Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
LINEヤフーの音声AIがもたらす未来:ASR/TTSと対話技術の新たな可能性 / LY Co...
Search
Sponsored
·
Ship Features Fearlessly
Turn features on and off without deploys. Used by thousands of Ruby developers.
→
LINEヤフーTech (LY Corporation Tech)
PRO
July 01, 2025
Technology
270
0
Share
LINEヤフーの音声AIがもたらす未来:ASR/TTSと対話技術の新たな可能性 / LY Corporation's Speech AI Vision: Towards Realtime Spoken Dialogue through Advanced ASR and TTS
LINEヤフーの音声認識と音声合成技術を活用した応用事例と、近年注目されているLLM基盤のリアルタイム音声対話技術の自社の取り組みについて紹介します。
LINEヤフーTech (LY Corporation Tech)
PRO
July 01, 2025
More Decks by LINEヤフーTech (LY Corporation Tech)
See All by LINEヤフーTech (LY Corporation Tech)
現場の負担は本当に減る?LINEヤフーの事例で紐解く、問い合わせ自動化の全プロセス
lycorptech_jp
PRO
0
90
「AIエージェントで変わる開発プロセス―レビューボトルネックからの脱却」
lycorptech_jp
PRO
0
900
LINEヤフーにおけるAIOpsの現在地
lycorptech_jp
PRO
6
3.5k
PMとしての意思決定とAI活用状況について
lycorptech_jp
PRO
1
240
Yahoo!ショッピングのレコメンデーション・システムにおけるML実践の一例
lycorptech_jp
PRO
1
330
Rollback from KRaft mode to ZooKeeper mode
lycorptech_jp
PRO
1
150
When an innocent-looking ListOffsets Call Took Down Our Kafka Cluster
lycorptech_jp
PRO
0
180
類似画像検索モデルの開発ノウハウ
lycorptech_jp
PRO
6
1.3k
メタデータ同期に潜んでいた問題 〜 Cache Stampede 時の Cycle Wait を⾒つけた話
lycorptech_jp
PRO
0
220
Other Decks in Technology
See All in Technology
Do Ruby::Box dream of Modular Monolith?
joker1007
1
350
AI時代における技術的負債への取り組み
codenote
1
1.7k
独断と偏見で試してみる、 シングル or マルチエージェント どっちがいいの?
shichijoyuhi
1
110
Practical TypeProf: Lessons from Analyzing Optcarrot
mame
0
930
Do Vibe Coding ao LLM em Produção para Busca Agêntica - TDC 2026 - Summit IA - São Paulo
jpbonson
3
150
M5Stack CoreS3とZephyr(RTOS)で Edge AIっぽいことしてみた
iotengineer22
0
280
ネットワーク運用を楽にするAWS DevOps Agent活用法!! / 20260421 Masaki Okuda
shift_evolve
PRO
2
220
Expiration of Secure Boot Certificates for vSphere Virtual Machines
mirie_sd
0
110
UIライブラリに依存しすぎないReact Native設計を目指して
grandbig
0
120
EMから幅を広げるために最近挑戦していること / Recent challenges I'm undertaking to expand my horizons beyond EM
hiro_torii
1
110
20260428_Product Management Summit_tadokoroyoshiro
tadokoro_yoshiro
12
13k
[最強DB講義]推薦システム | 基礎編
recsyslab
PRO
1
180
Featured
See All Featured
Agile Leadership in an Agile Organization
kimpetersen
PRO
0
140
Getting science done with accelerated Python computing platforms
jacobtomlinson
2
180
世界の人気アプリ100個を分析して見えたペイウォール設計の心得
akihiro_kokubo
PRO
69
39k
SERP Conf. Vienna - Web Accessibility: Optimizing for Inclusivity and SEO
sarafernandez
2
1.4k
Designing Dashboards & Data Visualisations in Web Apps
destraynor
231
54k
Leveraging LLMs for student feedback in introductory data science courses - posit::conf(2025)
minecr
1
240
Faster Mobile Websites
deanohume
310
31k
SEO in 2025: How to Prepare for the Future of Search
ipullrank
3
3.4k
Dealing with People You Can't Stand - Big Design 2015
cassininazir
367
27k
SEOcharity - Dark patterns in SEO and UX: How to avoid them and build a more ethical web
sarafernandez
0
170
The Director’s Chair: Orchestrating AI for Truly Effective Learning
tmiket
1
150
The Web Performance Landscape in 2024 [PerfNow 2024]
tammyeverts
12
1.1k
Transcript
-:$PSQPSBUJPOT4QFFDI"*7JTJPO 5PXBSET3FBMUJNF4QFFDIUP4QFFDI UISPVHI"EWBODFE"43BOE554 4QFFDIBOE"DPVTUJD"*%FQU %BUB4DJFODF(SPVQ +VNQFJ .JZBLF 5BJLJ,JOPTIJUB -*/&ϠϑʔͷԻ"*͕ͨΒ͢ະདྷɿ"43554ͱରٕज़ͷ৽ͨͳՄೳੑ
"HFOEB -:$PSQPSBUJPO`T4QFFDI"* -*/&ϠϑʔͷԻ"*ʹ͍ͭͯ -:$PSQPSBUJPO`T"43554 -*/&ϠϑʔͷԻೝࣝɾԻ߹ͷհ 3FBMUJNF4QFFDIUP4QFFDI ϦΞϧλΠϜ4QFFDIUP4QFFDIٕज़։ൃͷऔΓΈʹ͍ͭͯ
'VUVSF8PSLT ࠓޙͷల
-:$PSQPSBUJPO`T4QFFDI"* -*/&ϠϑʔͷԻ"*ʹ͍ͭͯ 7JEFPBOE"VEJP $POUFOU"OBMZTJT 4QFFDI 3FDPHOJUJPO 4QFFDI (FOFSBUJPO 7JEFP"VEJP$POUFOUT $BMM$FOUFS
.FFUJOH 7PJDF6TFS*OUFSGBDF 7JEFP"VEJP$POUFOUBOE$BMM"OBMZTJT ࣸਅૉࡐఏڙΞϑϩ
-*/&ϠϑʔͷԻೝࣝͱԻ߹ͷհ *OUSPEVDUJPOPG -:$PSQPSBUJPO`T"43554 "43"VUPNBUJD4QFFDI3FDPHOJUJPO 5545FYU5P4QFFDI
:+70*$&4USFBNJOH"43 :+70*$&ετϦʔϛϯάԻೝࣝ &GGJDJFOUMZBEBQUTUPUBSHFUEPNBJOT • "TUSBUFHZCBTFEPODPNQBDUNPEFMTXJUIPVU FYUFSOBMMBOHVBHFNPEFMT • %PNBJOBEBQUBUJPOXJUIPVUUBSHFUBVEJPEBUB 1BJSFETQFFDIUFYUEBUB 6OQBJSFEUFYUEBUB
#BTF.PEFM "EBQUBUJPO .PEFM 4QFFDI 5FYU 5FYU #PPTUTQISBTFXJUIVTFSEJDUJPOBSJFT 4QFFDI 3FDPHOJUJPO 8PVMEZPV MJLFUPTUBSU UIFOBWJHBUJPO WJBUIJTSPVUF 4FSWJDF4QFDJGJD %JDUJPOBSZ :FT /P 1SJPSJUJ[FFYQSFTTXBZT 1SJPSJUJ[FHFOFSBMSPBET ʷ :FBTU ˠ ˓ :FT ʷ ,OPX ˠ ˓ /P ʷ 1SJPSJUJ[FHFOFSBMMPBET ˣ ˓ 1SJPSJUJ[FHFOFSBM SPBET 3FTPMWFTIPNPOZNT ຊ ڮ χ ϗ ϯ ό γ · Ͱ Ϛ σ ͷ ϊʜ 4VSGBDF 3FBE 4VSGBDF 3FBE &OEUP&OE "43 4QFFDI • JF ɾຊڮ χϗϯόγ JTBMPDBUJPOJO5PLZP ɾຊڮ χοϙϯόγ JTBMPDBUJPOJO0TBLB • +PJOUQSFEJDUJPOPGCPUITVSGBDFBOESFBEJOH ಉදهҟԻޠ ޮతͳυϝΠϯదԠ ಈతϢʔβࣙॻʹΑΔϑϨʔζೝࣝڧԽ ˞"CPVU'FBUVSF 'FBUVSF)JHIBDDVSBDZGPSXFCTFBSDIBOE-:$PSQPSBUJPOEPNBJO 'FBUVSF 3FTPMWFTIPNPOZNTBOEDVTUPNJ[FTFBTJMZ 'FBUVSF1SPWJEFT8FC"1*BOEPOEFWJDFNPEVMFT
"DIPSJT&YQSFTTJWF554 "DIPSJT දݱྗ͕๛͔ͳԻ߹ 'FBUVSF$POUSPMFNPUJPOJOUFOTJUZXJUIFYQSFTTJPOTUZMFT 'FBUVSF QSFTFUTQFBLFSPQUJPOTXJUIIVNBOMJLFRVBMJUZ 'FBUVSF1SPWJEFT8FC"1* POEFWJDFNPEVMFTBOEFEJUJOHXFCUPPMT "DIPSJT &EJUPS5FYUUPTQFFDIFEJUJOHUPPM
"DIPSJT &YQSFTTJWFUFYUUPTQFFDI $POUSPM0WFS4QFBLFS &NPUJPO BOE*OUFOTJUZ
ԻೝࣝͷαʔϏε׆༻ࣄྫ • :BIPP +"1"/"QQ`T7PJDF4FBSDI • 7PJDF4FBSDIJTJNQMFNFOUFEJONPTU:BIPP+"1"/4FSWJDFT JFTFSWJDFTJODMVEJOH.BQT 5SBOTJU BOETIPQQJOH :BIPP+"1"/"QQ
J04"OESPJE &YBNQMFTPG"QQMJDBUJPO
Ի߹ͷαʔϏε׆༻ࣄྫ • /BWJHBUJPOWPJDFJO:BIPP+"1"/$BS/BWJHBUJPO"QQ • 0O%FWJDF/FVSBM5FYU5P4QFFDI&OHJOF DBMMFEl"DIPSJT -JUFz • (FOFSBUFBTFDPOEBVEJPXBWFGPSNJO
TFDPOET ˎ 3FBM5JNF'BDUPS 35' JTJOJ1IPOF • 5PNJOJNJ[FBQQTJ[F XFWFJNQMFNFOUFE WBSJPVTPQUJNJ[BUJPOTJOCPUIJOGFSFODFMJCSBSJFTBOENPEFMTJ[F :BIPP+"1"/$BS/BWJHBUJPO"QQ &YBNQMFTPG"QQMJDBUJPO
"MBCBQQGPSSFBMUJNFTQPLFOEJBMPHVF CBTFEPOB--. VOEFSEFWFMPQJOH IUUQTXXXMZDPSQDPKQKBUFDIOPMPHZEFTJHOMBCT &YBNQMFTPG"QQMJDBUJPO ϦΞϧλΠϜԻରͷ࣮ݧΞϓϦ ։ൃத
ϦΞϧλΠϜ4QFFDIUP4QFFDIٕज़։ൃͷऔΓΈ 3FBMUJNF4QFFDIUP4QFFDI
3FBMUJNF4QFFDIUP4QFFDI5SFOE ϦΞϧλΠϜ4QFFDIUP4QFFDIͷٕज़ಈ IUUQTPQFOBJDPNKB+1DIBUHQUPWFSWJFX IUUQTHFNJOJHPPHMFPWFSWJFXHFNJOJMJWF IUUQTNPTIJDIBU IUUQTOVEJBMPHVFHJUIVCJPKNPTIJ
0QFO"* $IBU(15 "EWBODFE7PJDF.PEF ,ZVUBJ .PTIJ /BHPZBVOJW +.PTIJ (PPHMF (FNJOJ-JWF
4QFFDIUP4QFFDI"SDIJUFDUVSF 4QFFDIUP4QFFDIͷϞσϧߏ -BSHF-BOHVBHF.PEFM 5FYU(VJEFE4QFFDI (FOFSBUJPO Low-latency speech generation using audio
tokens or a streaming TTS module "VEJP"EBQUFS 4QFFDI &ODPEFS .PEBMJUZBMJHONFOU CFUXFFOUFYUBOEBVEJP 1SPNQU 4QFFDI 4JOHMF4USFBN 6TFSTTQFFDIPOMZ .VMUJ4USFBN 6TFSTTQFFDI --.HFOFSBUFETQFFDI
1SPTPG*OUFHSBUJOH4QFFDI&ODPEFS XJUI--.T Իͱ--.Tͷ౷߹ʹΑΔར -FWFSBHF--. $BQBCJMJUJFT 1SPNQU%SJWFO 'MFYJCJMJUZ #ZQBTT "43&SSPST --.ͷߴͳج൫ೳྗͷ׆༻
ϓϩϯϓτʹΑΔߴ͍ΧελϚΠζੑ ԻೝࣝޡΓͷӨڹΛճආ
&WBMVBUJPOPG5BTL1FSGPSNBODF 4QFFDI--.ͷλεΫੑೳͷධՁ JOQVU +42V"% 2VFTUJPO"OTXFS DIBS@G "-5 5SBOTMBUJPOGSPNKQ UPFO
#FSU4DPSF (SPVOEUSVUIUFYU UFYUUPUFYU 5SBOTDSJCFEUFYU UFYUUPUFYU 4QFFDI TQFFDIUPUFYU --.HFNNBCJU "43NPEFMXIJTQFSTNBMM 4QFFDI&ODPEFSXIJTQFSTNBMM 5SBJOJOH5PPMLJU4-".--. &WBMVBUJPO5PPMLJUMMNKQFWBM 4-".--.IUUQTHJUIVCDPN9-"/$&4-".--. MMNKQFWBMIUUQTHJUIVCDPNMMNKQMMNKQFWBM $PNQBSBCMF QFSGPSNBODFPO USBOTMBUJPO UBTL #FUUFS QFSGPSNBODFPO 2"UBTL
&WBMVBUJPOPG*OGFSFODF4QFFE ਪͷධՁ 0 10 20 30 40 50
60 vllm slam-llm (transformers) Generated Characters per Second W--.,XPO 8PPTVL FUBM&GGJDJFOUNFNPSZNBOBHFNFOUGPSMBSHFMBOHVBHFNPEFMTFSWJOHXJUIQBHFEBUUFOUJPO1SPDFFEJOHTPGUIFUI4ZNQPTJVNPO0QFSBUJOH4ZTUFNT1SJODJQMFT 'BTUFS YGBTUFS (FOFSBUFT)PXDBO*IFMQZPVUPEBZ JOTFDPOET /VNCFSPG5PLFOT
'VUVSF8PSLT ࠓޙͷల 8FBSFEFWFMPQJOH • 3FBMUJNF4QFFDIUP4QFFDI JOUFHSBUJPOXJUI--. • .VMUJMJOHVBM4QFFDI5P5FYU5FYU5P4QFFDI 7PJDF$POUSPMJOB$BS )VNBOMJLFBOE/BUVSBM
$POWFSTBUJPOBM4FBSDI 4QPLFO%JBMPHVFWJB$BMM "*"HFOU 03 4FBSDI 8FBUIFS 1PEDBTU "*"HFOU "*"HFOU "*"HFOU
EOP