[
  {
    "startMs": 1309,
    "text": "[music]"
  },
  {
    "startMs": 13120,
    "text": "Let's just do a couple of quick"
  },
  {
    "startMs": 14480,
    "text": "questions and then we'll jump right in."
  },
  {
    "startMs": 16960,
    "text": "Uh, how many of us in the room here have"
  },
  {
    "startMs": 20880,
    "text": "built voice AI agents?"
  },
  {
    "startMs": 24720,
    "text": "Okay, that's a that's a pretty good"
  },
  {
    "startMs": 27199,
    "text": "audience here. And how many of you guys"
  },
  {
    "startMs": 28960,
    "text": "have built AI agents that have been"
  },
  {
    "startMs": 30640,
    "text": "deployed in production?"
  },
  {
    "startMs": 34640,
    "text": "Not bad. Okay, cool. So, uh we'll talk"
  },
  {
    "startMs": 38160,
    "text": "about what typically happens, right?"
  },
  {
    "startMs": 40399,
    "text": "Like everyone's talking about wise AI"
  },
  {
    "startMs": 42879,
    "text": "agents. Uh the"
  },
  {
    "startMs": 47120,
    "text": "you know, one pill solution to pretty"
  },
  {
    "startMs": 48480,
    "text": "much everything in the world today uh is"
  },
  {
    "startMs": 50320,
    "text": "is wise agents. So, everyone's building"
  },
  {
    "startMs": 52320,
    "text": "one and trying to deploy that. They"
  },
  {
    "startMs": 54719,
    "text": "sound great when you're sort of building"
  },
  {
    "startMs": 56239,
    "text": "that in your dev sort of landscape and"
  },
  {
    "startMs": 59920,
    "text": "then the moment you take this to from a"
  },
  {
    "startMs": 62000,
    "text": "proof of concept to production things"
  },
  {
    "startMs": 64080,
    "text": "start failing. Uh so we'll walk through"
  },
  {
    "startMs": 66640,
    "text": "these five different angles of like how"
  },
  {
    "startMs": 69280,
    "text": "uh or what we have seen uh at PO with"
  },
  {
    "startMs": 73520,
    "text": "VIA agents but just before that a quick"
  },
  {
    "startMs": 76000,
    "text": "uh intro from from my side uh I am Wenke"
  },
  {
    "startMs": 79759,
    "text": "the founder and CEO uh she used the"
  },
  {
    "startMs": 83040,
    "text": "title agent engineering manager u I'm"
  },
  {
    "startMs": 85520,
    "text": "calling myself chief agent officer uh"
  },
  {
    "startMs": 88159,
    "text": "from a from a title standpoint u okay so"
  },
  {
    "startMs": 92079,
    "text": "what is what is uh you know why why are"
  },
  {
    "startMs": 94720,
    "text": "even qualified for this this discussion"
  },
  {
    "startMs": 97040,
    "text": "and and like uh what what are we seeing"
  },
  {
    "startMs": 99920,
    "text": "that a lot of companies don't get to"
  },
  {
    "startMs": 101360,
    "text": "see? I I'll talk a bit about our journey"
  },
  {
    "startMs": 103840,
    "text": "in terms of like how we've uh come along"
  },
  {
    "startMs": 106399,
    "text": "so far and then jump right in. Uh you"
  },
  {
    "startMs": 109040,
    "text": "know we were we've been around for about"
  },
  {
    "startMs": 111680,
    "text": "14 years. Our journey has been a"
  },
  {
    "startMs": 113759,
    "text": "developer API platform and then now an"
  },
  {
    "startMs": 116399,
    "text": "uh you know an AI agent business. We"
  },
  {
    "startMs": 118079,
    "text": "started with voice and SMS APIs back in"
  },
  {
    "startMs": 120399,
    "text": "the day uh 2011 and then uh you know now"
  },
  {
    "startMs": 124479,
    "text": "we are primarily focused on our AI agent"
  },
  {
    "startMs": 127439,
    "text": "offering uh the the full stack on our"
  },
  {
    "startMs": 130800,
    "text": "platform. We we see over a billion voice"
  },
  {
    "startMs": 134000,
    "text": "calls each month across the globe. Uh"
  },
  {
    "startMs": 136800,
    "text": "and which is where we've seen a lot of"
  },
  {
    "startMs": 138400,
    "text": "these uh you know patterns emerge in"
  },
  {
    "startMs": 140400,
    "text": "terms of like how when we work with our"
  },
  {
    "startMs": 141920,
    "text": "customers what happens on their voice"
  },
  {
    "startMs": 144000,
    "text": "agents in in production. Uh we're a uh"
  },
  {
    "startMs": 147360,
    "text": "90 member team and uh we've we have uh"
  },
  {
    "startMs": 151599,
    "text": "50 million funding in the bank. Fun"
  },
  {
    "startMs": 153680,
    "text": "fact, this is not from external VC"
  },
  {
    "startMs": 156000,
    "text": "investors. This is all from being a"
  },
  {
    "startMs": 158000,
    "text": "profitable company having put that cash"
  },
  {
    "startMs": 159920,
    "text": "in the bank over over these years. Uh"
  },
  {
    "startMs": 164879,
    "text": "some customers we we power across the"
  },
  {
    "startMs": 166800,
    "text": "globe. Uh you know, we've just left some"
  },
  {
    "startMs": 168480,
    "text": "some logos in there. But primarily from"
  },
  {
    "startMs": 170800,
    "text": "an offering standpoint, u I would sort"
  },
  {
    "startMs": 173280,
    "text": "of cohort this into three different"
  },
  {
    "startMs": 174959,
    "text": "buckets. One is a programmable AI agent"
  },
  {
    "startMs": 178239,
    "text": "offering. We call it uh I mean it's a"
  },
  {
    "startMs": 180319,
    "text": "speech pipeline, not a true"
  },
  {
    "startMs": 181840,
    "text": "speech-to-pech product yet, but that's a"
  },
  {
    "startMs": 184080,
    "text": "that's a programmable offering. We also"
  },
  {
    "startMs": 185760,
    "text": "have an AI agent studio. It's a no code"
  },
  {
    "startMs": 188400,
    "text": "visual uh builder. And then we like I"
  },
  {
    "startMs": 191840,
    "text": "said, we started with voice APIs. So we"
  },
  {
    "startMs": 194239,
    "text": "obviously have built this out over the"
  },
  {
    "startMs": 196159,
    "text": "last 14 years, the SIP trunking and the"
  },
  {
    "startMs": 198400,
    "text": "audio streaming layers. So we don't rely"
  },
  {
    "startMs": 200319,
    "text": "on other folks for the telefony or the"
  },
  {
    "startMs": 202480,
    "text": "carrier layer. Like that's the"
  },
  {
    "startMs": 204319,
    "text": "breadandbut business we've built over"
  },
  {
    "startMs": 205920,
    "text": "all these years and and that's on top of"
  },
  {
    "startMs": 208640,
    "text": "which our AI uh agent platform sits."
  },
  {
    "startMs": 212400,
    "text": "Okay, with that uh let's get into this,"
  },
  {
    "startMs": 216480,
    "text": "right, which I'm I'm sure since you guys"
  },
  {
    "startMs": 218560,
    "text": "have all built AI agents, you've all"
  },
  {
    "startMs": 220640,
    "text": "seen this or you know built this in in"
  },
  {
    "startMs": 223040,
    "text": "one manner or another and we'll spend"
  },
  {
    "startMs": 224959,
    "text": "more time on this in terms of like how"
  },
  {
    "startMs": 227519,
    "text": "uh the entire pipeline looks, right? Uh"
  },
  {
    "startMs": 230640,
    "text": "what we see with customers is and and"
  },
  {
    "startMs": 232720,
    "text": "I'm sure you guys can all relate to this"
  },
  {
    "startMs": 234159,
    "text": "is you know anyone thinking about AI"
  },
  {
    "startMs": 236560,
    "text": "agents what they do is they pick a bunch"
  },
  {
    "startMs": 238720,
    "text": "of these orchestration frameworks and"
  },
  {
    "startMs": 241040,
    "text": "they do a pretty good job live kit or a"
  },
  {
    "startMs": 242959,
    "text": "pipecat you know build their AI oen on"
  },
  {
    "startMs": 245200,
    "text": "top of that uh they think they can just"
  },
  {
    "startMs": 247760,
    "text": "sort of orchestrate these different four"
  },
  {
    "startMs": 249760,
    "text": "layers speechtoext lm uh and and TTS"
  },
  {
    "startMs": 253920,
    "text": "with turn detection in between and we're"
  },
  {
    "startMs": 256079,
    "text": "off to the races like my AI engine agent"
  },
  {
    "startMs": 257840,
    "text": "works in a in a P and it's good to work"
  },
  {
    "startMs": 260479,
    "text": "in production. Uh typically that's what"
  },
  {
    "startMs": 262400,
    "text": "happens. They sort of measure their"
  },
  {
    "startMs": 264160,
    "text": "latencies and you can see some"
  },
  {
    "startMs": 266000,
    "text": "indicative latencies on on this slide at"
  },
  {
    "startMs": 268560,
    "text": "at each layer and they're like yeah this"
  },
  {
    "startMs": 270800,
    "text": "this uh seems good for me for what I"
  },
  {
    "startMs": 273840,
    "text": "need. So let let's position production"
  },
  {
    "startMs": 276320,
    "text": "and then the the production w uh sort of"
  },
  {
    "startMs": 279040,
    "text": "start to kick in and and you see all"
  },
  {
    "startMs": 281040,
    "text": "sort of failure modes which we are going"
  },
  {
    "startMs": 282560,
    "text": "to spend you know most of the time on on"
  },
  {
    "startMs": 285120,
    "text": "in this talk at least. Uh I've kept some"
  },
  {
    "startMs": 288479,
    "text": "time at the end for Q&A if you guys want"
  },
  {
    "startMs": 290080,
    "text": "to have uh you know questions but we'll"
  },
  {
    "startMs": 292800,
    "text": "jump right in from from this to uh you"
  },
  {
    "startMs": 295360,
    "text": "know different failure modes we see."
  },
  {
    "startMs": 297440,
    "text": "Let's start with"
  },
  {
    "startMs": 300240,
    "text": "you know the first one which everyone"
  },
  {
    "startMs": 301600,
    "text": "talks about like this is the most spoken"
  },
  {
    "startMs": 303759,
    "text": "about failure mode which is latency. U I"
  },
  {
    "startMs": 306880,
    "text": "think we have a few AI agent talks today"
  },
  {
    "startMs": 309039,
    "text": "or AI agent talks today. Um I'm pretty"
  },
  {
    "startMs": 311440,
    "text": "sure like everyone everyone's going to"
  },
  {
    "startMs": 312960,
    "text": "touch upon this specific failure mode"
  },
  {
    "startMs": 315039,
    "text": "which is why I'm bringing this right up"
  },
  {
    "startMs": 317199,
    "text": "uh in in terms of uh you know some like"
  },
  {
    "startMs": 321199,
    "text": "how this entire experience is for uh"
  },
  {
    "startMs": 324479,
    "text": "users right uh typically most folks"
  },
  {
    "startMs": 327680,
    "text": "measure this by time to first audio so"
  },
  {
    "startMs": 330960,
    "text": "the time when you user stop speaking to"
  },
  {
    "startMs": 334880,
    "text": "your agent starts speaking right and I"
  },
  {
    "startMs": 337680,
    "text": "think you've you've probably seen this"
  },
  {
    "startMs": 338880,
    "text": "if you guys have built voice agents on,"
  },
  {
    "startMs": 340720,
    "text": "you know, what uh good or natural feels"
  },
  {
    "startMs": 343360,
    "text": "like, what uh sort of annoying feels"
  },
  {
    "startMs": 347280,
    "text": "like or noticeable feels like, and then"
  },
  {
    "startMs": 348800,
    "text": "what annoying feels like, which is, you"
  },
  {
    "startMs": 351039,
    "text": "know, different tiered steps. Uh we"
  },
  {
    "startMs": 353919,
    "text": "notice, you know, most people want to be"
  },
  {
    "startMs": 357120,
    "text": "under 550 cuz that's what's advertised"
  },
  {
    "startMs": 360160,
    "text": "by, you know, platforms or uh you know,"
  },
  {
    "startMs": 363680,
    "text": "solutions or or or or layers. But I"
  },
  {
    "startMs": 366240,
    "text": "think most end up between 750 to 1.2. uh"
  },
  {
    "startMs": 368960,
    "text": "that's where most of the folks end up"
  },
  {
    "startMs": 370639,
    "text": "at. Uh the really bad performing ones"
  },
  {
    "startMs": 372639,
    "text": "end up you know more than 1.2 and then"
  },
  {
    "startMs": 375039,
    "text": "you start to see users uh hang up. Uh"
  },
  {
    "startMs": 378080,
    "text": "now I I'll share with you like what"
  },
  {
    "startMs": 380160,
    "text": "we've seen practically in in uh"
  },
  {
    "startMs": 383199,
    "text": "production with uh customers using this"
  },
  {
    "startMs": 386080,
    "text": "with at at different layers and then you"
  },
  {
    "startMs": 389039,
    "text": "know solutions to uh some of these. The"
  },
  {
    "startMs": 392800,
    "text": "way we want to think about this layer is"
  },
  {
    "startMs": 395600,
    "text": "sort of a balance between these three"
  },
  {
    "startMs": 397840,
    "text": "which is cost, intelligence and latency,"
  },
  {
    "startMs": 402800,
    "text": "right? And and and why do I bring these"
  },
  {
    "startMs": 405360,
    "text": "three up? Because they're sort of"
  },
  {
    "startMs": 407039,
    "text": "interrelated. I think one of the things"
  },
  {
    "startMs": 408720,
    "text": "I was just chatting with uh you know a"
  },
  {
    "startMs": 410160,
    "text": "couple of folks outside one of the"
  },
  {
    "startMs": 411520,
    "text": "things last one year we've seen lot of"
  },
  {
    "startMs": 413600,
    "text": "innovations lot of intelligence spike on"
  },
  {
    "startMs": 416080,
    "text": "the LLM side of uh things right and most"
  },
  {
    "startMs": 420639,
    "text": "of the you know intelligence has come in"
  },
  {
    "startMs": 422800,
    "text": "in terms of thinking or uh you know"
  },
  {
    "startMs": 425680,
    "text": "reinforcement learning and and so on and"
  },
  {
    "startMs": 427360,
    "text": "so forth the irony with voice agents is"
  },
  {
    "startMs": 430720,
    "text": "like almost always your the the LLM or"
  },
  {
    "startMs": 433840,
    "text": "the agent that's talking has to have"
  },
  {
    "startMs": 436240,
    "text": "thinking turned"
  },
  {
    "startMs": 437759,
    "text": "Right. So all the advancements we've had"
  },
  {
    "startMs": 440400,
    "text": "in the LLM layer in the last one year"
  },
  {
    "startMs": 443599,
    "text": "like none of that even apply here now."
  },
  {
    "startMs": 445440,
    "text": "Right? You obviously you have you know"
  },
  {
    "startMs": 447440,
    "text": "better models that can do you know"
  },
  {
    "startMs": 449120,
    "text": "better instruction following or tool"
  },
  {
    "startMs": 450560,
    "text": "calling but pretty much all of your"
  },
  {
    "startMs": 452240,
    "text": "intelligence that's been built in on the"
  },
  {
    "startMs": 454400,
    "text": "thinking layer is all off by default if"
  },
  {
    "startMs": 456560,
    "text": "you want it to be fast enough. So so"
  },
  {
    "startMs": 458720,
    "text": "that's one of the ironies that we come"
  },
  {
    "startMs": 460160,
    "text": "up with. So then how do you sort of"
  },
  {
    "startMs": 461840,
    "text": "balance intelligent cost and latency?"
  },
  {
    "startMs": 464560,
    "text": "Let's let's look at some of these uh you"
  },
  {
    "startMs": 466319,
    "text": "know options uh that are out there in"
  },
  {
    "startMs": 468479,
    "text": "the market right so and I'm specifically"
  },
  {
    "startMs": 470720,
    "text": "picking LLM because if you looked at the"
  },
  {
    "startMs": 472560,
    "text": "previous chart LLM is u you know sort of"
  },
  {
    "startMs": 477199,
    "text": "your highest latency bucket that adds to"
  },
  {
    "startMs": 480240,
    "text": "this right and uh if you look at you"
  },
  {
    "startMs": 482960,
    "text": "know frontier models which I think most"
  },
  {
    "startMs": 485599,
    "text": "folks start by default your your openi"
  },
  {
    "startMs": 489199,
    "text": "your clouds your geminis u you know p50"
  },
  {
    "startMs": 491919,
    "text": "ttfftd is roughly around 450 to 500 on"
  },
  {
    "startMs": 495599,
    "text": "on a good day and it can get spiky,"
  },
  {
    "startMs": 498160,
    "text": "right? It can it can uh you know P90 P95"
  },
  {
    "startMs": 501680,
    "text": "can go easily upwards of 1.2 1.3 seconds"
  },
  {
    "startMs": 504800,
    "text": "even uh and and that's not good for the"
  },
  {
    "startMs": 507759,
    "text": "overall agent experience."
  },
  {
    "startMs": 510080,
    "text": "So so that so that's your frontier"
  },
  {
    "startMs": 511599,
    "text": "model. Now there's another options which"
  },
  {
    "startMs": 514000,
    "text": "is your your cerebrus or or the gro that"
  },
  {
    "startMs": 517680,
    "text": "is famous and popular for spitting out a"
  },
  {
    "startMs": 519919,
    "text": "lot of tokens or or tokens very fast,"
  },
  {
    "startMs": 522320,
    "text": "right? uh these work but for you to get"
  },
  {
    "startMs": 525680,
    "text": "dedicated latency or time to first token"
  },
  {
    "startMs": 528000,
    "text": "on these you need dedicated capacity and"
  },
  {
    "startMs": 530080,
    "text": "that is really expensive that's where I"
  },
  {
    "startMs": 532080,
    "text": "spoke about the cost uh as as being one"
  },
  {
    "startMs": 534320,
    "text": "of the things to balance right it's"
  },
  {
    "startMs": 536160,
    "text": "really expensive and then like you talk"
  },
  {
    "startMs": 538080,
    "text": "to anyone from the gro team or the"
  },
  {
    "startMs": 539600,
    "text": "cerebrus team they'll tell you you need"
  },
  {
    "startMs": 541600,
    "text": "to book 12 months in advance for"
  },
  {
    "startMs": 543040,
    "text": "dedicated capacity they're booked out"
  },
  {
    "startMs": 544720,
    "text": "for the next 12 months so so that's"
  },
  {
    "startMs": 547200,
    "text": "that's a pretty expensive option and"
  },
  {
    "startMs": 548880,
    "text": "then you really need to be sure that the"
  },
  {
    "startMs": 551440,
    "text": "model you're deploying on some of these"
  },
  {
    "startMs": 552800,
    "text": "infra layers uh will be here 12 months"
  },
  {
    "startMs": 557040,
    "text": "from now and and it's a it's a big"
  },
  {
    "startMs": 558560,
    "text": "investment and a big unknown. So, so"
  },
  {
    "startMs": 561120,
    "text": "what's a realistic option for production"
  },
  {
    "startMs": 563920,
    "text": "grade uh"
  },
  {
    "startMs": 567680,
    "text": "agents that are that are good quality"
  },
  {
    "startMs": 569839,
    "text": "and end up balancing uh three of these u"
  },
  {
    "startMs": 573440,
    "text": "this is what has worked for us u which"
  },
  {
    "startMs": 577200,
    "text": "is the open source models u there are"
  },
  {
    "startMs": 580720,
    "text": "obviously a lot of them in terms of like"
  },
  {
    "startMs": 582320,
    "text": "the variety and and variations you can"
  },
  {
    "startMs": 584399,
    "text": "pick I'm specifically talking about the"
  },
  {
    "startMs": 586240,
    "text": "two we work with u quen 3.5 and gemma"
  },
  {
    "startMs": 590480,
    "text": "four. These are uh you know kind of"
  },
  {
    "startMs": 593600,
    "text": "cutting edge open source models right uh"
  },
  {
    "startMs": 596399,
    "text": "out in the market right now and we've"
  },
  {
    "startMs": 599360,
    "text": "done a lot of benchmarking around this"
  },
  {
    "startMs": 601920,
    "text": "in how they work. It it can be scary to"
  },
  {
    "startMs": 604640,
    "text": "think like okay I have the models now I"
  },
  {
    "startMs": 608080,
    "text": "have to host them you know run them on"
  },
  {
    "startMs": 610640,
    "text": "my own GPUs and so on and so forth but"
  },
  {
    "startMs": 612720,
    "text": "if you are consistently targeting under"
  },
  {
    "startMs": 615200,
    "text": "300 ms u this we've seen this to be a a"
  },
  {
    "startMs": 619279,
    "text": "great option to balance between latency"
  },
  {
    "startMs": 621440,
    "text": "cost and intelligence now some more deep"
  },
  {
    "startMs": 624720,
    "text": "dive here if you're doing only English"
  },
  {
    "startMs": 627279,
    "text": "uh quen 3.5 or GMA both work fine but if"
  },
  {
    "startMs": 630560,
    "text": "you're doing multilingual uh right"
  },
  {
    "startMs": 632560,
    "text": "international audiences different"
  },
  {
    "startMs": 633920,
    "text": "languages uh Gemma 4 is a much better"
  },
  {
    "startMs": 636959,
    "text": "model for that uh we've seen uh token"
  },
  {
    "startMs": 642079,
    "text": "fertility evals essentially what that"
  },
  {
    "startMs": 643920,
    "text": "means is if if I were to dejargonize"
  },
  {
    "startMs": 645600,
    "text": "that is like how many tokens does it"
  },
  {
    "startMs": 647760,
    "text": "take to generate one word in that"
  },
  {
    "startMs": 649279,
    "text": "language okay so Gemma is much much"
  },
  {
    "startMs": 652720,
    "text": "better at least 2.5 to 3x better than"
  },
  {
    "startMs": 655360,
    "text": "quen 3.5 from that perspective so your"
  },
  {
    "startMs": 658560,
    "text": "time to words is much faster on Gemma or"
  },
  {
    "startMs": 662240,
    "text": "everything else equal right on a on a"
  },
  {
    "startMs": 664480,
    "text": "multilingual basis. Now what sizes do"
  },
  {
    "startMs": 667519,
    "text": "you pick at the LLM layer? Uh the"
  },
  {
    "startMs": 670880,
    "text": "mixture of expert usually works fine. Uh"
  },
  {
    "startMs": 673760,
    "text": "the three or four billion mixture of"
  },
  {
    "startMs": 675200,
    "text": "expert usually works fine. The the"
  },
  {
    "startMs": 677440,
    "text": "problem with mixture of expert is like"
  },
  {
    "startMs": 678959,
    "text": "if anyone goes down wants to go down the"
  },
  {
    "startMs": 681040,
    "text": "direction of fine-tuning that can be a"
  },
  {
    "startMs": 683040,
    "text": "challenge uh because fine-tuning mixture"
  },
  {
    "startMs": 684880,
    "text": "of experts models are not easy. Uh you"
  },
  {
    "startMs": 687839,
    "text": "can end up breaking the model uh a a lot"
  },
  {
    "startMs": 690000,
    "text": "of times. So, so that's one challenge we"
  },
  {
    "startMs": 692000,
    "text": "see with Make sure experts, but usually"
  },
  {
    "startMs": 693920,
    "text": "out of the box, it gets you 90% closer"
  },
  {
    "startMs": 697839,
    "text": "to where you want to be like even"
  },
  {
    "startMs": 699360,
    "text": "without any fine-tuning or or or custom"
  },
  {
    "startMs": 702160,
    "text": "work done on the model. Uh so that's the"
  },
  {
    "startMs": 704880,
    "text": "advantage of mixer experts. Uh now, if"
  },
  {
    "startMs": 707440,
    "text": "you want to fine-tune and and you you"
  },
  {
    "startMs": 709120,
    "text": "want to go deeper and say like look, I'm"
  },
  {
    "startMs": 711040,
    "text": "working for a specific domain,"
  },
  {
    "startMs": 712720,
    "text": "healthcare, what have you, right? uh and"
  },
  {
    "startMs": 715200,
    "text": "I want to make sure I I'm able to"
  },
  {
    "startMs": 716560,
    "text": "fine-tune my model. You want to start at"
  },
  {
    "startMs": 718399,
    "text": "least with uh the 8 billion 12 billion"
  },
  {
    "startMs": 720800,
    "text": "at least uh from where we are today."
  },
  {
    "startMs": 722800,
    "text": "Maybe maybe six months from now a 4"
  },
  {
    "startMs": 724640,
    "text": "billion 4 billion model beats the 8"
  },
  {
    "startMs": 727279,
    "text": "billion model uh hands down. But for"
  },
  {
    "startMs": 729279,
    "text": "today uh what we've seen is you minimum"
  },
  {
    "startMs": 733120,
    "text": "need a 8 billion or 12 billion model. Uh"
  },
  {
    "startMs": 735760,
    "text": "cuz you're looking for two things in"
  },
  {
    "startMs": 737040,
    "text": "these models. One obviously fast tokens"
  },
  {
    "startMs": 739920,
    "text": "but uh good instruction following. Okay."
  },
  {
    "startMs": 743519,
    "text": "And the second thing is like very high"
  },
  {
    "startMs": 745839,
    "text": "uh success ratio in tool calling because"
  },
  {
    "startMs": 748480,
    "text": "if you can do these two things well then"
  },
  {
    "startMs": 750800,
    "text": "you are on to like 70 80% there from not"
  },
  {
    "startMs": 753519,
    "text": "even having to fine-tune it fine-tune"
  },
  {
    "startMs": 755440,
    "text": "any model like models will work out of"
  },
  {
    "startMs": 757279,
    "text": "the box right u so so that's uh been our"
  },
  {
    "startMs": 760560,
    "text": "recipe we've actually uh we run two"
  },
  {
    "startMs": 763120,
    "text": "flavors one a fine tune model"
  },
  {
    "startMs": 765760,
    "text": "for specific industries and then for uh"
  },
  {
    "startMs": 769920,
    "text": "you know most generic use cases uh MOE"
  },
  {
    "startMs": 773920,
    "text": "model just works out of the box. Uh"
  },
  {
    "startMs": 775519,
    "text": "there are a few more tips and tricks"
  },
  {
    "startMs": 776800,
    "text": "we'll talk about in the upcoming slides"
  },
  {
    "startMs": 778800,
    "text": "where we see failure models, but but"
  },
  {
    "startMs": 780639,
    "text": "that's where we stand from a from a"
  },
  {
    "startMs": 782959,
    "text": "latency LLM standpoint. Um all right,"
  },
  {
    "startMs": 786399,
    "text": "I'm running tight on time, so I'm going"
  },
  {
    "startMs": 788880,
    "text": "to fast track this. U now there are a"
  },
  {
    "startMs": 791120,
    "text": "couple of other flavors in this. Uh"
  },
  {
    "startMs": 792720,
    "text": "people build agents with a a mixture of"
  },
  {
    "startMs": 795360,
    "text": "models. What they do is you know for u"
  },
  {
    "startMs": 798720,
    "text": "the the talking part of it they have a"
  },
  {
    "startMs": 800399,
    "text": "conversational model which is a much"
  },
  {
    "startMs": 801760,
    "text": "lower smaller model and then you know"
  },
  {
    "startMs": 804399,
    "text": "maybe even a three billion model and"
  },
  {
    "startMs": 806240,
    "text": "then for tool calling they have a much"
  },
  {
    "startMs": 807600,
    "text": "larger model so that they have a"
  },
  {
    "startMs": 809600,
    "text": "improved tool calling success ratio"
  },
  {
    "startMs": 811040,
    "text": "there. Uh"
  },
  {
    "startMs": 815600,
    "text": "sorry the second one is uh assume your"
  },
  {
    "startMs": 819440,
    "text": "transcriptions are going to be brittle"
  },
  {
    "startMs": 821600,
    "text": "like that's that's uh something you want"
  },
  {
    "startMs": 824320,
    "text": "to sort of uh live by when you're"
  },
  {
    "startMs": 827360,
    "text": "building AI agents even if you have the"
  },
  {
    "startMs": 829680,
    "text": "best transcription engine out there and"
  },
  {
    "startMs": 831680,
    "text": "I I I'll show you why right like the the"
  },
  {
    "startMs": 834720,
    "text": "the state-of-the-art transcription"
  },
  {
    "startMs": 836560,
    "text": "engines out out in the market u you know"
  },
  {
    "startMs": 839360,
    "text": "sort of get you to four to 6% word error"
  },
  {
    "startMs": 843279,
    "text": "rate right and this is on known eval"
  },
  {
    "startMs": 845839,
    "text": "sets on real world noisy calls with you"
  },
  {
    "startMs": 850639,
    "text": "know sort of uh accents like people"
  },
  {
    "startMs": 852959,
    "text": "having different sort of accents uh"
  },
  {
    "startMs": 854800,
    "text": "domain vocabulary and so on and so forth"
  },
  {
    "startMs": 857040,
    "text": "like those usually end up in the double"
  },
  {
    "startMs": 858800,
    "text": "digits from a word erate perspective"
  },
  {
    "startMs": 861360,
    "text": "right uh now you obviously you can"
  },
  {
    "startMs": 862880,
    "text": "fine-tune you know pick up an open"
  },
  {
    "startMs": 864399,
    "text": "source model and fine-tune uh but we see"
  },
  {
    "startMs": 867600,
    "text": "typically like what breaks here often"
  },
  {
    "startMs": 869839,
    "text": "and there are patterns s here in terms"
  },
  {
    "startMs": 871279,
    "text": "of what breaks. So, proper nouns,"
  },
  {
    "startMs": 873680,
    "text": "jarens, uh phone numbers like random"
  },
  {
    "startMs": 877040,
    "text": "missing digits with phone numbers, uh"
  },
  {
    "startMs": 879120,
    "text": "wrong substitutions. I I'll walk through"
  },
  {
    "startMs": 880800,
    "text": "some examples of like how you solve for"
  },
  {
    "startMs": 882800,
    "text": "these addresses when you're trying to"
  },
  {
    "startMs": 884639,
    "text": "collect a long address. Uh you know, the"
  },
  {
    "startMs": 887600,
    "text": "the transcription engine could just end"
  },
  {
    "startMs": 889440,
    "text": "up missing some parts of it."
  },
  {
    "startMs": 892320,
    "text": "Code switch languages. I I'll just take"
  },
  {
    "startMs": 894320,
    "text": "a example of a language I speak because"
  },
  {
    "startMs": 896959,
    "text": "that's was easy for me to put on the"
  },
  {
    "startMs": 898639,
    "text": "slide. uh where you know like if you"
  },
  {
    "startMs": 901440,
    "text": "were to sort of take English but written"
  },
  {
    "startMs": 904079,
    "text": "in a different script uh that's what's"
  },
  {
    "startMs": 905920,
    "text": "used for Hindi right like this is"
  },
  {
    "startMs": 908000,
    "text": "English written in that script right"
  },
  {
    "startMs": 910160,
    "text": "whereas like the actual English version"
  },
  {
    "startMs": 911760,
    "text": "of this is hello how are you so if if"
  },
  {
    "startMs": 914079,
    "text": "I'm addressing an audience in a"
  },
  {
    "startMs": 915600,
    "text": "different country where I have code"
  },
  {
    "startMs": 917360,
    "text": "switched languages and I start getting"
  },
  {
    "startMs": 919279,
    "text": "my English in a different uh sort of"
  },
  {
    "startMs": 921839,
    "text": "script everything starts breaking from"
  },
  {
    "startMs": 924399,
    "text": "the transcription engine to the LLM"
  },
  {
    "startMs": 926160,
    "text": "layer and then beyond because your LLM"
  },
  {
    "startMs": 928240,
    "text": "starts then producing output in that"
  },
  {
    "startMs": 929839,
    "text": "sort of script a lot of times and then"
  },
  {
    "startMs": 932240,
    "text": "your TTS messes up. Okay. So, so this is"
  },
  {
    "startMs": 935040,
    "text": "uh very important to be careful about"
  },
  {
    "startMs": 937120,
    "text": "and if you want to build your agent"
  },
  {
    "startMs": 939760,
    "text": "independent of the transcription engine,"
  },
  {
    "startMs": 941600,
    "text": "you need to build a layer that"
  },
  {
    "startMs": 943040,
    "text": "normalizes all of this, right? We'll"
  },
  {
    "startMs": 944720,
    "text": "talk about solutions in a minute. And"
  },
  {
    "startMs": 946160,
    "text": "there is the other case which is Hindi"
  },
  {
    "startMs": 948240,
    "text": "in Latin or or or you know Roman, right?"
  },
  {
    "startMs": 951279,
    "text": "Which is like this is Hindi but it reads"
  },
  {
    "startMs": 954240,
    "text": "English which again messes up everything"
  },
  {
    "startMs": 956880,
    "text": "uh you know downstream. Those are just"
  },
  {
    "startMs": 958320,
    "text": "examples. This applies to, you know,"
  },
  {
    "startMs": 959839,
    "text": "Arabic, Mandarin, uh, Japanese, what"
  },
  {
    "startMs": 962320,
    "text": "have you. Uh, pretty much any language."
  },
  {
    "startMs": 964480,
    "text": "So, what actually moves the needle with"
  },
  {
    "startMs": 967440,
    "text": "a at the transcription layer? Uh, for"
  },
  {
    "startMs": 970240,
    "text": "prop proper nouns, we recommend uh you"
  },
  {
    "startMs": 973120,
    "text": "using not just keyword boosting. I think"
  },
  {
    "startMs": 975199,
    "text": "a lot of transcription engine engines"
  },
  {
    "startMs": 977120,
    "text": "provide you keyword boosting where you"
  },
  {
    "startMs": 978800,
    "text": "can put in specific words into their"
  },
  {
    "startMs": 980800,
    "text": "engine, but doing dynamic keyword"
  },
  {
    "startMs": 982560,
    "text": "boosting. What that means is don't keep"
  },
  {
    "startMs": 984720,
    "text": "the keyword for the entire state of the"
  },
  {
    "startMs": 986480,
    "text": "call. just add that dynamically when you"
  },
  {
    "startMs": 989600,
    "text": "think you need that as an answer so that"
  },
  {
    "startMs": 992320,
    "text": "you get the highest accuracy. Meaning at"
  },
  {
    "startMs": 994399,
    "text": "different states of the call, the"
  },
  {
    "startMs": 995920,
    "text": "transcription engine will have different"
  },
  {
    "startMs": 998320,
    "text": "uh keywords boosted during different"
  },
  {
    "startMs": 1000560,
    "text": "phases, right? Uh and that's what we've"
  },
  {
    "startMs": 1002320,
    "text": "seen works best because if you just"
  },
  {
    "startMs": 1004079,
    "text": "pollute your context of the"
  },
  {
    "startMs": 1005360,
    "text": "transcription engine with tons of"
  },
  {
    "startMs": 1007680,
    "text": "keywords, it'll start hallucinating"
  },
  {
    "startMs": 1009040,
    "text": "again, right? So, so that's what we see"
  },
  {
    "startMs": 1010720,
    "text": "typically working best. Uh"
  },
  {
    "startMs": 1014000,
    "text": "yeah, post-process post-process your"
  },
  {
    "startMs": 1016160,
    "text": "transcripts with an LLM, right? Cuz your"
  },
  {
    "startMs": 1018399,
    "text": "LLM has domain context. Your"
  },
  {
    "startMs": 1020480,
    "text": "transcription engine does not. So a lot"
  },
  {
    "startMs": 1022720,
    "text": "of words that it would say uh I'll give"
  },
  {
    "startMs": 1024959,
    "text": "you some examples may not make sense."
  },
  {
    "startMs": 1026799,
    "text": "This is transcription like a phone"
  },
  {
    "startMs": 1028640,
    "text": "number from a transcription engine."
  },
  {
    "startMs": 1030480,
    "text": "Right? Like what do you think that E is?"
  },
  {
    "startMs": 1033439,
    "text": "Right? If you give it to an LM, it knows"
  },
  {
    "startMs": 1035520,
    "text": "that's a three. Similarly, like what"
  },
  {
    "startMs": 1037199,
    "text": "that one is, it's a digit one. So, so"
  },
  {
    "startMs": 1040000,
    "text": "your transcription engine a lot of times"
  },
  {
    "startMs": 1041439,
    "text": "could mess that up, but when you"
  },
  {
    "startMs": 1043600,
    "text": "postprocess it with the LLM layer, it'll"
  },
  {
    "startMs": 1046000,
    "text": "instantly correct that from a collection"
  },
  {
    "startMs": 1048240,
    "text": "standpoint. I mean, uh, and and the last"
  },
  {
    "startMs": 1050880,
    "text": "one, like I said, uh, transliteration is"
  },
  {
    "startMs": 1053200,
    "text": "your ST output that's sort of u, you"
  },
  {
    "startMs": 1056320,
    "text": "know, multilingual also gets normalized"
  },
  {
    "startMs": 1059840,
    "text": "using either an NLM you first"
  },
  {
    "startMs": 1062320,
    "text": "transliterated or, you know, use some"
  },
  {
    "startMs": 1064480,
    "text": "kind of a neural uh, transliteration"
  },
  {
    "startMs": 1067440,
    "text": "engine. There are a lot of them open"
  },
  {
    "startMs": 1068640,
    "text": "source. You can just pick one of them,"
  },
  {
    "startMs": 1070160,
    "text": "right? Uh that would do all of that work"
  },
  {
    "startMs": 1072080,
    "text": "for you. Send cleaned transcripts"
  },
  {
    "startMs": 1074320,
    "text": "consistently independent of the"
  },
  {
    "startMs": 1076000,
    "text": "transcription engine to your LLM."
  },
  {
    "startMs": 1080559,
    "text": "All right. The third one we typically"
  },
  {
    "startMs": 1081840,
    "text": "see is collecting data. This is where I"
  },
  {
    "startMs": 1084480,
    "text": "think 50 to 60% of AI agents mess up"
  },
  {
    "startMs": 1086960,
    "text": "pretty badly. Uh and like we like to"
  },
  {
    "startMs": 1090559,
    "text": "think of it as"
  },
  {
    "startMs": 1092799,
    "text": "a UX problem. Uh but just for voice. So"
  },
  {
    "startMs": 1096559,
    "text": "think data models uh and not a"
  },
  {
    "startMs": 1099440,
    "text": "transcript coming into an LLM and and"
  },
  {
    "startMs": 1101200,
    "text": "trying to figure out what the transcript"
  },
  {
    "startMs": 1102559,
    "text": "said. So let's take some inspiration"
  },
  {
    "startMs": 1104799,
    "text": "from uh I'm assuming most of us are"
  },
  {
    "startMs": 1108000,
    "text": "developers here um you know take"
  },
  {
    "startMs": 1109600,
    "text": "inspiration from Python's data classes"
  },
  {
    "startMs": 1111679,
    "text": "pantic zod from Typescript or form"
  },
  {
    "startMs": 1115280,
    "text": "fields in the UI right like if you start"
  },
  {
    "startMs": 1117679,
    "text": "thinking of it from that problem"
  },
  {
    "startMs": 1119200,
    "text": "statement we have seen accuracy grow up"
  },
  {
    "startMs": 1121919,
    "text": "from grow from 30% to like 95% from a"
  },
  {
    "startMs": 1125840,
    "text": "data collection standpoint when you"
  },
  {
    "startMs": 1127360,
    "text": "start thinking in that manner. So like"
  },
  {
    "startMs": 1129919,
    "text": "decide your shape before you ask, right?"
  },
  {
    "startMs": 1132559,
    "text": "Like instead of keeping it open-ended,"
  },
  {
    "startMs": 1134240,
    "text": "can you keep it constrained? So can can"
  },
  {
    "startMs": 1137200,
    "text": "a phone number be a phone number type"
  },
  {
    "startMs": 1139360,
    "text": "field? The moment you do that, right,"
  },
  {
    "startMs": 1141760,
    "text": "you know like how many digits it needs"
  },
  {
    "startMs": 1144080,
    "text": "to have. You can do validation on on top"
  },
  {
    "startMs": 1146640,
    "text": "of that, right? And then what sort of"
  },
  {
    "startMs": 1148720,
    "text": "allowed values can even be there. So in"
  },
  {
    "startMs": 1151440,
    "text": "the previous example we saw if an E"
  },
  {
    "startMs": 1153760,
    "text": "comes in in middle of a phone number and"
  },
  {
    "startMs": 1155679,
    "text": "you know it's a phone number you"
  },
  {
    "startMs": 1157200,
    "text": "instantly know like either you smart"
  },
  {
    "startMs": 1159039,
    "text": "guess that to three and confirm that"
  },
  {
    "startMs": 1160559,
    "text": "with a user or you know that's an error"
  },
  {
    "startMs": 1163679,
    "text": "and then you validated that and asked"
  },
  {
    "startMs": 1164960,
    "text": "the user to repeat again right so so"
  },
  {
    "startMs": 1167440,
    "text": "that's I think one of the common"
  },
  {
    "startMs": 1168640,
    "text": "patterns we've seen here from from a a"
  },
  {
    "startMs": 1172000,
    "text": "collection pattern name I think is the"
  },
  {
    "startMs": 1174000,
    "text": "is the interesting one I've just picked"
  },
  {
    "startMs": 1176000,
    "text": "a you know a a hard to pronounce name"
  },
  {
    "startMs": 1180000,
    "text": "like There's no way a human is going to"
  },
  {
    "startMs": 1181840,
    "text": "get this right and and no way a"
  },
  {
    "startMs": 1183760,
    "text": "transcription engine will get this"
  },
  {
    "startMs": 1184960,
    "text": "right. How many ever times you do this"
  },
  {
    "startMs": 1186640,
    "text": "right? So the moment you start thinking"
  },
  {
    "startMs": 1188240,
    "text": "of this as fields and then have rules"
  },
  {
    "startMs": 1190640,
    "text": "and then confirmation mechanisms on on"
  },
  {
    "startMs": 1193520,
    "text": "spelling this uh you know sort of uh"
  },
  {
    "startMs": 1195919,
    "text": "letter by letter only then you kind of"
  },
  {
    "startMs": 1198080,
    "text": "get it right otherwise it's going to"
  },
  {
    "startMs": 1199280,
    "text": "mess up pretty badly in terms of how you"
  },
  {
    "startMs": 1200880,
    "text": "collect this on a voice call and and"
  },
  {
    "startMs": 1203039,
    "text": "that's just an example of you know what"
  },
  {
    "startMs": 1206799,
    "text": "u I'm talking about in terms of the the"
  },
  {
    "startMs": 1208480,
    "text": "data collection piece of it."
  },
  {
    "startMs": 1211520,
    "text": "Another place where it goes badly"
  },
  {
    "startMs": 1214480,
    "text": "dramatically is relative uh values. Date"
  },
  {
    "startMs": 1217360,
    "text": "being one of the examples. If somebody"
  },
  {
    "startMs": 1219600,
    "text": "says next week uh Wednesday 8, 8 could"
  },
  {
    "startMs": 1223520,
    "text": "mean 8:00 a.m. 8:00 p.m. and then"
  },
  {
    "startMs": 1225440,
    "text": "figuring out what that date actually is."
  },
  {
    "startMs": 1227520,
    "text": "Again, now becomes a very constrained"
  },
  {
    "startMs": 1229280,
    "text": "problem. If you knew this was a datetime"
  },
  {
    "startMs": 1231280,
    "text": "field and I I'm collecting a datetime"
  },
  {
    "startMs": 1233840,
    "text": "field and then you take the current date"
  },
  {
    "startMs": 1235520,
    "text": "and then figure out what this value"
  },
  {
    "startMs": 1236960,
    "text": "would be bases that, right? So, so"
  },
  {
    "startMs": 1238799,
    "text": "that's how you want to make sure like uh"
  },
  {
    "startMs": 1240880,
    "text": "you do this with a combination of the"
  },
  {
    "startMs": 1242640,
    "text": "LLM with the tool calling and the tool"
  },
  {
    "startMs": 1244640,
    "text": "calling is doing a lot of this heavy"
  },
  {
    "startMs": 1245919,
    "text": "lifting for you from a from a field"
  },
  {
    "startMs": 1248080,
    "text": "standpoint."
  },
  {
    "startMs": 1251120,
    "text": "Yeah. And then you make you you run like"
  },
  {
    "startMs": 1253200,
    "text": "this from a unit test perspective. So"
  },
  {
    "startMs": 1256080,
    "text": "all of your u evals need to start"
  },
  {
    "startMs": 1259200,
    "text": "treating these fields as unit tests. And"
  },
  {
    "startMs": 1262320,
    "text": "as long as your unit tests uh sort of"
  },
  {
    "startMs": 1265039,
    "text": "validate and pass, you know, your agent"
  },
  {
    "startMs": 1267280,
    "text": "is going to be uh sort of reliable and"
  },
  {
    "startMs": 1269200,
    "text": "repeatable. You don't, you know, run uh"
  },
  {
    "startMs": 1271600,
    "text": "hundreds of end to end agent test cases"
  },
  {
    "startMs": 1273600,
    "text": "just to find out, you know, one field"
  },
  {
    "startMs": 1275360,
    "text": "collection is broken. You do your eval"
  },
  {
    "startMs": 1278320,
    "text": "at a field level and a unit test uh"
  },
  {
    "startMs": 1280320,
    "text": "level."
  },
  {
    "startMs": 1285120,
    "text": "And then yeah, like I said, I think u"
  },
  {
    "startMs": 1287200,
    "text": "you this this mindset makes everything"
  },
  {
    "startMs": 1289600,
    "text": "more structured instead of hoping I'll"
  },
  {
    "startMs": 1292480,
    "text": "put a ton of prompt, keep changing, you"
  },
  {
    "startMs": 1294880,
    "text": "know, the prompt by a few uh characters"
  },
  {
    "startMs": 1297440,
    "text": "every time and somehow my prompt"
  },
  {
    "startMs": 1298960,
    "text": "engineering is going to make LLM much"
  },
  {
    "startMs": 1301120,
    "text": "more instruction tuned and sort of"
  },
  {
    "startMs": 1303280,
    "text": "magically start following some of these"
  },
  {
    "startMs": 1304720,
    "text": "things. So in fact u like I said right"
  },
  {
    "startMs": 1307520,
    "text": "like we have seen us get to 95 97%"
  },
  {
    "startMs": 1310559,
    "text": "accuracy without having to fine-tune a"
  },
  {
    "startMs": 1312720,
    "text": "model right and then and the trick is"
  },
  {
    "startMs": 1314960,
    "text": "basically like just breaking down your"
  },
  {
    "startMs": 1316559,
    "text": "context of what the agent is doing at"
  },
  {
    "startMs": 1318720,
    "text": "that point with specific u states of"
  },
  {
    "startMs": 1321919,
    "text": "what the agent is going through."
  },
  {
    "startMs": 1325039,
    "text": "All right u I'm just going to quickly u"
  },
  {
    "startMs": 1327360,
    "text": "skip through this from a"
  },
  {
    "startMs": 1330480,
    "text": "time standpoint. I just see I got three"
  },
  {
    "startMs": 1332400,
    "text": "more minutes. Um hopefully that's a bug"
  },
  {
    "startMs": 1335679,
    "text": "but but we'll leave it at that. Okay. Um"
  },
  {
    "startMs": 1339360,
    "text": "so so this is the fourth area where we"
  },
  {
    "startMs": 1341200,
    "text": "see issues coming in. Most folks take"
  },
  {
    "startMs": 1344320,
    "text": "the LLM output and then we send it to a"
  },
  {
    "startMs": 1347120,
    "text": "TTS. Obviously I think there are a lot"
  },
  {
    "startMs": 1349120,
    "text": "of good TTS's in the market that take"
  },
  {
    "startMs": 1350880,
    "text": "care of a lot of heavy lifting but a lot"
  },
  {
    "startMs": 1353440,
    "text": "of times it it messes up. Uh what we"
  },
  {
    "startMs": 1356559,
    "text": "recommend and what we've seen is you"
  },
  {
    "startMs": 1358400,
    "text": "usually want to have a normalization"
  },
  {
    "startMs": 1360720,
    "text": "layer between your LLM and what is fed"
  },
  {
    "startMs": 1363679,
    "text": "to a TTS. You don't send your LLM output"
  },
  {
    "startMs": 1367360,
    "text": "directly to a TTS, right? And and we'll"
  },
  {
    "startMs": 1369919,
    "text": "just walk through some examples. The"
  },
  {
    "startMs": 1372559,
    "text": "basics which is strip emojis uh markdown"
  },
  {
    "startMs": 1376960,
    "text": "before before any synthesis into the"
  },
  {
    "startMs": 1378559,
    "text": "TTS. Most orchestration pipelines do"
  },
  {
    "startMs": 1380799,
    "text": "this like you know a live kit or a"
  },
  {
    "startMs": 1382159,
    "text": "pipecat would do that for you if you"
  },
  {
    "startMs": 1383919,
    "text": "just set a few flags. So I but but just"
  },
  {
    "startMs": 1386720,
    "text": "make sure if you're not using them or"
  },
  {
    "startMs": 1388400,
    "text": "buildings from scratch that you've set"
  },
  {
    "startMs": 1390080,
    "text": "this explicitly because you don't want"
  },
  {
    "startMs": 1391760,
    "text": "an emoji showing up on on on something"
  },
  {
    "startMs": 1394080,
    "text": "read out or you know markdown showing up"
  },
  {
    "startMs": 1396080,
    "text": "there."
  },
  {
    "startMs": 1398000,
    "text": "Okay. I think I think some more common"
  },
  {
    "startMs": 1399919,
    "text": "ones uh custom uh dictionaries most TTS"
  },
  {
    "startMs": 1403440,
    "text": "engines provide this to you like how to"
  },
  {
    "startMs": 1405600,
    "text": "pronounce custom words whether it's you"
  },
  {
    "startMs": 1408640,
    "text": "know proper nouns brands uh acronyms and"
  },
  {
    "startMs": 1412080,
    "text": "so on and so forth. So set those in uh"
  },
  {
    "startMs": 1414159,
    "text": "when you go from your LLM to your TTS"
  },
  {
    "startMs": 1416240,
    "text": "output because if you don't, you're"
  },
  {
    "startMs": 1417520,
    "text": "going to mess that up. And I I'll I'll"
  },
  {
    "startMs": 1419200,
    "text": "show you an example of like how we test"
  },
  {
    "startMs": 1420559,
    "text": "that. Uh the the other one is like most"
  },
  {
    "startMs": 1424400,
    "text": "engines also give you speed. So if you"
  },
  {
    "startMs": 1426799,
    "text": "know you're pronouncing an entity, slow"
  },
  {
    "startMs": 1428880,
    "text": "down. Have your agent slow down. So at"
  },
  {
    "startMs": 1431200,
    "text": "point 8x or 7x so that it it's able to"
  },
  {
    "startMs": 1434240,
    "text": "like inunciate on that specific entity"
  },
  {
    "startMs": 1437520,
    "text": "and and doesn't mess up how it's"
  },
  {
    "startMs": 1439200,
    "text": "pronouncing an email or a phone number"
  },
  {
    "startMs": 1441440,
    "text": "or a name letter by letter"
  },
  {
    "startMs": 1445760,
    "text": "and yeah just normalize all the messy"
  },
  {
    "startMs": 1448000,
    "text": "stuff right like emails currency dates"
  },
  {
    "startMs": 1450480,
    "text": "don't leave it to the TTS to do it uh"
  },
  {
    "startMs": 1453679,
    "text": "most of them do it but don't leave it to"
  },
  {
    "startMs": 1455919,
    "text": "the TTS to do it like build your"
  },
  {
    "startMs": 1457840,
    "text": "normalization layer at your end so that"
  },
  {
    "startMs": 1460559,
    "text": "tomorrow you think you need to switch"
  },
  {
    "startMs": 1461919,
    "text": "TTS or you know for whatever reason the"
  },
  {
    "startMs": 1464159,
    "text": "first one's down and you want to use"
  },
  {
    "startMs": 1465600,
    "text": "another TTS you're able to sort of not"
  },
  {
    "startMs": 1468320,
    "text": "rely natively on the TTS's engine but"
  },
  {
    "startMs": 1471600,
    "text": "you are building this in-house uh for"
  },
  {
    "startMs": 1474000,
    "text": "for this to be managed"
  },
  {
    "startMs": 1476720,
    "text": "and then yeah u I think I don't have my"
  },
  {
    "startMs": 1479200,
    "text": "batch here but I I don't have my last"
  },
  {
    "startMs": 1481520,
    "text": "name on that so my first test is if it"
  },
  {
    "startMs": 1483600,
    "text": "cannot pronounce my last name or my"
  },
  {
    "startMs": 1485600,
    "text": "company's name it's already dropping the"
  },
  {
    "startMs": 1487520,
    "text": "ball so my last name is uh Balas"
  },
  {
    "startMs": 1489679,
    "text": "Subramanion and if you cannot pronounce"
  },
  {
    "startMs": 1491760,
    "text": "that using a voice AI agent uh like"
  },
  {
    "startMs": 1495520,
    "text": "that's a check for me. I I know like uh"
  },
  {
    "startMs": 1498400,
    "text": "you know the agent will mess up a lot of"
  },
  {
    "startMs": 1500080,
    "text": "words that uh you know need to be"
  },
  {
    "startMs": 1503919,
    "text": "spelled out day by day. The second one"
  },
  {
    "startMs": 1506559,
    "text": "is our company name Po. So a lot of"
  },
  {
    "startMs": 1508799,
    "text": "engines pronounce pronounce it pivo or"
  },
  {
    "startMs": 1511279,
    "text": "uh pleo and and so on and so forth. But"
  },
  {
    "startMs": 1513840,
    "text": "but I think specifically being able to"
  },
  {
    "startMs": 1515840,
    "text": "control this in your pipeline is super"
  },
  {
    "startMs": 1517840,
    "text": "critical. And then if you're building a"
  },
  {
    "startMs": 1520880,
    "text": "if you're building a customerf facing"
  },
  {
    "startMs": 1522400,
    "text": "product then then um you know sort of"
  },
  {
    "startMs": 1525279,
    "text": "give this option to your customers. All"
  },
  {
    "startMs": 1526880,
    "text": "right I'm just going to skim through the"
  },
  {
    "startMs": 1528640,
    "text": "the the last two slides. U I'm I'm"
  },
  {
    "startMs": 1531120,
    "text": "running badly over time. Uturn"
  },
  {
    "startMs": 1533919,
    "text": "detection. I think this is it own"
  },
  {
    "startMs": 1535279,
    "text": "separate topic but I'm just going to"
  },
  {
    "startMs": 1536960,
    "text": "quickly pull up all the points so you"
  },
  {
    "startMs": 1538559,
    "text": "guys can skim through that and if if you"
  },
  {
    "startMs": 1540640,
    "text": "need a chat u after this we can we can"
  },
  {
    "startMs": 1543279,
    "text": "talk about this. Right. Uh"
  },
  {
    "startMs": 1547520,
    "text": "I'm just going to leave that for like"
  },
  {
    "startMs": 1548960,
    "text": "five seconds and then and then we can"
  },
  {
    "startMs": 1550480,
    "text": "chat about this offline. I'm quite over"
  },
  {
    "startMs": 1552320,
    "text": "time. And then the the the last one is"
  },
  {
    "startMs": 1555520,
    "text": "uh bargin and and back channeling. I"
  },
  {
    "startMs": 1557600,
    "text": "think there's a lot of talk around"
  },
  {
    "startMs": 1558559,
    "text": "speech to speech models that do some of"
  },
  {
    "startMs": 1560240,
    "text": "this, but we've been able to see how we"
  },
  {
    "startMs": 1562559,
    "text": "could do all of this in speech to speech"
  },
  {
    "startMs": 1564000,
    "text": "pipelines. You really don't need a"
  },
  {
    "startMs": 1565840,
    "text": "speech to speech model to do all of this"
  },
  {
    "startMs": 1567360,
    "text": "up. Uh again, I'll just I just put put"
  },
  {
    "startMs": 1569840,
    "text": "this up on the slide and and sort of"
  },
  {
    "startMs": 1572400,
    "text": "close at that. Um"
  },
  {
    "startMs": 1575200,
    "text": "all right I don't think we have time for"
  },
  {
    "startMs": 1577360,
    "text": "questions we can take them offline if"
  },
  {
    "startMs": 1578799,
    "text": "you have any time but uh hopefully this"
  },
  {
    "startMs": 1580480,
    "text": "was helpful and gave you some insights"
  },
  {
    "startMs": 1582400,
    "text": "on uh what we are seeing in productions"
  },
  {
    "startMs": 1584960,
    "text": "uh with billions of calls at scale. All"
  },
  {
    "startMs": 1586960,
    "text": "right thanks"
  }
]