࡞ࡾࠊసᴗࡢຠ⋡໬ࢆࡣ. ࠸㸬ࡑࡇ࡛ࠊࣔࢽࢱࣜࣥࢢసᴗࡢຠ⋡໬ࡀ᭱⤊ⓗ࡞┠ⓗ࡛. ࠿ࡿࡇ࡜ࡀ౯್ࢆ⏕ࡳฟࡍ࡜࠸࠺ほⅬࡀ᭱ࡶ⌧ᐇⓗ࡛࠶ࡿ㸬. ࠶ࡿ㸬. ࡑࡇ࡛ࠊ㈝⏝ᑐຠᯝࡢഃ㠃ࡶྵࡵ⌧≧ࡢᢏ⾡࡛࡝ࡢࡼ࠺࡞. ᐇ㦂࡛ࡣࢥ࣮ࣝࢭࣥࢱ࣮ࡢ⣙ 8000 ಶࡢ㏻ヰࢹ࣮ࢱࡢ୰ࠊ ఱ௳ࡢࠕၥ㢟࿧ࠖࢆ᳨ฟ࡛ࡁࡿ࠿᳨ドࡋࡓ㸬ၥ㢟࿧ࡢᐇ㝿. ࢯ࣮ࣜࣗࢩࣙࣥࡀᐇ⌧ྍ⬟࡛࠶ࡿ࠿᥈ࡿ㸬 (1) ⡆᫆ⓗ࡞ࢺࣞࣥࢻศᯒ. ࡢ௳ᩘࡣᮍ▱࡛࠶ࡿࡀࠊ㏻౛ࡢẚ⋡࡛ࡣ㸯㸪㸰௳࡯࡝Ꮡᅾ. ㏻ヰෆᐜࡢ⡆༢࡞ά⏝ἲ࡜ࡋ࡚⪃࠼ࡽࢀࡿࡢࡣࠊୡ㛫ࡢ. ࡍࡿ࡜࠸࠺๓ᥦ࡛࠶ࡗࡓ㸬㡢ኌࢹ࣮ࢱࡢ㘓㡢ࡣ㢳ᐈ࡜࢜࣌. ᪂ࡋ࠸࣮࣮࢟࣡ࢻࡸ⮬♫࡛㔜せどࡉࢀࡿ⏝ㄒ㸦♫ྡࠊ〇ရ. ࣮ࣞࢱࡢኌࢆศ㞳ࡋ࡚㘓㡢࡛ࡁࡿᙜ♫ࡢ㏻ヰ㘓㡢ࢩࢫࢸ࣒. ྡ࡞࡝㸧ࡀ㏻ヰ୰࡟࡝ࡢ㢖ᗘ࡛ฟ⌧ࡍࡿ࠿ࢆศᯒࡍࡿࡇ࡜. VoiceTracking ࡛⾜ࡗࡓ㸬. ࡛࠶ࡿ㸬㢖ᗘࡢ᫬⣔ิࡢ⤫ィศᯒࢆ⾜࠺ࡇ࡜࡟ࡼࡗ࡚ࠊ᭷. ࢹ࣮ࢱࢭࢵࢺࡢࡍ࡭࡚ࡢ㏻ヰࢹ࣮ࢱࡢࢸࣥࢩࣙࣥࣃࢱ. ຠ࡞㛵ಀᛶࢆᑟࡃࡇ࡜ࡀྍ⬟࡜࡞ࡿ㸬. ࣮ࣥࢆồࡵࠊࢸࣥࢩࣙࣥࡢᖹᆒ್ཬࡧᐇ㦂ⓗ࡟ㄪᩚࡋࡓࡋ. ౛࠼ࡤࠊ≉ᐃ࣮࣮࢟࣡ࢻ࡟࠾ࡅࡿ᭙᪥ูࠊ᫬㛫ᖏูࡢฟ. ࡁ࠸್ࢆ㉸࠼ࡿᅇᩘ࡟ᇶ࡙ࡃᩘᘧࢆࠕࢸࣥࢩࣙࣥࢫࢥ࢔ࠖ. ⌧㢖ᗘࢆㄪ࡭ࡓ࠸࡜ࡍࡿ㸦ᅗ 5㸧㸬ࡇࡢࡼ࠺࡞ศᯒࡣ▐᫬࡟. ࡜ࡋࡓ㸬ࡇࡢࢫࢥ࢔ࢆ⏝࠸࡚ࠊᅗ 4 ࡢࡼ࠺࡟㏻ヰࢹ࣮ࢱࢆ. ⤖ᯝฟຊࡀồࡵࡽࢀࡿࡓࡵࠊ㡢ኌ࠿ࡽணࡵ᝟ሗࢆᢳฟࡋ࡞. ࢫࢥ࢔㡰࡟୪࡭ࡓ㸬ࢫࢥ࢔ࡣ᭱㧗್࡛ṇつ໬ࡋࡓ㸬ࡑࡢᚋࠊ. ࡅࢀࡤ࡞ࡽ࡞࠸㸬ᙜ♫ࡣࠊࡇࡢࡼ࠺࡞⏝㏵࡟ྥࡅࠊ᪂ㄒ࡞ ࡝࡟ࡶᑐᛂ࡛ࡁࡿ㧗㏿ࡢ᳨⣴࢚ࣥࢪࣥࢆᥦ౪ࡋ࡚࠸ࡿ[f]㸬 800. 160. 700. 140. 600. 120. 500. 100. 400. 80. 300. 60. 200. 40. 100. 20. 0. ᪥. ᭶. ⅆ. Ỉ. ᮌ. 㔠. ᅵ. 0 䠍䠌᫬䡚䠍䠍᫬䡚䠍䠎᫬䡚䠍䠏᫬䡚䠍䠐᫬䡚䠍䠑᫬䡚. ᅗ 5 ᭙᪥ู࣭᫬㛫ᖏูࡢ࣮࣮࢟࣡ࢻࡢฟ⌧ᩘ Figure 5 Number of keyword utterance by day of the week and by time frame. (2) 㡢ኌࢸ࢟ࢫࢺኚ᥮ ㏻ヰࢹ࣮ࢱࡢⓎኌෆᐜࢆࢸ࢟ࢫࢺ࡟ኚ᥮࡛ࡁࢀࡤ࣐࣮ࢣ ࢸ࢕ࣥࢢࡸࢥࣥࣉࣛ࢖࢔ࣥࢫ⟶⌮࡞࡝ᵝࠎ࡞ศ㔝࡛㛗ᮇⓗ ᅗ 4 ṇつ໬ࢸࣥࢩࣙࣥࢫࢥ࢔ࡢ㡰࡟㏻ヰࢆ୪࡭ࡓᵝᏊ㸬. ࡞ᜠᜨࡀࡶࡓࡽࡉࢀࡿࡇ࡜ࡣඹ㏻ࡢㄆ㆑࡛࠶ࡿ㸬ࡓࡔࡋࠊ. ࢫࢥ࢔ࡢ㧗࠸㏻ヰ࠿ࡽ⫈ྲྀ☜ㄆࢆࡍࡿ㸬. ௒ࡢ࡜ࡇࢁࡍ࡭࡚ࡢ㏻ヰෆᐜࢆࢸ࢟ࢫࢺ࡟᭩ࡁ㉳ࡇࡍࡢ࡟. Figure 4 Calls arranged in descending order of tension scores.. ኱ࡁ࡞ࢥࢫࢺࡀ࠿࠿ࡿࡓࡵࠊᡞᝨ࠺࣮ࣘࢨ࣮ࡀከ࠸㸬. Monitoring is done starting from the calls that yielded the. ⌧ᅾࠊᑟධࢥࢫࢺࢆᢚ࠼࡞ࡀࡽ㡢ኌࢸ࢟ࢫࢺኚ᥮ᢏ⾡ࢆ. highest scores.. ᑟධࡋࠊ᪂ࡋ࠸ά⏝᪉ἲࢆ㛤Ⓨࡋࡓ࠸࣮ࣘࢨ࣮࡟ࡣࠊࣇࣟ ࣥࢺ࢚ࣥࢻࣉࣟࢭࢵࢧ࡜ࡋ࡚࣮࣡ࢻࢫ࣏ࢵࢸ࢕ࣥࢢᆺ㡢ኌ. ࢫࢥ࢔ࡢ㧗࠸㏻ヰ࠿ࡽᐇ㝿࡟⫈ྲྀ☜ㄆࢆ⾜ࡗࡓ࡜ࡇࢁࠊࢫ. ㄆ㆑ࢆ฼⏝ࡍࡿࡇ࡜࡛ᑐ㇟࡜ࡍࡿࢱ࣮ࢤࢵࢺࢆ⤠ࡾࠊ㔜せ. ࢥ࢔ࡢࢺࢵࣉ 50 ㏻ヰࡢ࠺ࡕࠊ㸲௳ࡢၥ㢟࿧ࡀ☜ㄆࡉࢀࡓ㸬 ඲㏻ヰࢹ࣮ࢱࡢ 0.7%࡟┦ᙜࡍࡿ㏻ヰᩘࢆㄪ࡭ࡿࡔࡅ࡛ከ. ‫ף‬2013 Information Processing Society of Japan. f) 㧗㏿㡢⣲ᆺ᳨⣴࣭ㄪᰝࢩࢫࢸ࣒ VoiceTracking VOC Analyzer. 5.

(6) ᝟ሗฎ⌮Ꮫ఍◊✲ሗ࿌ IPSJ SIG Technical Report. Vol.2013-SLP-98 No.2 2013/10/25. ࡜ࡉࢀࡿ㏻ヰࡢࡳ࡟ᑐࡋ࡚኱ㄒᙡ㐃⥆㡢ኌㄆ㆑ࢆ⾜࠺࡜࠸ ࠺㑅ᢥ⫥ࡀ࠶ࡿ㸦ࣁ࢖ࣈࣜࢵࢻࢱ࢖ࣉࡢ㡢ኌㄆ㆑[3]㸧㸬 ْৢਵਫ਼ดનੳ ઺એॖও‫ش‬४ٓ. ᅗ 6 ࣮࣡ࢻࢫ࣏ࢵࢸ࢕ࣥࢢ[g]࡜኱ㄒᙡ㐃⥆㡢ኌㄆ㆑ࢩ ࢫࢸ࣒[h]࡛ᵓᡂࡉࢀࡿࠕࣁ࢖ࣈࣜࢵࢻᆺ㡢ኌㄆ㆑ࢩࢫࢸ࣒ࠖ Figure 6 “Hybrid” speech recognition system formed by a word-spotting front end and a large-vocabulary continuous speech recognition system.. 4. ࠾ࢃࡾ࡟ ࢥ࣮ࣝࢭࣥࢱ࣮࡟࠾ࡅࡿ㏻ヰ㘓㡢ࡢᾐ㏱ࡀ㐍ࡴ୰ࠊ㡢ኌ ᢏ⾡ࡢά⏝ࡢሙࡣࡲࡍࡲࡍᗈࡀࡾࡘࡘ࠶ࡿ㸬ࡲࡓࠊᚑ᮶⪃ ࠼ࡽࢀ࡚ࡁࡓ㡢ኌㄆ㆑࡜㡢ኌྜᡂࡢ࡯࠿ࠊࣃࣛゝㄒ࣭㠀ゝㄒ᝟ሗࡢά⏝ࡶྍ⬟࡛࠶ࡿࡇ࡜ࡀࢃ࠿ࡗࡓ㸬௒ᚋࡣࡼࡾᖜ ᗈ࠸⏝㏵࡛ࡢ㡢ኌᢏ⾡ࡢά⏝ࢆࡵࡊࡍ࡜࡜ࡶ࡟ࠊ㡢ኌᢏ⾡ ࡀࢥ࣮ࣝࢭࣥࢱ࣮ࡢ㹇㹒ࢩࢫࢸ࣒ࡢḞ࠿ࡏ࡞࠸㒊ศ࡟࡞ࡿ ࡇ࡜ࢆ㢪ࡗ࡚࠸ࡿ㸬. ཧ⪃ᩥ⊩ 1) http://ja.wikipedia.org/wiki/㔜せᴗ⦼ホ౯ᣦᶆ 2) ㉺௰Ꮥᩥ,⠛⏣ᾈ୍, “ヰ⪅ㄆ㆑ࡢᅜ㝿ືྥ,” ᪥ᮏ㡢㡪Ꮫ఍ㄅ 69 ᕳ 7 ྕ, 2013. 3) ͆〇ရࢺࣞࣥࢻ㸭㡢ኌㄆ㆑㸸ᮇᚅ್࡜ά⏝ࣞ࣋ࣝ࡟ࢠࣕࢵࣉ,͇ Computer Telephony, 㸷᭶ྕ, 2013.. g) ࣮࣮࢟࣡ࢻ᳨⣴ࢩࢫࢸ࣒ VoiceTracking KeywordFinder h) 㧗⢭ᗘ㡢ኌㄆ㆑㸦ASR㸧ࢯ࣮ࣜࣗࢩࣙࣥ SpeechRec (NTT-IT ♫). ‫ף‬2013 Information Processing Society of Japan. 6.

(7)