1 00:00:00,05 --> 00:00:03,00 - [Instructor] I'm going to show you a demo 2 00:00:03,00 --> 00:00:06,06 of a machine learning audio classification model. 3 00:00:06,06 --> 00:00:08,00 And we are going to use this platform 4 00:00:08,00 --> 00:00:09,06 called Teachable Machine. 5 00:00:09,06 --> 00:00:13,07 So let's pick audio project. 6 00:00:13,07 --> 00:00:16,08 So a machine learning model 7 00:00:16,08 --> 00:00:21,02 is basically classifying different set of data. 8 00:00:21,02 --> 00:00:24,09 So we are going to focus on an AI machine learning model 9 00:00:24,09 --> 00:00:27,05 of sound, voice, or audio. 10 00:00:27,05 --> 00:00:30,07 And Teachable Machine that I'm using here is from Google. 11 00:00:30,07 --> 00:00:32,08 It's free and experimental part of Google. 12 00:00:32,08 --> 00:00:34,05 So you don't need to sign into Google 13 00:00:34,05 --> 00:00:37,04 and they're not collecting any data. 14 00:00:37,04 --> 00:00:41,07 So we're going to do a live training of an Edge AI model. 15 00:00:41,07 --> 00:00:43,02 Right now when we do training, 16 00:00:43,02 --> 00:00:45,05 it is going to be on a computer. 17 00:00:45,05 --> 00:00:48,05 It can be hosted on an Edge device 18 00:00:48,05 --> 00:00:50,04 or it can be hosted on the cloud. 19 00:00:50,04 --> 00:00:54,05 So when it sits as inference model on an Edge device, 20 00:00:54,05 --> 00:00:55,04 it becomes Edge AI. 21 00:00:55,04 --> 00:01:00,02 So we're going to classify different sounds. 22 00:01:00,02 --> 00:01:04,04 So sound is nothing but digitized data 23 00:01:04,04 --> 00:01:06,03 of different frequencies. 24 00:01:06,03 --> 00:01:08,04 Let's organize and group different sounds 25 00:01:08,04 --> 00:01:12,09 and build a model to recognize one particular sound. 26 00:01:12,09 --> 00:01:15,09 I want you to think about your environment, 27 00:01:15,09 --> 00:01:18,03 be it a factory or a home setting. 28 00:01:18,03 --> 00:01:20,07 Sound is everywhere. 29 00:01:20,07 --> 00:01:24,00 How can you use it to create a useful use case? 30 00:01:24,00 --> 00:01:28,04 Say water is dripping and a pipe is leaking 31 00:01:28,04 --> 00:01:30,08 or water is overflowing in a dam. 32 00:01:30,08 --> 00:01:35,01 Dripping versus overflowing water are different sounds. 33 00:01:35,01 --> 00:01:37,00 A squeaky floor in construction 34 00:01:37,00 --> 00:01:38,09 would make a totally different sound. 35 00:01:38,09 --> 00:01:41,05 So think about the sound in your environment 36 00:01:41,05 --> 00:01:43,06 and what do you want to separate out? 37 00:01:43,06 --> 00:01:45,07 In this demo, we want to build a model 38 00:01:45,07 --> 00:01:48,03 to identify, alert you, 39 00:01:48,03 --> 00:01:50,05 or predict something that's going to break 40 00:01:50,05 --> 00:01:53,01 based on identifying a sound. 41 00:01:53,01 --> 00:01:55,00 So materials could be delivered 42 00:01:55,00 --> 00:01:56,08 by an autonomous mobility robot 43 00:01:56,08 --> 00:01:59,02 because it's filled a bin or a pallet 44 00:01:59,02 --> 00:02:00,09 and you may want to stop it. 45 00:02:00,09 --> 00:02:02,09 That's a different kind of sound. 46 00:02:02,09 --> 00:02:06,02 So you could automate this or notify someone. 47 00:02:06,02 --> 00:02:08,00 Those are all product features. 48 00:02:08,00 --> 00:02:09,06 And these are all my examples. 49 00:02:09,06 --> 00:02:10,09 Later, we will do a challenge 50 00:02:10,09 --> 00:02:14,07 and you get to build on it with your own examples. 51 00:02:14,07 --> 00:02:17,01 So coming back to machine learning, 52 00:02:17,01 --> 00:02:20,02 machine learning is about organizing datasets 53 00:02:20,02 --> 00:02:22,09 into multiple groups or classes. 54 00:02:22,09 --> 00:02:28,03 So here, we have one class called background noise. 55 00:02:28,03 --> 00:02:31,07 So all I'm going to do is click on mic. 56 00:02:31,07 --> 00:02:34,02 I've already given permission to use my mic. 57 00:02:34,02 --> 00:02:35,04 So I'm going to be quiet, 58 00:02:35,04 --> 00:02:38,09 so my background is going to be recorded for 20 seconds. 59 00:02:38,09 --> 00:02:42,09 So then I just press extract sample 60 00:02:42,09 --> 00:02:46,05 and it says teachable machine wants 20 minimum samples. 61 00:02:46,05 --> 00:02:49,00 So it's extracted the samples for me. 62 00:02:49,00 --> 00:02:51,04 So I'm done. 63 00:02:51,04 --> 00:02:54,09 So next, I'm going to create a class called bell. 64 00:02:54,09 --> 00:02:57,04 See, here what we are going to do, 65 00:02:57,04 --> 00:02:59,08 the class can be called any name, it's just a name, 66 00:02:59,08 --> 00:03:02,02 but it's a group of one type of data 67 00:03:02,02 --> 00:03:04,01 that I'm going to collect. 68 00:03:04,01 --> 00:03:07,03 What I want to do is live training with you. 69 00:03:07,03 --> 00:03:09,06 So I have actually brought a real bell. 70 00:03:09,06 --> 00:03:11,09 Do you hear that? 71 00:03:11,09 --> 00:03:13,00 Yeah. 72 00:03:13,00 --> 00:03:17,00 So you have a choice to click upload 73 00:03:17,00 --> 00:03:21,03 and you could have your dataset of some sound that you want 74 00:03:21,03 --> 00:03:23,08 to be classified and identified. 75 00:03:23,08 --> 00:03:26,03 You could very well use that. 76 00:03:26,03 --> 00:03:30,05 Or you can click on mic and it will record, 77 00:03:30,05 --> 00:03:32,07 just like we did with the background sound, 78 00:03:32,07 --> 00:03:36,03 it can record a sound that you're going to produce. 79 00:03:36,03 --> 00:03:40,09 So I have brought a bell with me and live in this demo, 80 00:03:40,09 --> 00:03:43,06 I'm going to generate eight minimum samples, 81 00:03:43,06 --> 00:03:47,00 we can do more, of the bell sound. 82 00:03:47,00 --> 00:03:48,06 So think of what do you have. 83 00:03:48,06 --> 00:03:51,03 I could have brought my ukulele here, 84 00:03:51,03 --> 00:03:53,08 it won't fit in this room for the demo, 85 00:03:53,08 --> 00:03:56,06 but you can pretty much bring anything. 86 00:03:56,06 --> 00:03:59,05 Think of an object around you, what noise you're looking at, 87 00:03:59,05 --> 00:04:00,09 or you look at your work and say, 88 00:04:00,09 --> 00:04:03,00 I want to capture this noise 89 00:04:03,00 --> 00:04:05,06 and record that sound, create a dataset 90 00:04:05,06 --> 00:04:07,05 and bring it and upload it. 91 00:04:07,05 --> 00:04:13,08 Okay, so let's do the live collection of the bell sound. 92 00:04:13,08 --> 00:04:14,09 Here we go. 93 00:04:14,09 --> 00:04:17,08 And it's saying that it will record 94 00:04:17,08 --> 00:04:19,06 only two seconds at a time. 95 00:04:19,06 --> 00:04:22,05 So I'm going to do that multiple times. 96 00:04:22,05 --> 00:04:25,06 So right when I'm clicking record, 97 00:04:25,06 --> 00:04:29,01 I'm going to make the bell sound. 98 00:04:29,01 --> 00:04:31,04 (bell ringing) 99 00:04:31,04 --> 00:04:35,06 Okay, and extract two samples and it says eight minimum, 100 00:04:35,06 --> 00:04:38,07 so I'm going to do that more times. 101 00:04:38,07 --> 00:04:41,03 (bell ringing) 102 00:04:41,03 --> 00:04:43,05 Extract sample. 103 00:04:43,05 --> 00:04:47,02 (bell ringing) 104 00:04:47,02 --> 00:04:48,03 Here we go. 105 00:04:48,03 --> 00:04:52,03 (bell ringing) 106 00:04:52,03 --> 00:04:53,07 One more time. 107 00:04:53,07 --> 00:04:56,06 (bell ringing) 108 00:04:56,06 --> 00:04:59,01 I just want to make 20 samples, 109 00:04:59,01 --> 00:05:01,06 so I'm going to do one more time. 110 00:05:01,06 --> 00:05:04,00 (bell ringing) 111 00:05:04,00 --> 00:05:05,02 Here we go. 112 00:05:05,02 --> 00:05:08,01 So we have about 10 samples of the bell 113 00:05:08,01 --> 00:05:10,08 and we're ready with two classes. 114 00:05:10,08 --> 00:05:14,07 So machine learning is about classifying 115 00:05:14,07 --> 00:05:18,00 multiple sounds in different buckets called classes. 116 00:05:18,00 --> 00:05:19,02 And that's what we've done. 117 00:05:19,02 --> 00:05:22,04 If you want, we can add a third class and a fourth class 118 00:05:22,04 --> 00:05:25,02 and test the model, right? 119 00:05:25,02 --> 00:05:29,01 So we're just going to click on train. 120 00:05:29,01 --> 00:05:31,08 I want to show you one thing before we do train. 121 00:05:31,08 --> 00:05:36,07 Here, the defaults are set and epoch is a number. 122 00:05:36,07 --> 00:05:39,04 By default, it's best practice heuristics 123 00:05:39,04 --> 00:05:42,00 to have 50 as the epoch count. 124 00:05:42,00 --> 00:05:43,03 So I'm going to leave it at that. 125 00:05:43,03 --> 00:05:46,06 Epoch is basically the number of times 126 00:05:46,06 --> 00:05:49,07 each data in here, each sound in here. 127 00:05:49,07 --> 00:05:51,06 Can you see how beautifully it is digitized 128 00:05:51,06 --> 00:05:54,04 and you can see the digital sound difference? 129 00:05:54,04 --> 00:05:58,08 Each of this, how many times is it scanned 130 00:05:58,08 --> 00:06:01,01 for learning by the machine learning model 131 00:06:01,01 --> 00:06:03,09 is the epoch count, okay? 132 00:06:03,09 --> 00:06:05,07 So I'm going to leave it at that. 133 00:06:05,07 --> 00:06:08,02 I'm just going to say train the model. 134 00:06:08,02 --> 00:06:10,08 And it says it's preparing the training data. 135 00:06:10,08 --> 00:06:11,08 It's very less data. 136 00:06:11,08 --> 00:06:14,06 It should do that very, very quickly. 137 00:06:14,06 --> 00:06:18,02 And did the FFT box. 138 00:06:18,02 --> 00:06:19,08 So what do you think is happening 139 00:06:19,08 --> 00:06:21,09 right here on the right side? 140 00:06:21,09 --> 00:06:26,06 So this is where the output of this model is produced. 141 00:06:26,06 --> 00:06:29,00 I have left the microphone on. 142 00:06:29,00 --> 00:06:32,08 So as I'm talking, it is capturing sound 143 00:06:32,08 --> 00:06:33,09 and it's running inference and say, 144 00:06:33,09 --> 00:06:36,04 I have a train model, let me test it. 145 00:06:36,04 --> 00:06:39,04 So it is testing my sound 146 00:06:39,04 --> 00:06:42,09 and it's trying to place it as background or bell. 147 00:06:42,09 --> 00:06:44,05 Why is it? 148 00:06:44,05 --> 00:06:46,04 What's going on here? 149 00:06:46,04 --> 00:06:50,04 So if I'm quiet, 150 00:06:50,04 --> 00:06:54,04 okay, 99% confidence it's background noise. 151 00:06:54,04 --> 00:06:57,05 And if I ring a bell, (bell ringing) 152 00:06:57,05 --> 00:07:00,03 Ooh, 100% confidence it's bell. 153 00:07:00,03 --> 00:07:03,00 So it's able to recognize the bell and background sounds. 154 00:07:03,00 --> 00:07:07,09 Those are in its learning, but every AI is narrow AI, 155 00:07:07,09 --> 00:07:11,01 which means it knows only what it's been trained 156 00:07:11,01 --> 00:07:12,06 in its training data. 157 00:07:12,06 --> 00:07:15,00 So since I've given only background sound and bell sound, 158 00:07:15,00 --> 00:07:16,02 it knows only that. 159 00:07:16,02 --> 00:07:18,06 So when I'm talking, it hears my voice 160 00:07:18,06 --> 00:07:22,07 and it is thinking, is this background or is this bell? 161 00:07:22,07 --> 00:07:24,01 And it is trying to fit that 162 00:07:24,01 --> 00:07:26,09 into the model's learning brain. 163 00:07:26,09 --> 00:07:29,04 So that's what is happening. 164 00:07:29,04 --> 00:07:32,09 So we have collected our data. 165 00:07:32,09 --> 00:07:35,01 We have organized them in classes. 166 00:07:35,01 --> 00:07:41,08 We have trained the model and we have now tested that. 167 00:07:41,08 --> 00:07:45,08 So one thing to remember is since it's recognizing my voice 168 00:07:45,08 --> 00:07:50,01 and as one of the two classes it has, 169 00:07:50,01 --> 00:07:52,04 you'll have to think about this in work. 170 00:07:52,04 --> 00:07:56,02 If you want a particular sound to be captured, 171 00:07:56,02 --> 00:07:59,01 like many examples that I gave you earlier, 172 00:07:59,01 --> 00:08:01,06 you might want to create an extra class 173 00:08:01,06 --> 00:08:03,04 of other things that happens. 174 00:08:03,04 --> 00:08:06,04 If a person would be walking, that walking sound, 175 00:08:06,04 --> 00:08:08,01 or there will be a train whistle 176 00:08:08,01 --> 00:08:09,06 that happens at certain time. 177 00:08:09,06 --> 00:08:12,01 If there is going to be other sounds 178 00:08:12,01 --> 00:08:14,08 from which you want to isolate the sound, 179 00:08:14,08 --> 00:08:18,09 you might want to train the model in a lot of other sounds, 180 00:08:18,09 --> 00:08:21,02 the background disturbances and people walking 181 00:08:21,02 --> 00:08:22,05 and that kind of thing 182 00:08:22,05 --> 00:08:25,01 so that it makes your model more accurate 183 00:08:25,01 --> 00:08:27,03 when you actually give your sample. 184 00:08:27,03 --> 00:08:30,04 It's not trying to force fit into that one particular bucket 185 00:08:30,04 --> 00:08:32,07 because now it's learned other things 186 00:08:32,07 --> 00:08:34,09 and it will say, no, no, this is a person walking. 187 00:08:34,09 --> 00:08:36,01 This is the train whistle. 188 00:08:36,01 --> 00:08:38,00 It's not what I'm looking for. 189 00:08:38,00 --> 00:08:40,07 So you can create a lot of different classes 190 00:08:40,07 --> 00:08:44,01 simply by doing add a class. 191 00:08:44,01 --> 00:08:46,04 And the more data that you give, 192 00:08:46,04 --> 00:08:49,07 the model becomes more accurate, okay? 193 00:08:49,07 --> 00:08:51,05 So the next step, the final step 194 00:08:51,05 --> 00:08:52,08 is we're going to export the model. 195 00:08:52,08 --> 00:08:55,00 So we train the model. 196 00:08:55,00 --> 00:08:56,06 So typically at work, 197 00:08:56,06 --> 00:08:59,03 one person will not be doing all of this. 198 00:08:59,03 --> 00:09:01,04 I want you to learn Edge AI. 199 00:09:01,04 --> 00:09:03,02 I want you to learn the different models. 200 00:09:03,02 --> 00:09:05,03 I want you to learn training and inference 201 00:09:05,03 --> 00:09:07,02 and think about the product and everything 202 00:09:07,02 --> 00:09:09,04 so that you get a very solid foundation 203 00:09:09,04 --> 00:09:11,01 in Edge AI in this course. 204 00:09:11,01 --> 00:09:12,06 So when we export the model, 205 00:09:12,06 --> 00:09:14,01 what we are doing is we're just trying 206 00:09:14,01 --> 00:09:15,09 to use the model we created. 207 00:09:15,09 --> 00:09:18,05 So when we're training, it's training model. 208 00:09:18,05 --> 00:09:19,08 Then we tested it. 209 00:09:19,08 --> 00:09:22,03 And then now we're actually using this 210 00:09:22,03 --> 00:09:23,04 as an inference model. 211 00:09:23,04 --> 00:09:26,04 Inference model is nothing but a ready model 212 00:09:26,04 --> 00:09:28,06 ready for production, right? 213 00:09:28,06 --> 00:09:32,00 So we could upload the model. 214 00:09:32,00 --> 00:09:33,09 So what's happening when we upload the model, 215 00:09:33,09 --> 00:09:35,08 it is actually sent to the cloud 216 00:09:35,08 --> 00:09:38,05 and Teachable Machine actually hosts this somewhere 217 00:09:38,05 --> 00:09:39,08 in this URL. 218 00:09:39,08 --> 00:09:41,01 You can copy this. 219 00:09:41,01 --> 00:09:42,04 What do you think you can do with this? 220 00:09:42,04 --> 00:09:43,08 You train the model. 221 00:09:43,08 --> 00:09:46,01 You want it to recognize the bell sound 222 00:09:46,01 --> 00:09:50,02 or the bin sound or a dripping water sound. 223 00:09:50,02 --> 00:09:52,09 And you could actually put this on the cloud 224 00:09:52,09 --> 00:09:55,05 and share this link with somebody 225 00:09:55,05 --> 00:09:57,09 who is taking this to your customer 226 00:09:57,09 --> 00:09:59,07 and say, hey, let's test this out 227 00:09:59,07 --> 00:10:01,02 with real sound in the field. 228 00:10:01,02 --> 00:10:04,01 You could do that remotely in a cloud setting 229 00:10:04,01 --> 00:10:06,02 because you're going to share this link. 230 00:10:06,02 --> 00:10:08,09 So that's one thing you can do by putting it in the cloud. 231 00:10:08,09 --> 00:10:13,06 Or you could actually click download and download this model 232 00:10:13,06 --> 00:10:15,02 and then put it in a mobile app 233 00:10:15,02 --> 00:10:17,04 or somewhere else where you want it to integrate 234 00:10:17,04 --> 00:10:18,08 into your workflow. 235 00:10:18,08 --> 00:10:20,07 And they also give you the source code. 236 00:10:20,07 --> 00:10:24,03 You can copy that or you have your data scientists copy that 237 00:10:24,03 --> 00:10:26,05 and then build on this 238 00:10:26,05 --> 00:10:29,08 or integrate into other products that you're using 239 00:10:29,08 --> 00:10:31,09 to create the right kind of experience you want. 240 00:10:31,09 --> 00:10:36,03 So when this AI model recognizes a sound, 241 00:10:36,03 --> 00:10:38,09 what kind of automation do you want it to do? 242 00:10:38,09 --> 00:10:40,05 What kind of things do you want to do? 243 00:10:40,05 --> 00:10:42,08 So I'm excited to see what you're going to do 244 00:10:42,08 --> 00:10:45,03 with the audio classification on an Edge device. 245 00:10:45,03 --> 00:10:47,05 The minute you put it on an Edge device, 246 00:10:47,05 --> 00:10:50,00 it becomes Edge AI running inference. 247 00:10:50,00 --> 00:10:50,08 Just remember that. 248 00:10:50,08 --> 00:10:54,07 Rest of it, we trained a beautiful model and we are ready. 249 00:10:54,07 --> 00:10:57,00 So let me close this window. 250 00:10:57,00 --> 00:10:59,01 Next, I'm going to give you a challenge 251 00:10:59,01 --> 00:11:02,04 and you're going to be able to do this for yourself. 252 00:11:02,04 --> 00:11:04,04 There's also a handout with the same thing. 253 00:11:04,04 --> 00:11:06,03 So with the same steps, 254 00:11:06,03 --> 00:11:09,06 so you can actually go learn the steps to do this 255 00:11:09,06 --> 00:11:11,02 and practice more. 256 00:11:11,02 --> 00:11:14,00 And I'm more excited to see 257 00:11:14,00 --> 00:11:16,05 what kind of things you're going to do at work. 258 00:11:16,05 --> 00:11:18,04 So at the end of this course, 259 00:11:18,04 --> 00:11:20,04 like when you scroll down at the bottom of the course, 260 00:11:20,04 --> 00:11:23,06 when you finish, you will see an option saying Sudha Live 261 00:11:23,06 --> 00:11:27,01 because I do a live stream on LinkedIn Live 262 00:11:27,01 --> 00:11:29,05 once a month and talk to my students. 263 00:11:29,05 --> 00:11:30,09 So you can come there, 264 00:11:30,09 --> 00:11:32,09 share the different challenges you face, 265 00:11:32,09 --> 00:11:34,01 the solutions you've built, 266 00:11:34,01 --> 00:11:36,09 ask questions or just continue your learning. 267 00:11:36,09 --> 00:11:39,09 And you can also join my Business School of AI 268 00:11:39,09 --> 00:11:41,00 learner community 269 00:11:41,00 --> 00:11:43,02 where you can just come and tag or ask questions 270 00:11:43,02 --> 00:11:45,05 or the easiest thing, you can even post a question 271 00:11:45,05 --> 00:11:49,00 in the question tab of LinkedIn Learning app.