咽喉音を利用した会話・摂食行動の認識

全文

(1)「マルチメディア，分散，協調とモバイル (DICOMO2017)シンポジウム」平成29年6月. '(X"?J3YG#LO ďȑǱħ ɿȑɣ Ġğî ƶƊɨ ƟǃĈɍ ĞɘÔÝ țƂɨÝ 9H[ȫ"/ɸ8D/4%E&EñƓȊʄȮǕƓȊ2ƨɟGCŔ+ʅƁǗǢ.4ɩɰtLT/ê ítLT3 2 ayh3ɰĄBȫʄšɸȒÌ3ƍ¼HȒʅśƉƢ4 GMM H¢*-ȫšɸ2ɟɎ"D ȬǳȒÌ3ȶÂHn vÓ.Ȓʄ%3ȶÂǶƅH LSTM 2A*-Ƿà"D/.Ŋž3ȫšɸȒÌHķ Dʅđǀā2-ȶÂǶƅHlz ~\cKbT\2þ,ŵɞɃ.ħƮÎ"DŭƢ/ƛ8-ȒÌ3ƍ¼Ǯ ĪHŦì.D/HǘȮ 'ʅ. Conversational and Eating Behavior Recognition by Leveraging Throat Sound JUMPEI ANDO†1 TAKATO SAITO†2 SATOSHI KAWASAKI†2 MASAJI KATAGIRI†2 DAIZO IKEDA†2 HIROSHI MINENO†3, †4 MASAFUMI NISHIMURA†3. 4ɛſ¥Č3ĎȂÜȊ1gz }hbe T.. 1. ȿɾʀǚƾđ/1CʄɾʀȄ3ĺɁ3ƹłʄB2 4%3ēů@ǁƤǀā;.Hä>'ɟGCHĦĩƖǘ2 ƿȡ"D/.ʄɾʀȄ3®ĭǨƿ/ȸY m\3ËƼ ÎHđƾ. '/ȜžDʅ%3.?Ƹ2ɾʀȄ. 3ɸ8D/4ñ3ȊÉ2ʄȫ"/4ȮǕȊ. Cʄn vÓ3ć²ǎ1ȬǳȶÂǶƅHǷà. ʄŊž. 3ǭĪ2Ɉĳ.ȒÌʂ.4ȫ@ɸ103ȒÌʃ H¼É.DÜȊĽDʅƁǠ.4ʄđǀā.Ùɩ 'd _Hǂ-ɸ×6ȫȒÌ3ƍ¼HȒ/.Ľ ȊȪ¦. '3.ÿå"Dʅ. É2ɟGDɗȜ1ȒÌȜǲ.Dʅň4E;.ȷɰ Ą103ȥɚ2ǂDɩɰtLT2ÊʄhbTjfû 3êítLT(÷ 1)BķBEDʇayh3ɰɱľÿ Hǂ-ɸʄȫȒƳ2ɥ"DȬǳ1ȒÌ3ȶÂ2ɟ "DǗǢHȒ*-'[1-3]ʅ.4ǌȫ@ñ/* 'ȒÌ2ɟɎ"DɰÏɞHƍ¼Ķ2ȒÌ3ȶÂHȒ1*'ʄđɦ3űĥǁƤ2DćƐ.șɪ1ƹơ. 4ʄ3A1ɰÏɞ3ƍ¼ȋĈąõɫ.D/ Ȱɵ*'ʅ ƁǗǢ.4ʄÈ»ƿ/. -3ɰÏɞ3ƍ¼4ȒG#ʄn. ÷ 1 ƁǗǢ.ǂ'tLT3ƏŇ/%3ȗǓ£. vÓ.Ȓ*'ȶÂǶƅ3ŵǯ¿d _HĶǎ2Ƿ à"D/.ʄȒÌHƍ¼"DŭƢHƍȤ"Dʅ3A 1Jp a3Āàʄn vÓ3ȶÂǶƅHŵɞɃ. 2. VR@A. .ħƮÎ"D»ƿȎǎ2ȒGE-Dʅ%3'>2ʄ. űĥǁƤȒÌHȮȶ"DǗǢ4ĩȒGE-DʅE. Ɛ1lz ~\cKbT ǂŀEDʄćƐ. ;.3´ȒǗǢ4ʄ Ē¸ǀā2ü>ɇ>BE'^Y2. 1ƹơHĿĐ. ACÃǂȄ3ȒÌHȮȶ"DŭƢ[5-7]ʄ ÃǂȄ3Ɂ. - Hź34ĔŲ.41ʅŭʄ. ň4ʇayh3ľÿ¡ǂ"D/.ȒÌȶÂȊÉ. 2ȗǓ. ɾ;D/HǘȮ. 2 ǟɷ2ĈÂ.DʅÈȄ 4Ʒë2ü>ɇ=^Yɾ. -Cʄ3A1ć²ǎľÿ?ȃŅ. ' Hǂŀ"D34B2õɫ.*'ʅEB3. '^YHǂ-ȒÌHȮȶ"DŭƢ[8-15]/ʄ. X\e.D2ʄ>^YHȦȀ. 'ǣɞ.. ȒÌ. șɪ1ƹơ2Ė»"D'>ʄƁǗǢ.4 RNN 3ǟ.. HȮȶ"D/.1ʅ%3'>ʄɛŵɞʄēĬ¸2. D LSTM(Long Short-Term Memory)[4]HÃǂ"DʅLSTM. DHɢ-ÈȄ3ŭƢHɓǂ"D/4ɫ. 1 ɮĝĈĎĈĎɡǽàǜĎōȓǗǢǜ Graduate School of Integrated Science and Technology, Shizuoka University 2 NTT fXx´ɏōȓǗǢŊ ǚ^ZUǗǢU p Urban Sensing Research Group, Research Laboratories, NTT DOCOMO, INC.. 3 ɮĝĈĎĎȓɡľÿĎɳý College of Informatics, Shizuoka University 4 JST / JST, PRESTO. ʅĶȄ. Copyright (c) 2017 by the Information Processing Society of Japan. ― 116 ―.

(2) 4ÃǂȄ3Ɂ2ȗǓ. '^YHǂ-D3.ʄµ. ƦŵɞǧHɢ-9:ű3ȒÌHɎǺ. -^Z. ;'ʄūƻ[13]3A2ȒÌȶÂ2-gz }h be T3Ƥǂ"DǗǢ4ćʄ%3ć4ĎȂd . U.Dʅ;'ʄɛŵɞ^YHɁ2ȗǓ"D/HĿ. _ƏǪHȰɵ/. Đ. ǔǒ10/ƛ8-eLŗɢ10Ùɩɫ. -CʄĘû.ȗǓȻŒ3ę1ÊɌĪ^Y@ɩɰ. -Dʅūƻ[13].4ʄÙɩĔŲ1. tLTAƤǂEDʅ%3'>ʄÈȄ /ƛ8-ď¦. ĖȺȒÌ3ȶÂǮĪ1*-Dʅ2?ʄūƻ. 2ɐǂ.D'>2ǁƤȂń@®ĭǨƿ3A1}Ln. [8].4\t enOBķBEDʈɃÊɌĪľÿHþ. UHǏǎ/. 2ÃǂȄǥ*-Dī*-DƗ-D. 'ȒÌȮȶ.4ć¢GE-Dʅ. }LnU3A1ɛŵɞ+ćƐ1ǀā.Ùɩ _HĖȺ/. 'd. 'd _ƛ8-ĥ2Ɛ1ǟɷ3ɪɰɗǈ. Ƚ*-DɤƙHǍ*-D/*-DÌ 2 ɟ. 'ȒÌȶÂ3ĀàʄɠBE'ǣɞ.Ùɩ -Dʅ/. -ʈěk ^peHǂ-Ì ȶÂHȒ*-. Dʅ. . 1Bʄ)B?ĎȂd _HĈə2Ùɩ.. CGʄɰd _3Āà4%3°ãıʅ%3'>ʄ. 1*'Ì 2ɟ. >ŵɞɃ.ĖȺȒÌ3ÏɞHƍ¼. ʄ%3ÏɞH¾C¼. 1*-Dʅ%3ŭʄĎȂd _YL]3ȟƲBgz. ʅ/CGʄĖȺȒÌ3ǹǺŵ. }hbe TH¢G1ȒÌȶÂ?ȒGE-Dʅ. -ȶÂ"D/4ɫ. -ƠĐƀÁÂ/ƛ8-ȶÂǮĪ. ɞǇ1DĀà@đǀād _2ȝBEDșɪ2ȒÌ¾. ūƻ[14].4 SVM H¢*-ȶÂ. CŻGDĀà4%3āǆŵÅHȋÌǎ2ÁŬ"D/ɫ. žBn vĦH 10 ǝɞ/. '>Ɩǘ2Ïɞƍ¼HȒ34õɫHƎ>Dʅň ɑÖđŮ. -Dʅđɐǂŵ3Ŋ -Dʅ%3'>ʄęə. 3ĎȂd _.ȶÂH"DĻȜ*'3.½ɷƓ2ƞÎ ĽȊ3ɾ SVM Hǂ-Dʅ. 'ǗǢ[2]2-?êítLT2µDɪɰ@. ĖȺȒÌ3ǹǺŵɞġ3ĴɱHÛ-Ïɞƍ¼3ǮĪ . -Dʅ;'ʄǮĪãHǏǎ/. Ïɞƍ¼. - DNN Hǂ-. 'Āà2+-?ɪɰ3ĴɱHƪ.'. ŭʄǹǺŵɞ3ġ2Ⱦô"Dƍ¼ȯC3Ŧì4ǘȮ.1 *'[3]ʅ -ʄÈ»ƿ/. ȥ3ŋƢ2GDŭƢ/. -3Ïɞ3. ƍ¼4ȒG#ʄn vÓ.Ȓ*'ȶÂǶƅ3ŵǯ¿d _HĶǎ2Ƿà"D/.ʄȒÌHƍ¼"DŭƢČ ù"D[8-15].3ŭƢ3ÃƲ/. -ʄɪɰǧHȯƍ¼. -. ?Ķǎ2Ƿà"DɑǞ.ȯƍ¼HŘɢ.Dʅ;'ʄǹ Ǻŵɞġ@șɪ2ȒÌ¾CąGDĀà2+-?ʄǖſ ɞ3n vÓ.ȶÂ ĖļÜȊ.Dʅ D/4ɫ. . 'ǶƅHɓ¾2Ƿà"D/. 1Bʄɓ¾1Ƿà HƠ>. ʅ/CGʄđǀād _2ȝBEDćƐ. 1ƹơHĿĐ. - Hź34ĔŲ.41ʅūƻ. [13].4ʄ]r3sVbe2µE'\t enO/ȇ 2ȦȀ. 'ɩɰtLTH¢*- 48 ŵɞɎǺ. /æö3ǀāɰHÙɩ. -ÊɌĪ. ʄLSTM Hǂ-ɸ@ǔǒ10. 9 ǟɷ3ȶÂHȒ*-Dʅ.4ʆ½ɞ3ɛn v.ȶÂHȒ'>2ʄn v¸.żɴ/1*'ȶÂǶ ƅHżǴǶƅ/"D Hǂ-Dʅ %3ŭʄƁǗǢ.4ɸ3ȫ10ʄȫšɸ3 Ȭǳ1ȒÌ½ƄHȒ/HǏǎ/. -CʄĀà2A*. -B2ǖÓ.3ÁÂĻȜ21Dʅ. . 1Bʄ. ȒÌ½Ƅ2-Ŋž3ǶƅķBED/ſĵ.Dn vɛHĐ>D34ɫ. ʅ;'ʄūƻ[13]3A1øĐ. 3n vɛ.Ƿà"DĀàʄȒÌ3ǹǺŵɞġ@șɪ2 ȒÌ¾CGDĀà2ɟ. -Ɔɂ2Ėļ.1ʅ%3. '>ʄĺɁƹơHŎŝ"D/Ǐǎ2àȌ"DA1Ȭ ǳ1ȒÌľÿHķD/4Ǝ>-ɫ. /ȃBEDʅ. 3. Y JG#O! 3.1 O!.PG# ƁǗǢ.4ƐDűĥǁƤ3ȒÌ3Bɸ8D /ȫ"/2ɟɎ"DȒÌ/. -ɸ×6ȫȒ. ÌHĖȺ/"Dʅɸ×6ȫ4%E&Eñ3ȊÉʄȮ ǕȊÉ/ƨɟ¨D/ǕBE-Dʅ ñ3ȊÉ 4ȯñ2+1CȈƱ3Õô/1D/ǕBE- DʅÊ-ʄ®ĭǎ1ɸǁƤ4Ɂ3®ĭH©+32Ɣ $1ʅɸə3ƪę2ADɾʀȄ3ƈɺ4ɗǫ1ǉǊ H?'B"ʅáƐ2ȮǕƓȊ3³¬/. -ȫ3ƪę. @ƺȫ3ĂÊDʅŹ24XuzgV Z|3ƪę2 ADɾʀȄ3+Ǌ2+1Dʅ%3'>ʄɸ×6ȫ 3ȒÌHƿȡ"D/.ɾʀȄ3®ĭǻŔǨƿ3ɾĪÎ ſĵ.D'>ĖȺ/. 'ʅ. 3.2 9H ´ȒǗǢ3Ǖȝ/ňɑÖȒ*-'êítLT/ɩ ɰtLTH¡ǂ"D/3±ĽHɀȚ. ʄʇ+3tLT. 3ªÞBķBE'ʇ+3ȶÂǶƅ3ŵǯ¿HĶǎ2 LSTM .Ƿà"D/.Ŋž3ȒÌHƍ¼"DŭƢHƍȤ 'ʅ´ȒǗǢ2DA1n vÓ3ȶÂǶƅHŵ ɞɃ.Ƿà"DĀàʄn vÓ3ǶƅHŵɞɃ. ħƮÎ"D»ƿȎǎ2ȒGE-Dʅ%3'>2ʄ Ɛ1lz ~\cKbT ǂŀE-Dʄ d _@ƍ¼3ǏǎąŹ2ļ!- HƠ>ǐ"ĻȜ ǁ!ʄǮĪ3ɾ ǌȝ24Ĉą1ŋɞDʅ ;'ʄ}LnU3A1ɛŵɞ+ćƐ1ƹơ.Ùɩ 'd _2Ė. -4n vÓ3ȶÂǮĪ35B+. Ĉ1D°ãDʅ%3'>ʄ"8-3ƹơ2Ė. Copyright (c) 2017 by the Information Processing Society of Japan. ― 117 ―. -.

(3) Ėļ.D Hǌȝ"D/4õɫ.Dʅŭʄ. .Dʅ%3Ƕƅ/. ň4ʇayh3ľÿ¡ǂ"D/.ȒÌȶÂȊÉ. ;'ʄêítLT4šɸȒÌ3ȶÂ2ŽË.D[16]ʅ%. ɾ;D/HǘȮ. -Cʄ3A1ć²ǎľÿ?ȃŅ. 3'>ʄɸ3ȶÂ?ſĵ.Dʅ1ʄķBEDǶƅ4. ' Hǂŀ"D34B2õɫ.*'ʅEBș. n vÓ.CʄɎǺĽ4©ɧE1ʄ3Ƕƅ HħƮÎ»ƿ"D34ĔŲ.D/ȃ-Dʅ. ɪ1ƹơ2Ė»"D'>śƉƢ.4 LSTM Hǂ-ʇ+ 3ȶÂǶƅ3ŵǯ¿ľÿʂêítLT3ȶÂǶƅ3ŵǯ¿. . /ɩɰtLT3ȶÂǶƅ3ŵǯ¿ʃHǷà"DŋƢHƍȤ. . 'ʅLSTM 4 RNN 3ǟ.Cʄɛŵɞ3ǑɟHȃŅ. . ƬBķBE'ñʄèòʄǌȫ103ćƐ.șɪ1ŵǯ. A. ¿ǁȾk_ BʄĻȜ/"DǭĪ.ɸ@ȫHȶÂ. Z. .Dʅ. . .

(4). ÜȊ1gz }hbe T.Dʅ%3'>ʄßľÿ. GMM. ;'ʄgz }hbe TH¢*'ȶÂHȒĀ àʄĎȂd _3ÙɩȰɵ/1Dʅ/CGʄűĥǁƤ ǀā.Ùɩ. -ȫ/ƺȫ3ȶÂ2?ǿDʅ. A. .

(5) . . …. Z. GMM LSTM.

(6) . . ÷ 2 śƉŋƢ. 'ćƐ1ĎȂd _2n vÓ3Ȭǳ. 1ȒÌ}qH"D34Ĉąõɫ.DʅśƉƢ. 4ʄGMM .B!>ȶÂ /.ɸ/ȫHȶÂ. 'ǶƅH LSTM .Ƿà"D. -DʅǶƅ/. 3.4 GMM

(7) KBG#O! GMM Hǂ'ȬǳȒÌ3n vÓȶÂ.4ʄµÉ. -ʄĪ.ɰ. ɱǎ1ƸĹəBɸ/ȫHĎȂ"D/ƛ8-ʄgz . HêítLT ɩɰtLT%E&E3ȟƫªÞ/. }hbe T3ĎȂHËƼÎ.Dʅ%3'>ʄĎȂ. v/2ő¼. d _3Ùɩ3X\eHBEDʅ%3ŭ.ʄɸ81. dH Ň. Bȫ. -DcmHȝ1Bɸ8-D/*. 4ʄșũ3ƖȞ½ĢHƩà"D/.ȕEDxd.ʄ. 'șɪ1ȒÌ3ȶÂ?ĎȂ2A*-ÜȊ21D?3/ſĵ. ɰH¢*'ȒÌȮȶ.4ĩ¢GE-D[17,18]ʅĎȂd. 'ƸĹəBtLTƚ2 GMM .ȒÌx. ʄǖŵɞ3n v/2ȶÂHȒ GMM. _YL]2ļ!-ƩàũHɓ¾2ȱŪ"D/.ęə3. -Dʅ. ĎȂd _HĖȺ/ 3.3. ʄn. 'Āà.?ɑĎȂHŏBEDʅĎ. Ȃd _ x 3ǘƼĕĪ½ĢH p(x)/. =9H. śƉ"DȒÌȶÂŋƢH÷ 2 2Ǚ"ʅµÉ4ɩɰtL. '/ʄƩàũ K. 3 GMM M(x)43į.ȕ$Dʅ. T×6êítLT3ʇayh3ľÿ.Cʄƺǥ2Ù ɚEDʅEB3d _4n v3ǤĦ 25msecʄn. væſ 10msec .æƣũ½ƄEʄ%E&E 39 ƕ². .

(8) . 3 MFCC ƸĹəŵǯ¿(ɌĪʄÊɌĪŇ½?ä=)2ąŜ EDʅ%EBƸĹəHǂ- GMM 2ADn vÓ3 ȶÂHȒʅȶÂ2gz }hbe T3Ãǂ?ƍȤ 'ʄĈə3ĎȂd _HÈ2ǂŀ"D3ɫ. ʅ. 3.5

(9) O!C6D& .ʄGMM Hǂ-ȶÂ. 'n vÓ3ȶÂǶ. %3'>ʄ´ȒǗǢ[17,18].?¢GE--ʄęə3ĎȂ. ƅHB!>Ơ>BE'lz ~\cKbT 2. d _.?ɑĎȂȾ2/ȃBED GMM HɔŐ. þ,ɸʄȫȒÌ3ȶÂǶƅ7ąŜ"DŭƢ2+-. 'ʅɩɰtLT¯2+-4ǌȫɪɰƴɰ 3ʈ+3ƹł3ȶÂHÈ2ĎȂ. ɉ8DʅƁǗǢ.ɓǂ. ' GMM 2A*-Ȓ. ʅŭʄêítLT¯2+-4ʄEBʈ+3ƹł2. ĽȊ3ŦìH÷*-DʄLSTM 73µÉ/ ɷ3ɪɰ/. -Ō*-Dʅ3A2. êítLT¯3ǌȫ/ȶÂ. -ȶÂ. 2ʄñʄèò/ȶÂ. -4ʆǟ ʄŊž. 'n v4ȫ. 'ɕ½Hɸ2ź. -ąŜ. -ķBE'. ʇǟɷ3ȬǳȶÂǶƅ3ŵǯ¿H LSTM 2µÉ. 'n vHȫ. 2ąŜ. Ê-ʄñèò3ȣʉ+3ƹł3ȶÂHȒʅ 1ʄɪɰ2+-4șũ3ǟɷ3xdHǂŀ. 'ąŜȞÆ4ƕ3/C.Dʅ. ɩɰtLT¯.ǌȫ/ȶÂ.

(10) ǶƅHŵɞɃ.ħƮÎ 3ȞÆ4ɸ2ɟ. 3ǭĪ.3ȒÌȶÂǶƅHķDʅɩɰtLT24æɆȫȄ. EDʅŭʄȫ2ɟ. -êítLT¯3ȶÂǶƅɓǂ -ʄȋǌȫ4êítLT¯3Ƕ. 3ǌȫHä=ɰĄȥɚEDʄŭ.êítLT4Ɓ. ƅHɓǂ"D/ʄȄǌȫHɩɰtLT¯3ǶƅHɓǂ. 3ǌȫɰĄ(ȥɚED'>ʄʇ+3tLTHÃǂ". "D/.ȫƍ¼.D/ſĵ.Dʅ;'ʄąŜ. D/.ʄȄ3ǌȫ/Ɓ3ǌȫHƖǘ2ÏÂ"D/. 'ǶƅHŵɞɃ.ħƮÎ"D/.Ŋž3ǭĪ7ąŜ.. Copyright (c) 2017 by the Information Processing Society of Japan. ― 118 ―.

(11) DʅħƮÎ/. -4ʄūƻ[12,13]/áƐ2ʄ60 ǝɞ3. Ïɞƚ2n vÓ3ȶÂǶƅżɴ/1*'?3H%. PCM 2ȦĐ n2ȦĐ. 'ʅ;'ʄiL]Qbeǧ3ƓȊ4¶-P. 'ʅ. 3Ïɞ3ȶÂǶƅ/"D»ƿHȒ*'ʅ 4.2 !G#X 3.6 LSTM

(12) O!C6D&. đɽĒ¸.«Â2ɸ@ȫH. LSTM 2ADȶÂǶƅǷà.4ʄÈƙ.êítLTʄɩ. 'ʅɸ2+-4ʈó½ʂɸ8Ċ>BǴGC;.Hʆ. ɰtLT%E&E GMM Hǂ-n v/2ȶÂ. ó/ĐȁʃÙɩ. ʇ+3ǶƅHʆ+3 LSTM 2µÉ. '. -ɸʄȫȒÌ3. -Dɦ3ɰHÙɩ. 'ʅɸ8Ʒ4¶-ĮĲ2Ƿ. -. Dʅd _438 20 ½ǞĪɩ;*'ʅȫ2+-4 3. ƍ¼HȒʅLSTM 4 Recurrent Neural Network(RNN)3œ. â 5 ½ɞ3ɞʄɪȲ"Dɦ3ɰĄHÙɩ. İ/. 38 60 ½ǞĪɩ;*'ʅÙɩ. -śƉE'xd.DʅRNN 3ɞě3{g. beHʄɑÖ3ƹłHȥņ"Dwx~^/ʄ%3ȥņƓ. 'ʅd _4. 'd _4 GMM 3ĎȂ. 2Ãǂ"Dʅ. ɟHÄĸ"DW eB1D LSTM obT2ȀŜ D/.ʄRNN 3Ȱɵ.*'ÍɖƧĉÍɖƵǌƾȺ. 4.3 +Z,G#X. HŏD/.DʅƁǗǢ.4 Forget Gate Hėµ. đɽĒ¸2-ɸ@ȫH. ' LSTM[17]Hǂ'ʅLSTM 2µÉ"DƸĹə4ʄGMM. â½ʄǰ 120 ½ɞ½ǂŀ. 2ADȶÂǶƅ/. H. -êítLT¯4ñèò. 1B~}bT\. -Dɦ3d _.ʄ12. 'ʅÙɚ4Ē¸.ȏċ@ȫ -DǀāHƒŤ. -Ȓ*'ʅȖ. ǌȫɪɰ3 4 ƕ²3 one-hot qTeʄɩɰtL. ɽȄ2ƌċ2ī*-?Bʄ\t enOHŢ "D1. T¯4ǌȫɪɰ3 2 ƕ²3 one-hot qTeķ. 0Ŷƙ3ŖDȍH. BE-Cʄ%EBHɎǶ. Ó3ȶÂ2¢ GMM 3ĎȂ2ǂDʅ%3'>ʄñ. 6 ƕ²3qTe/. '?3. -?B*'ʅ3d _4n v. H¢*'ʅ1ʄGMM 3ȶÂ.ķBE'ñè. ʄèòʄǌȫʂȋǌȫʄȄ3ǌȫHä="8-3ǌ. òǌȫĆ3Ƕƅ4¶-ɪɰ/. ȫʃ3Ȭǳ1ȒÌ}qHȒ*'ʅ38 120 ½ɞ3ɰ. -Ō*'ʅ;. 'ʄȶÂǶƅ.41 GMM 3\XJHǐřµÉƸĹə2. Ąd _2. ¢/?ƍȤ. ítLT¯24ñ èò A6tLTȗǓȄȋɁ3ǌ. 'ʄŋ.}q. ĎȂ2¢Ðŧģ1 D/HǏǎ/. '?3Hɕ. ĎȂ2ADǮĪŦìHĶ3đŮ". -ƁǗǢ.4Ãǂ. 1*'ʅɞě4. 'ȒÌ}q3¸ȨHȕ 1 2Ǚ"ʅê. ȫʂȋǌȫʃ3}qH. ɩɰtLT¯24ȋǌȫ. A6æöȫȄ3ǌȫ"8-2Ė. -ǌȫ}qH. 1 ě. LSTM ě/1*-Dʅ¼Éě4 3 +3i fHŔ. ';'ʄLSTM 3ĎȂ2Ƥǂ"D/HǏǎ/. )ʄßi f4ɸʄȫʄA6ɸʄȫĆ3ƹł. Ɋ ɸʄȫȒÌ3}q?Ȓ*'ʅ38 120 ½ɞ. /Ėļ. HǘƼ.. 3d _3¸ʄɸ4¶3 31%(ǰ 37 ½)ʄȫ4¶. ¼ÉEDʅó4ʄ3 +3i f3ǘƼżĈ/1D. 3 21%(ǰ 25 ½).*'ʅɸ}q4ɸɝĊŵB. -Cʄ%E&E3ȒÌʄƹłB. i f3ȒÌHżǴǎ1¼É/. 'ʅ. ɸǴŵ;.. 'ʅɊʄɸ.1ȫ103Ï. ɞ2+-4}qHĆ. 4. %W d _4ȒÌȶÂ2¢ GMM ×6ȶÂǶƅ3Ƿà2¢ LSTM 3ĎȂ×6%3Ȫ¦3'>2 3 ǟɷǂŀ. 'ʅ. GMM 3ĎȂ4 4.2 ǩ3«ÂȒÌɰ×6 4.3 ǩ3đɽĒȒ. -ʄĪ.?ȫ3@CÚCE5ȫ}q. H. 'ʅ1ʄɸ×6ȫ}q4Jic _ɰ. 3<Hȅ-ÁŬ. -ʄđǀāȒÌɰ3ɕʄ1 ŵɞ½H¢*'ʅ 4.1 %S:) 3 ǟɷ3d _Ùɩ4á3ƓðH¢*-ƕ2Ǚ"ȦĐ. 'ʅIC X `4 Sony Ș ICD-. UX544F H¢ʄYp~Uæƣũ 44.1kHz / Linear. -Dʅ%3'>ʄđɦ3Ȗ. ȕ 1 ȬǳȒÌ}q3¸Ȩ(đɽĒȒÌɰ). Â×6 LSTM 2ADȶÂǶƅ3ǷàHȪ¦"D2ɦ. X `H¢*-đŮ. -. ɽȄ3ȒÌ/Ǉ1DĀà?ȝBE'ʅ. Ìɰ3 2 ǟɷH¢ʅLSTM 3ĎȂ24đɽĒȒÌɰ×6. ítLT4ÒȹƴǾɬƜȘ SH12-iK H¢*'ʅÙɚ4 IC. -Dʅȫ}q4+3. ȫ2Ė. 4.4 ǩ3đǀāȒÌɰH¢ʅ1ʄGMM 2ADȒÌȶ. .Ȓ*'ʅɩɰtLT4 Sony Ș ECM-PC50 H¢*'ʅê. -Â. . . . . . . . .

(13) . . 4.4 +>*G#X đǀāȒÌɰ4ɾʀȄ 4 ŵɞ3ɞʄēĬ¸HƒŤ Ē¸.ɸHÚC1BȫH. '. -Dɦ3ɰĄHÙɚ. '?3.DʅÙɚ¸Ĕ4 10 ŵB 14 ŵ;.Ȓʄ%3. Copyright (c) 2017 by the Information Processing Society of Japan. ― 119 ―.

(14) ɞ4ȏċ@ɹ<ƷHɸ81B 4-6 â.ɪȲH. -D. ƐċÙɚE-Dʅ;'ʄɊ2ŴɸH/*-Dʅ. ƍ¼ĽȊHȪ¦"Dʅ;'ʄ5.3 ǩ.4ƍ¼2Ė"D LSTM 3ŵɞĚɝĦ3ĴɱHȱƇ. 'ʅ. ɸ3ȖɽȄáă3ȫ?ȥɚE-Dʅ%3'>ʄ ȫ. 1BɸH"Dd _?ä;E-DʅÙɚ4ēĬ. ¸ɰǀāHƒŤ ƥ. 'C. 5.1 KBG#O!1FI đɽ4 4.2 ǩ.Ȣ×. -Ȓ*'ʅ%3'>ʄÙɚ2}[PH. 'đɽĒȒÌɰHĎȂ2ʄđǀ. āȒÌɰ3ɕ 1 ŵɞǞĪHȪ¦2¢*'ʅGMM 4"8. -Ɛ1ēĬ¸3ɼɰHɗǈ$'ʅ%39. ʄNJX3ºŸŉ3nI3ɰ fJ3ɝɜɰ10ć. -3xdHƩàũ 256 øĐ. Ň. ǟćƐ1ȉŷɰɗǈ. ·½ŨȒ¿4ĖȠŇ½2Äɠ. â3ŭ2ÑÉ. -Dʅd _Ùɚ2438 11. -'( Ùɚ4àȣ 12 óȒ*'%. 3Ƕƅʄ38 288 ŵɞ3ɰĄķBE'3Ùɩ _3BŴɸÈĶ3ɰĄH2¾C¼ ɞǞHÃǂ H. 'ʅ¾C¼. 'd. -ʄ38 8 ŵ. 'ɰĄ24ɸ ȫ3}q. 'ɸ ȫ}q4đɽĒȒÌɰ/áƐ3þ. ƭ.}qHȒ*'¾C¼. 'ɰĄ2. 'ɸ. 'ʅȪ¦ŕƑ4 F-. measure(¹ƾƼ/ɓàƼ3ȱéħú)Hǂ'ʅȪ¦4n vÓ3ȶÂǶƅ2Ė. -Ȓ*'ʅêítLTʄɩɰ. tLT3%E&E3ǶƅH÷ 4ʄ÷ 5 2Ǚ"ʅ÷ 4 BʄêítLT¯4ñʄA6èò3ȶÂǮĪ /GDʄñ2ɟ. -4Ȫ¦d _3ǸĖũ. 29 ó/ę1/ʄA6êítLTȗǓȄÌ'C. ȫ3}q3¸ȨH÷ 3 2Ǚ"÷ 3 BGD A2ʄ¾C¼. ʄRMZJ3½Ũ. ɻHÌ. 'C. 'ɦ2ǁ!DȔţEɰHǑĖǎ2ȯƍ¼. -'> Recall 2ƛ8- Precision 1*-Dʅ. 'd _ʂ38ʊŵɞ½ʃ3 22ʁʂǰ. 1.8 ŵɞʃ4ȫʄ21ʁʂǰ 1.7 ŵɞʃ4ɸ/1*- Dʅ;' ¾C¼. 'ɰĄ3)ɕʆŵɞ2ñʄè. òʄǌȫ3Ȭǳ1ȒÌ}qH. . '}q3¸ȨH.

(15) . ȕ 2 2Ǚ"ʅ. . 05. 21. . ,*.#". . .

(16). . . 7/. . 68. . . . 35 . ÷ 4 êítLT¯3ȶÂǶƅ. ÷ 3 8 ŵɞ½3d _2.

(17) . '. ɸʄȫ}q3¸Ȩ. . . . . . . .

(18) . .

(19) .

(20). . ȕ 2 ȬǳȒÌ}q3¸Ȩ(đǀāȒÌɰ). . . . 35 . . . ÷ 5 ɩɰtLT¯3ȶÂǶƅ 5.2 28;1FI. 5. I +Z. śƉŋƢHǂ-ɸʄȫ3ƍ¼ǮĪHȪ¦"Dʅ. Ȫ¦4 2 ǟɷȒʅÀ>2ʄ5.1 ǩ.ñʄèòʄǌȫ ȒÌ3ȶÂĽȊHȪ¦"Dʅƕ2ʄ5.2 ǩ. LSTM 2AD ȶÂǶƅHǷà"DŋƢHǂ'ɦ3ɸ×6ȫȒÌ3. ĎȂd _4đɽĒȒÌɰ/đǀāȒÌɰ3 7 ŵɞ½H¢ *'ʅȪ¦24đǀāȒÌɰ3 1 ŵɞ½H¢*'ʅȪ¦ŕ Ƒ4 F-measure Hǂ'ʅLSTM 3µÉƸĹə4ßtLT 3 GMM Hǂ-n v/2ñʄèòʄǌȫ/ȶÂ 'ǶƅH¢*'ʅȕ 3 2 LSTM 3ƏɍA6ĎȂk}. Copyright (c) 2017 by the Information Processing Society of Japan. ― 120 ―.

(21) w _HǙ"ʅ1ʄƛɄĖȺ/. - 3.5 ǩ.Ȣ×. 'l. z ~\cKbT 2þ,ɸʄȫ3ƍ¼?á Ɛ2đŮ. ȕ 3 LSTM 3Əɍ/ĎȂk}w _. -Ȫ¦"Dʅ. ÷ 6 2lz ~\cKbT 2þ,ŭƢʄ A6śƉŋƢ3ǶƅHǙ"ʅ;'ʄƍ¼Ƕƅ3£/. -. µ Éƕ²ũ. 6. ɞě. 1. ¼É{gbeũ. 3. ÷ 7 2Ȫ¦2¢*'ʆŵɞ½3d _3¸ʄ30 ½ɞ½¾. ĎȂƼ. 0.001. C¼. '?3HǙ. Drop out Ƽ. 0.2. "ʅLSTM Hǂ'śƉŋƢ4lz ~\cKbT . ƤĽÎɟũ. ReLu[18]. 2þ,ŋƢ2ƛ8-ɸʄȫ/?2ĈĦ2ǮĪŦì. żɓÎŋƢ. Adam. ŵɞĚɝĦ. 10 sec. 'd _2Ė. -ƍ¼. -DʅƸ2ɸ2ɟ. 'ǶƅHŚǅ. -ĈĦ2ȶÂǮĪã. -. Dʅ%E4÷ 7 B?Dʅ;'ʄƖȡ}q4Ĉ; 2. '?3.D'>2Ɩȡ}q.4ɸ/1* -DÏɞČù". . Dʅ%*'ɸ3ȫH LSTM 4đɦ3ȒÌ0.

(22). C2ƍ¼. . -Dđɦ4ɸ. 1Bȫ. -'ʅ3Ƕƅ4Ĉ;1Ɩȡ}q.ĎȂ ș. . -Dʅ%3ŭ. . .ʄACƖǘ2Ɩȡ}qH"D/.ACƖǘ2ș. . -?ɸ3ȫ/*'ȒÌ3¾CŻGC3ư ɪ1ȒÌHȶÂ.-'/Hŀç ɪ1ȒÌHȶÂ.D?. . . 21. E1ʅ. 05. )+!-&($'% .!.. . ÷ 6 ɸʄȫ3ƍ¼Ƕƅ ŰČŋƢʂlz ~\cKbT ʃ/śƉƢ ʂLSTMʃ3ƛɄ. ". . . %. . . #$. &'% '. LSTM

(23) . 30! ÷ 7. ½ɞ3đǀāȒÌɰ2Ė"Dƍ¼Ƕƅ3£. - ŰČŋƢʂlz ~\cKbT ʃ/śƉƢʂLSTMʃ3ƛɄ . Copyright (c) 2017 by the Information Processing Society of Japan. ― 121 ―.

(24)

(25) . . . . ,. 6. 7$. 1 *.

(26)

(27)

(28) . . %. 25. . 896 .9,. #+ . . . 21 . 05 4. 4. 4. . ! 0 / . . ( ( ( (. 4. ÷ 8 MLP ×6 LSTM 3ŵɞĚɝĦÂ3ȶÂǶ ƅʂF-Measureʃ. 89 497$ .9', &935.

(29) - . : ) " 25. . ÷ 9 ȒÌƍ¼Ƕƅ3£ - LSTM ŵɞĚɝĦ3Ĵɱ . ADŋƢ2ɟ LSTM Hǂ-ȶÂǶƅHǷà"D2'C LSTM 3. ŵɞĚɝĦ3ĴɱHȱƇ. 'ʅ3đɽ.4 LSTM 3ŵ. ɞĚɝĦH 0.05 ǝB 10 ǝ;.Ì 2+- LSTM HĎȂ. ʄĎȂ. -?áƐ2ǮĪʅE4 GMM 2A. DȬǳȒÌȶÂ3ǮĪÕô.D/ȃBEDʅɸƍ. 5.3 LSTM 5U/T01FM7. ¼ǮĪ1DƾȺ4đǀāȒÌɰHȪ¦ĖȺ/ à2ɶȐ2ƾE'ʅ%3ƿǄ/. -%E&E3Āà. 'hbe TH¢*. 'Ā. -ʄđǀāȒÌɰ3ɹɸ. Ʒ4ćǟćƐCʄ%E2ļ!-èòɰ?ćƐÎ. -'. /ȃBEDʅ;'ʄ«ġ?Cï=ɌĪ@ï<ŭ. -ɸʄȫ3ƍ¼HȒ*-ʄ%3ǮĪH?/2ŵɞĚɝ. ę. ąGD(.ȶÂ.11DĀà?ȝBE'ʅ%. Ħ2+-Ȫ¦HȒ*'ʅ;'ʄŵɞƏɍH¢G1gz. E2Ė. }hbe T/. Ìɰ3èòƍ¼ǮĪ4đǀā/ƛ8-ɾ/HǘȮ. Ɛ2Ȫ¦. -ʈěk ^pe(MLP)?á. 'ʅMLP A6 LSTM 3ĎȂʄȪ¦24đɽ. -ʄsceabp\3<Hś¤. Dʅ%3Ƕƅ/. - 5.3 ǩ.Ǚ. -'đɽĒȒ -. 'żǴǎ1ɸƍ¼Ǯ. ĒȒÌɰH¢*'ʅ%3) 2/3 3 90 ½ɞ½HĎȂ2ʄ. Ī?đǀāȒÌɰH¢*'Āà/ƛ8-ǮĪɾ1*-. ƘC3 30 ½ɞ½HȪ¦2¢*'ʅȪ¦ŕƑ4 F-measure. Dʅ. Hǂ'ʅLSTM 3Əɍ/ĎȂk}w _4óƍȩ"D. 2Dèò@ñ3ȶÂǮĪãĻȜ.D/ȃB. '*-ʄɸƍ¼ǮĪHŦì"D24đǀā. ŵɞĚɝĦ4ɢ-ȕ 3 /áƐ.Dʅ1ʄMLP 2+. EDʅ;'ʄèòȶÂ2ɟ. -4ÈĶ 2 n v3Ƕƅ?Ƕà. ȼĻ#. -µÉ2Ãǂ. 'ʅ. -4êítLT3Ʒƿǎ1Ľ. ?ĥ2ȶÂÜȊ.D/4#ʄDǞĪ3. ȯCHȧĔ"DA1ǵ<ĻȜ.D/ȃ-Dʅ. ÷ 8ʄ÷ 9 2 MLP A6 LSTM 3ŵɞĚɝĦÂ3ƍ ¼Ƕƅ/%3£(7 ½ɞ½)HǙ"ʅ÷ 8 BGDA. ;'ʄ5.2 ǩ2-ʄɸ3ȫHđɦ3ȒÌ0. 2ʄŵɞĚɝĦHɛ1E51D90ƍ¼3ǮĪã. C2ƍ¼.'/HǘȮ. -D/GDʅ;'ƁǗǢ.4ʄšɸȒƳHG1. ;1Ɩȡ}q.ĎȂ. ɹƝ2ɟ. 3¾CŻGC3ư. -4ɸ}qH. -1÷ 9 H. 'ʅ%.ķBE'Ƕƅ4Ĉ -?ɸ3ȫ/*'ȒÌ. șɪ1ȒÌHȶÂ.-'/H. ȝD/GDA2ŵɞĚɝĦHɛ"D2+E-ȯƍ¼. ŕ. ƪę. -Dʅ3Ƕƅ4ŵɞĚɝĦHĩD/.n. ȶÂ"DĀà4Ȭǳ1Ɩȡ}qĻȜ21Dʅŭ.ʄ. vÓ3ȶÂǶƅ3k_ 10ŵɞǎXcS\e. śƉƢ4Ȭǳ1Ɩȡ}qH"D/1ɸ3. HĎȂ. -D/Hŀç. -DʅƁƃ4ʄɸ3ȫ/*'șɪ1ȒÌH. -D/ȃBEDʅ;'ʄ. ȫHȶÂ.-CʄE4ĎȂd _ƏǪ3'>2ĻȜ. ɹƝ2ÂɊƖȡ}qH"D/.šɸȒƳHG1. 1}q3X\eÇƪ2ǿDʅ;'ʄ%3ØĖ2A. ɹƝ/ɋĥ3ɸ2ɟ. CȬǳ1Ɩȡ}qH"D/.ACșɪ1ȒÌ3ȶ. Ȃ.Dʅ%3Ƕƅ/ ?. -ȶÂǶƅ3ŵǯ¿3ɒHĎ -ɹƝ/ɋĥ3šɸHȶÂ.D. ÂHſĵ.Dʅ'(. E1ʅ. ʄǏǎ@}q3ȻŒHȃŅ. -Ńɗ2ƍȤ"DĻȜD(Fʅ. 5.4 E śƉƢ3ƍ¼ǶƅH<D/ȫ2ƛ8-ɸ3ƍ¼ǮĪ 1*-Dʅ;'ʄlz ~\cKbT 2. Copyright (c) 2017 by the Information Processing Society of Japan. ― 122 ―.

(30) 6. ƁǗǢ.4ñ èò ǌȫ/*'ɸ ȫȒÌ 2ɟɎ3DȒÌHn vÓ.ȶÂ. 'Ƕƅ2 LSTM. Hɓǂ. -Ƿà"D/.ɸʄȫȒÌHƍ¼"DŋƢ. HśƉ. 'ʅn vÓ3ȶÂǶƅHB!>Đ>'. ȞÆ2ļ!-ąŜ"Dlz ~\cKbT 2þ, ŋƢ/ƛ8-ĈĦ1ǮĪãHǘȮ.'ʅ Ķ3Ȱɵ/. -ʄACćƐ1ƹơHä=d _2Ė. -Ȫ¦"D/BEDʅ/CGʄň{ \ V \/. -ȃ-DɾʀȄ3®ĭǨƿ.4ĺɁ3ƹơ. /ɟ¨3ƨACșɪ1ȒÌʄ£5ʄȫ2+-ȋ½ ǡƎǎ2ȫ ǧ2ȫ. -DʄȄǡƎǎ2ȫ. -DʄHȶÂ. 3'>ʄĎȂd _ƏǪ3X\eHȃŅ *'ACȬǳ1}qH. -DʄTV. '/ɭȜDʅ% ++ó.1. -ACșɪ1ȒÌ3ȶÂ. .D0H½Ƅ"DĐ.Dʅ;'ʄgz } hbe T3Əɍ[19,20]HąŹ"D/.ƍ¼ĽȊã H÷D/?ƍȤ. -Dʅ. ƁǗǢ.4 LSTM 3ǂ'Ƿà2ɟ 2Ȱɵ/1*'ʅ%3Õô/. -ɸƍ¼ǮĪ. -ɸ2ɟɎ"DȬǳȒÌ. 3ȶÂǮĪȃBEDʅ2 Ǧ.ÚC'ūƻ[14]. 4ʄȒÌ3ǹǺŵɞ/n vĦ3ɟ¨Ľ2+-Ȣ× -Cʄň4n vĦHĖȺ3ȒÌ2ļ!-żɓÎ" D/.ȶÂ2ŽË1ƸĹHő¼.D/ȃ-Dʅ Ķ4ȶÂǮĪ3*'ñʄèòHĺ2ʄn vĦ @ƸĹə10HżɓÎ"D/.ƍ¼ǮĪãHǏŕ" Đ.Dʅ. NQ đɽd _Hś¤. -ɲ'ɮĝĈĎľÿĎɕƊĜ´ǁʄ. Ȥȳɲ'ɮĝĈĎľÿĎɕțǃ´ǁʄǼĠ´ǁ2Łȵ. [7] Tapia, Emmanuel Munguia et.al., "Activity recognition in the home using simple and ubiquitous sensors." International Conference on Pervasive Computing, pp.158-175 (2004) [8] Kwapisz, Jennifer R et.al., "Activity recognition using cell phone accelerometers." ACM SigKDD Explorations Newsletter, Vol.12, No.2, pp.74-82 (2011) [9] Ravi, Nishkam et.al., "Activity recognition from accelerometer data." Aaai, Vol.5, No.2005, (2005) [10] Bao, Ling et.al., "Activity recognition from user-annotated acceleration data." International Conference on Pervasive Computing, pp.1-17 (2004) [11] Peng, Ya-Ti et.al., "Healthcare audio event classification using hidden markov models and hierarchical hidden markov models." Multimedia and Expo, 2009. ICME 2009. IEEE International Conference on. IEEE, pp.1218-1221 (2009) [12] ƽƋȆ , "Recurrent Neural Network 2þ,űĥǁƤȒÌ Ȯȶ." ɬċľÿɋªĎōȓÿå, Vol.116, No.189, pp.7-12 (2016) [13] ƽƋȆ , "űĥǁƤȒÌȮȶ3'>3 Recurrent Neural Network Əɍ3ȱƇ." űƁɰɱĎ 2016 ĨųčǗǢǌȕȴƯ ȳūɩ, 2-4-1 pp.1-2 (2016) [14] Ĉ¸Ň , "ŠĤɬȫşɅ^Y2AD~J_LvǁƤ ȒÌȮȶ." ľÿ»ƿĎȳūȭ, Vol.53, No.7, pp.1675-1686 (2012) [15] Zhang, Zhengyou et.al., "Multi-sensory microphones for robust speech detection, enhancement and recognition." ICASSP2004, pp.781784 (2004) [16] Bi, Yin et.al., "Autodietary: A wearable acoustic sensor system for food intake recognition in daily life." IEEE Sensors Journal, Vol.16, No.3, pp.806-816 (2016) [17] Ying, Dongwen et.al., "Voice activity detection based on an unsupervised learning framework." IEEE Transactions on Audio, Speech, and Language Processing, Vol.19, No.8, pp.2624-2633 (2011) [18] Gers, Felix A et.al., "Learning to forget: Continual prediction with LSTM." Neural computation, Vol.12, No.10, pp.2451-1471 (2000) [19] Nair, Vinod et.al., "Rectified linear units improve restricted boltzmann machines." Proceedings of the 27th international conference on machine learning (ICML-10), pp.807-814 (2010) [20] Sutskever, Ilya et.al., "Sequence to sequence learning with neural networks." Advances inneural information processing systems., pp.1-9 (2014) [21] LeCun, Yann et.al., "Neural machine translation by jointly learning to align and translate." Nature, Vol.521, No.7553, pp.436-444 (2015). ;"ʅ. $E4< [1] țƂɨÝ , "ǁɰ/ǀāɰ3áŵÙɚ2ADɾʀȄ3Ȓ ÌA6ɁƹłȶÂ2ɟ"DƍȤ." űƁɰɱĎ 2015 Ĩųč ǗǢǌȕȴƯȳūɩ, 2-4-9 pp.1309-1310 (2015) [2] ďȑǱħ , "ɯ§ȚǬŲûɁƹơȮȶZ\cv2ɟ"D ǗǢ." űƁɰɱĎ 2016 ĨųčǗǢǌȕȴƯȳūɩ, 1-4-4 (2016) [3] ďȑǱħ , "ɁƹơȮȶZ\cv2DɰLqeƍ ¼ŭƢ2ɟ"DƍȤ." űƁɰɱĎ 2016 ĨǛčǗǢǌȕȴƯ ȳūɩ, 2-6-2 (2016) [4] Hochreiter, Sepp et.al., "Long short-term memory." Neural Computation, Vol.9, No.8, pp.1735-1780 (1997) [5] Ġ® , "đǚŕãJp a2ADȮǕǋɾʀȄ3 '>3ÑȱûȸťŞZ\cv3ǗǢɝǌ." ľÿ»ƿĎȳū ȭ , Vol.49, No.1, pp.2-10 (2008) [6] Philipose, Matthai et.al., "Inferring activities from interactions with objects." IEEE pervasive computing, Vol.3, No.4, pp.50-57 (2004). Copyright (c) 2017 by the Information Processing Society of Japan. ― 123 ―.

(31)