WEBVTT 00:00:00.000 --> 00:00:02.100 You know that feeling when you're typing a frantic 00:00:02.100 --> 00:00:05.080 text, maybe you're walking, and you mistype a 00:00:05.080 --> 00:00:08.660 word, just completely butcher it, but autocorrect 00:00:08.660 --> 00:00:11.380 just silently fixes it? Oh yeah. You hit send, 00:00:11.519 --> 00:00:14.019 the message looks perfect, the other person has 00:00:14.019 --> 00:00:16.519 no idea you originally typed, you know, ducking. 00:00:16.600 --> 00:00:19.179 Right, the classic save. The meaning got there, 00:00:19.500 --> 00:00:23.239 but the input was a total disaster. Exactly. 00:00:23.320 --> 00:00:25.699 But here's the thing, if you were speaking that 00:00:25.699 --> 00:00:27.559 message, say, in a language you're trying to 00:00:27.559 --> 00:00:30.250 learn, you would have sounded completely wrong. 00:00:30.829 --> 00:00:32.350 The meaning might have gotten there, but the 00:00:32.350 --> 00:00:35.149 delivery, broken. And for the longest time, that 00:00:35.149 --> 00:00:38.549 has been the hidden trap of using AI to learn 00:00:38.549 --> 00:00:41.229 a language. It's the auto -correct trap. You 00:00:41.229 --> 00:00:43.590 speak into an app, it checks your grammar, maybe 00:00:43.590 --> 00:00:46.030 fixes your syntax, but it's completely deaf to 00:00:46.030 --> 00:00:48.630 how you actually sound. Right. It fixes the text, 00:00:48.729 --> 00:00:51.250 not the voice. Which is fine for, you know, writing 00:00:51.250 --> 00:00:54.229 an email. But it's catastrophic if you're trying 00:00:54.229 --> 00:00:56.130 to learn how to speak. You think you're practicing 00:00:56.130 --> 00:00:58.609 pronunciation, but really, you're just practicing 00:00:58.609 --> 00:01:03.009 dictation. But today, we are looking at a massive 00:01:03.009 --> 00:01:05.170 shift in the technology underneath all this. 00:01:05.609 --> 00:01:08.750 We're moving from an AI that reads to an AI that 00:01:08.750 --> 00:01:11.290 actually listens. We have a stack of sources 00:01:11.290 --> 00:01:14.790 here about using Google Gemini Pro specifically 00:01:14.790 --> 00:01:17.670 for pronunciation training. And I got to say. 00:01:18.489 --> 00:01:22.870 Looking at what it can do, some of it is borderline 00:01:22.870 --> 00:01:25.349 spooky. Spooky is the right word. I mean, we 00:01:25.349 --> 00:01:27.650 are not talking about robot voices anymore. We're 00:01:27.650 --> 00:01:30.310 talking about a tool that can tell if you sound 00:01:30.310 --> 00:01:32.730 confident. Or nervous. Or if you're just rushing 00:01:32.730 --> 00:01:34.109 because you want to get the sentence over with. 00:01:34.170 --> 00:01:36.189 See, that's the part that hooked me. Because 00:01:36.189 --> 00:01:38.129 I have tried the language apps, I've done the 00:01:38.129 --> 00:01:41.170 owl, and that frustration of, am I saying this 00:01:41.170 --> 00:01:44.319 right? Yeah. The number one reason I quit. It 00:01:44.319 --> 00:01:46.640 is. It's that total lack of real feedback. It's 00:01:46.640 --> 00:01:48.680 the biggest hurdle. And what we have here is 00:01:48.680 --> 00:01:50.780 basically a roadmap. We're going to cover why 00:01:50.780 --> 00:01:53.500 Gemini Pro is fundamentally different from tools 00:01:53.500 --> 00:01:56.299 like ChatGPT or Claude when it comes to audio. 00:01:56.359 --> 00:01:58.459 The physics of it. The physics of it, exactly. 00:01:58.480 --> 00:02:00.680 Yeah. Then we're going to break down this power 00:02:00.680 --> 00:02:03.760 prompt recipe that supposedly turns the AI into 00:02:03.760 --> 00:02:06.819 a 20 -year veteran coach. And then look at a 00:02:06.819 --> 00:02:10.000 daily routine to actually fix the problems it 00:02:10.000 --> 00:02:12.599 finds. Yep. So let's start with the technology 00:02:12.599 --> 00:02:15.360 itself, because I think most people, and I include 00:02:15.360 --> 00:02:17.639 myself here, we just assume AI is AI. If I talk 00:02:17.639 --> 00:02:20.360 to ChatGPT or I talk to Gemini, it's all just 00:02:20.360 --> 00:02:23.319 processing data. Not quite. But the source material 00:02:23.319 --> 00:02:26.979 draws a really hard line in the sand here, especially 00:02:26.979 --> 00:02:29.639 with how they handle sound. It's a massive distinction. 00:02:30.120 --> 00:02:32.539 And it all comes down to how the machine perceives 00:02:32.539 --> 00:02:34.879 reality. So most of the tools we've used for 00:02:34.879 --> 00:02:37.699 the last few years, ChatGPT, the voice assistants 00:02:37.699 --> 00:02:41.360 on your phone, They all rely on this legacy system 00:02:41.360 --> 00:02:46.039 called speech to text. STT. STT. Yeah. STT. I 00:02:46.039 --> 00:02:47.860 know the acronym, but let's slow down a bit. 00:02:48.060 --> 00:02:50.319 How does that actually work under the hood? OK. 00:02:50.319 --> 00:02:52.419 So imagine you're speaking to a stenographer 00:02:52.419 --> 00:02:54.939 who is trying to be a little too helpful. You 00:02:54.939 --> 00:02:58.360 say a sentence into the mic. The AI's first job 00:02:58.360 --> 00:03:01.039 isn't to critique how you sound. Its first job 00:03:01.039 --> 00:03:03.020 is to figure out what words you intended to say. 00:03:03.039 --> 00:03:05.580 Right. It takes the audio, strips out the noise, 00:03:05.759 --> 00:03:07.819 turns it into text tokens, and then it analyzes 00:03:07.819 --> 00:03:10.460 the text. So it's discarding the audio almost 00:03:10.460 --> 00:03:13.139 immediately. Precisely. Once it has the text, 00:03:13.300 --> 00:03:17.139 the audio is trash. It's gone. Wow. So if you 00:03:17.139 --> 00:03:19.360 mispronounce a word, but the context makes it 00:03:19.360 --> 00:03:21.280 obvious what you meant. The AI just fixes it. 00:03:21.460 --> 00:03:23.259 It acts like that helpful friend who finishes 00:03:23.259 --> 00:03:26.340 your sentences for you. If you say, I would like 00:03:26.340 --> 00:03:29.840 an apple, and the context is fruit, the speech 00:03:29.840 --> 00:03:33.500 detect system just writes apple. It hands the 00:03:33.500 --> 00:03:36.729 text apple to the brain of the AI. The AI looks 00:03:36.729 --> 00:03:39.270 at it and says, grammar is perfect. Meanwhile, 00:03:39.289 --> 00:03:41.349 you're still walking around saying Opal. Exactly. 00:03:41.949 --> 00:03:44.110 I mean, that's fatally flawed for a learner. 00:03:44.590 --> 00:03:47.169 It's prioritizing meaning over the mechanics. 00:03:47.370 --> 00:03:49.689 It's optimizing for communication, not for correction. 00:03:50.430 --> 00:03:52.729 And there's another layer to this that's mentioned 00:03:52.729 --> 00:03:55.909 in the sources, the issue of training bias. The 00:03:55.909 --> 00:03:59.270 profiling aspect. In a way, yeah. These STT systems 00:03:59.270 --> 00:04:01.990 are trained on these massive data sets, so they 00:04:01.990 --> 00:04:04.409 use demographic probabilities to make guesses. 00:04:04.569 --> 00:04:06.870 So if it knows I'm from a certain country. Or 00:04:06.870 --> 00:04:09.990 if you tell it, say you're Vietnamese, it accesses 00:04:09.990 --> 00:04:13.050 a database of common errors made by Vietnamese 00:04:13.050 --> 00:04:15.409 speakers. So it's already looking for missing 00:04:15.409 --> 00:04:17.970 ending sounds before I even open my mouth. It's 00:04:17.970 --> 00:04:20.779 a confirmation bias engine. It might flag a missing 00:04:20.779 --> 00:04:23.120 ending sound because statistically, that's what 00:04:23.120 --> 00:04:25.600 it expects to hear, even if you actually nailed 00:04:25.600 --> 00:04:28.199 it. So it's giving me generic advice based on 00:04:28.199 --> 00:04:30.879 my profile, not personal feedback. Right, not 00:04:30.879 --> 00:04:32.899 based on your actual performance. That explains 00:04:32.899 --> 00:04:35.680 so much about why those generic language apps 00:04:35.680 --> 00:04:39.560 feel so repetitive. Okay, so how is Gemini Pro 00:04:39.560 --> 00:04:42.800 different? The source keeps using this word, 00:04:43.259 --> 00:04:46.500 multimodal. Multimodal is the game changer. It 00:04:46.500 --> 00:04:49.019 means the AI isn't just looking at a text transcript, 00:04:49.220 --> 00:04:52.860 it is processing the raw audio file. The actual 00:04:52.860 --> 00:04:55.639 sound waves. The actual waveforms, yeah. It's 00:04:55.639 --> 00:04:57.819 listening to the length of your vowels in milliseconds. 00:04:58.600 --> 00:05:00.459 It's hearing where you place the stress in a 00:05:00.459 --> 00:05:03.199 word. It's detecting the micropauses between 00:05:03.199 --> 00:05:05.740 syllables. So it's listening to the physics of 00:05:05.740 --> 00:05:08.100 the sound, not just the definition of the word. 00:05:08.259 --> 00:05:10.180 Right. It connects the audio directly to the 00:05:10.180 --> 00:05:12.920 processing core. It can feel the energy. One 00:05:12.920 --> 00:05:15.379 of the key insights here is that Gemini can tell 00:05:15.379 --> 00:05:18.180 if you're speaking confidently or shyly. A text 00:05:18.180 --> 00:05:20.639 transcript can't show shyness. But sound waves 00:05:20.639 --> 00:05:23.620 can. That is the aha moment for me. It's the 00:05:23.620 --> 00:05:26.000 difference between sending someone an email and 00:05:26.000 --> 00:05:28.540 leaving them a voicemail. Yeah. The emotional 00:05:28.540 --> 00:05:32.189 context is just there. And for pronunciation, 00:05:32.689 --> 00:05:34.689 that emotional contest, the rhythm, the hesitation, 00:05:34.889 --> 00:05:37.129 the prosody, that is where the accent lives. 00:05:37.209 --> 00:05:40.149 That's where fluency lives. So the fatal flaw 00:05:40.149 --> 00:05:42.649 of standard speech to text is that it cleans 00:05:42.649 --> 00:05:45.910 up the mess before analyzing it. Exactly. It 00:05:45.910 --> 00:05:48.509 prioritizes meaning over sound, fixing your errors 00:05:48.509 --> 00:05:51.029 instead of flagging them. OK. So we have a tool 00:05:51.029 --> 00:05:53.930 that has ears, metaphorically speaking, but the 00:05:53.930 --> 00:05:56.529 source material is very clear. You can't just 00:05:56.529 --> 00:05:59.189 turn it on and say, help me. No. that just leads 00:05:59.189 --> 00:06:01.889 to chaos, you need a specific setup. Right. You 00:06:01.889 --> 00:06:04.129 have to control the variables. And this is practically 00:06:04.129 --> 00:06:06.529 very simple, but it's crucial. First, the hardware. 00:06:06.649 --> 00:06:08.829 OK. You don't need a studio mic. Your smartphone 00:06:08.829 --> 00:06:12.589 is fine. But the room. Yeah. The room matters. 00:06:12.689 --> 00:06:15.709 Quiet room. Door closed. Essential. Because remember, 00:06:15.970 --> 00:06:17.790 we're dealing with sound waves now, not just 00:06:17.790 --> 00:06:20.550 text tokens. If you have a fan whirring in the 00:06:20.550 --> 00:06:23.709 background, traffic noise, Gemini might interpret 00:06:23.709 --> 00:06:26.170 that white noise as a phoneme. It might think 00:06:26.170 --> 00:06:28.420 that whoosh. from the fan is you trying to make 00:06:28.420 --> 00:06:30.560 a nice sound. So it's too sensitive for its own 00:06:30.560 --> 00:06:32.720 good sometimes. It can create hallucinations 00:06:32.720 --> 00:06:36.160 in the audio processing. So silence is key. But 00:06:36.160 --> 00:06:38.860 the bigger setup rule, and this is where most 00:06:38.860 --> 00:06:41.579 people fail, is about what you actually say. 00:06:41.959 --> 00:06:44.420 The no single words rule. This is a mistake everyone 00:06:44.420 --> 00:06:45.939 makes. They pick up the app and they just say, 00:06:45.939 --> 00:06:49.319 Apple, banana. Hello. I do this constantly. I 00:06:49.319 --> 00:06:50.819 just want to check if I can say the word. Why 00:06:50.819 --> 00:06:52.939 is that bad? Because that's not how language 00:06:52.939 --> 00:06:57.129 works. The force suggests a paragraph like 150 00:06:57.129 --> 00:06:59.829 to 200 words, a short story, a recap of your 00:06:59.829 --> 00:07:03.069 day, whatever. The reason is a technical concept 00:07:03.069 --> 00:07:06.209 called co -articulation. Co -articulation? It's 00:07:06.209 --> 00:07:08.870 how sounds influence each other. When you speak 00:07:08.870 --> 00:07:11.670 a single word, you say it in isolation. But in 00:07:11.670 --> 00:07:14.589 a real sentence, the end of one word blends into 00:07:14.589 --> 00:07:16.970 the start of the next one. The rhythm changes. 00:07:17.069 --> 00:07:18.769 If you only practice single words, you sound 00:07:18.769 --> 00:07:21.649 like a robot. The AI needs to hear how you link 00:07:21.649 --> 00:07:24.120 words together. Do you blend? Is your rhythm 00:07:24.120 --> 00:07:26.660 robotic or fluid? You only get that data from 00:07:26.660 --> 00:07:28.899 a full paragraph. That's like trying to judge 00:07:28.899 --> 00:07:31.579 a dancer by looking at a single photo versus 00:07:31.579 --> 00:07:34.279 watching a video of them actually moving. That 00:07:34.279 --> 00:07:36.600 is a perfect analogy. The paragraph provides 00:07:36.600 --> 00:07:38.779 the movement. OK, so we're in a quiet room. We 00:07:38.779 --> 00:07:40.819 have our paragraph about our day. Now we get 00:07:40.819 --> 00:07:43.180 to the power prompt. This is the recipe mentioned 00:07:43.180 --> 00:07:44.600 in the source material. And I have to admit, 00:07:44.720 --> 00:07:46.740 whenever I see these long detailed prompts, I 00:07:46.740 --> 00:07:50.040 get a little skeptical. Skeptical. Well, it feels 00:07:50.040 --> 00:07:53.120 like I'm, you know, LRPing with a computer, please 00:07:53.120 --> 00:07:55.740 pretend you are a coach. It just feels weird 00:07:55.740 --> 00:07:58.060 to give a machine a personality. Does it actually 00:07:58.060 --> 00:08:00.300 change the output? It changes everything. You 00:08:00.300 --> 00:08:02.439 have to remember these large language models 00:08:02.439 --> 00:08:06.519 are these vast, generic oceans of text. They've 00:08:06.519 --> 00:08:09.139 read everything from Reddit threads to astrophysics 00:08:09.139 --> 00:08:11.579 textbooks. Right. If you don't narrow their focus, 00:08:11.980 --> 00:08:15.040 they drift. They just refer to the mean. By assigning 00:08:15.040 --> 00:08:17.720 a persona like... a native English pronunciation 00:08:17.720 --> 00:08:20.879 coach with 20 years of experience, you are telling 00:08:20.879 --> 00:08:25.120 the AI which part of its latent space to activate. 00:08:25.699 --> 00:08:28.139 You're forcing it to adopt a specific standard 00:08:28.139 --> 00:08:30.459 of critique. So it's not just flavor text, it's 00:08:30.459 --> 00:08:33.500 a functional constraint. Absolutely. And the 00:08:33.500 --> 00:08:36.820 prompt in the source has four very strict rules 00:08:36.820 --> 00:08:38.779 that I think are brilliant. Okay, let's run through 00:08:38.779 --> 00:08:40.419 them because this seems to be the secret sauce, 00:08:40.539 --> 00:08:42.960 the first one and one we've touched on. No stereotypes. 00:08:43.289 --> 00:08:46.049 This is critical for accuracy. You literally 00:08:46.049 --> 00:08:49.230 command the AI. Only report what is actually 00:08:49.230 --> 00:08:52.830 heard. Do not use demographic data. You are forcing 00:08:52.830 --> 00:08:55.629 it to ignore its training on general population 00:08:55.629 --> 00:08:58.929 stats and focus purely on your audio file. It 00:08:58.929 --> 00:09:01.129 prevents that confirmation bias we talked about. 00:09:01.389 --> 00:09:03.370 Exactly. That's empowering. It's telling the 00:09:03.370 --> 00:09:06.350 machine to look at the data, not the trend. Exactly. 00:09:06.669 --> 00:09:10.009 The second rule is specific quotes. The AI has 00:09:10.009 --> 00:09:13.269 to provide timestamps and the exact words. Right. 00:09:13.570 --> 00:09:15.690 It can't just say, work on your vowels. It has 00:09:15.690 --> 00:09:18.649 to say, at point one five, you said sheep, but 00:09:18.649 --> 00:09:21.320 it sounded more like ship. Which is so helpful, 00:09:21.600 --> 00:09:23.000 because otherwise you're just guessing where 00:09:23.000 --> 00:09:24.779 you went wrong. And frankly, without timestamps, 00:09:24.919 --> 00:09:26.840 I'd probably just assume the AI was hallucinating. 00:09:27.100 --> 00:09:29.639 Then there's stress analysis. English is a stress 00:09:29.639 --> 00:09:32.220 -timed language. If you say phototorlography 00:09:32.220 --> 00:09:34.340 instead of photogorheriaphy. And the meaning 00:09:34.340 --> 00:09:37.419 gets lost or just sounds wrong. Exactly. So the 00:09:37.419 --> 00:09:40.519 prompt explicitly asks the AI to listen for that 00:09:40.519 --> 00:09:43.580 emphasis. And the last one was confusing sounds. 00:09:44.679 --> 00:09:46.259 But I want to go back to the stress analysis 00:09:46.259 --> 00:09:48.379 for a second, because it leads into the most 00:09:48.379 --> 00:09:49.940 surprising part of the source material for me, 00:09:50.019 --> 00:09:52.460 the case study. The Vietnamese student test. 00:09:52.620 --> 00:09:56.659 Yes. This really highlights the emotional intelligence 00:09:56.659 --> 00:10:00.159 of this multimodal system. So the author used 00:10:00.159 --> 00:10:03.480 this prompt with a student. They didn't tell 00:10:03.480 --> 00:10:05.539 Gemini where the student was from. But Gemini 00:10:05.539 --> 00:10:08.789 guessed it. Instantly. Based purely on the intonation, 00:10:09.090 --> 00:10:10.929 the rise and fall of the voice gemini correctly 00:10:10.929 --> 00:10:13.029 identified the speaker as likely Vietnamese. 00:10:13.429 --> 00:10:15.950 Whoa! That is nuanced listening. That's detecting 00:10:15.950 --> 00:10:17.830 the music of the mother tongue bleeding into 00:10:17.830 --> 00:10:20.429 the target language. That's wild. But what stood 00:10:20.429 --> 00:10:23.110 out to me even more in that case study was the 00:10:23.110 --> 00:10:26.289 feedback on speed. It wasn't just, you said this 00:10:26.289 --> 00:10:28.529 word wrong. Oh, the nervous comment, yeah. Yeah, 00:10:28.570 --> 00:10:30.970 the AI told the student, you are speaking too 00:10:30.970 --> 00:10:33.629 fast. This makes you sound nervous. And it suggested 00:10:33.629 --> 00:10:36.269 pausing at commas to sound more confident. I 00:10:36.269 --> 00:10:37.950 mean, just think about the implication of that. 00:10:38.070 --> 00:10:40.889 That isn't pronunciation advice. That is psychological 00:10:40.889 --> 00:10:43.450 coaching. It is. It's soft skills. The standard 00:10:43.450 --> 00:10:45.909 speech -to -text tool would just process the 00:10:45.909 --> 00:10:48.330 words. If the words were right, it would give 00:10:48.330 --> 00:10:52.289 a thumbs up. Gemini heard the pace, the milliseconds 00:10:52.289 --> 00:10:55.309 between words, and correlated it with an emotional 00:10:55.309 --> 00:10:58.360 state, anxiety. It's teaching you how to command 00:10:58.360 --> 00:11:00.919 a room, not just how to pronounce a vowel. It's 00:11:00.919 --> 00:11:03.320 connecting the dots between how we sound and 00:11:03.320 --> 00:11:05.480 how we are perceived, which is, you know, the 00:11:05.480 --> 00:11:08.120 whole point of learning a language, really. We 00:11:08.120 --> 00:11:10.120 want to be understood, but we also want to project 00:11:10.120 --> 00:11:12.860 who we are. Precisely. It's closing the gap between 00:11:12.860 --> 00:11:15.580 your internal voice and your external reality. 00:11:15.840 --> 00:11:18.840 I'm curious, why is the no -stereotypes instruction 00:11:18.840 --> 00:11:22.500 so critical for the user? It forces the AI to 00:11:22.500 --> 00:11:24.740 listen to your individual voice, not just guess, 00:11:25.019 --> 00:11:27.240 based on a demographic textbook. Now, before 00:11:27.240 --> 00:11:28.820 we get too carried away thinking this is all 00:11:28.820 --> 00:11:31.299 magic, we have to pay the bills. Okay, we are 00:11:31.299 --> 00:11:33.519 back. We've praised the ghost in the machine, 00:11:33.840 --> 00:11:38.000 but the source material also throws a, uh, a 00:11:38.000 --> 00:11:41.080 bucket of cold water on us. Gemini Pro is good, 00:11:41.080 --> 00:11:43.899 but it has limits. It does. It's not a magic 00:11:43.899 --> 00:11:46.350 wand. It's not a human. That's so important to 00:11:46.350 --> 00:11:48.580 remember. We talked about the Hallucinations 00:11:48.580 --> 00:11:51.139 from background noise, that's a big one. But 00:11:51.139 --> 00:11:53.779 there's also the issue of accent confusion. Right, 00:11:53.919 --> 00:11:56.980 the water versus wata problem. Exactly. If you 00:11:56.980 --> 00:11:59.340 don't specify in your prompt whether you're aiming 00:11:59.340 --> 00:12:02.340 for a general American or receive pronunciation, 00:12:02.600 --> 00:12:05.720 that's the standard British accent. Gemini just 00:12:05.720 --> 00:12:08.279 defaults to the average. Which is usually a generic 00:12:08.279 --> 00:12:10.500 American accent. Usually, yeah. So if I'm trying 00:12:10.500 --> 00:12:12.659 to sound like I'm from London and I say schedule 00:12:12.659 --> 00:12:15.480 the British way. Gemini might flag it as an error. 00:12:15.600 --> 00:12:17.809 It might, yeah. because it's comparing you to 00:12:17.809 --> 00:12:21.250 a database of mostly American speakers. You have 00:12:21.250 --> 00:12:23.149 to be specific in your prompt. You got to say 00:12:23.149 --> 00:12:26.210 act as a British English coach. That seems like 00:12:26.210 --> 00:12:29.309 a simple fix, but definitely good to know. What 00:12:29.309 --> 00:12:32.909 about the hardware limitation? The source mentioned 00:12:32.909 --> 00:12:35.580 something about. sounds getting lost. This is 00:12:35.580 --> 00:12:37.559 a physics problem and it's actually really interesting. 00:12:37.980 --> 00:12:40.159 So the microphone on your smartphone is incredible, 00:12:40.539 --> 00:12:43.159 but it's designed for phone calls. Okay. So it 00:12:43.159 --> 00:12:45.139 often runs these noise cancellation algorithms 00:12:45.139 --> 00:12:48.139 to cut out background hiss. Right. The problem 00:12:48.139 --> 00:12:51.620 is, some English sounds, specifically the unvoiced 00:12:51.620 --> 00:12:54.759 flickatives like the the and them or the F and 00:12:54.759 --> 00:12:58.000 fish, they occupy the same high frequency range 00:12:58.000 --> 00:13:00.679 as that background hiss. So the phone thinks 00:13:00.679 --> 00:13:03.200 my the sound is just air conditioner noise and 00:13:03.200 --> 00:13:05.679 it deletes it. Exactly. It scrubs the audio to 00:13:05.679 --> 00:13:08.360 clean it. And in the process, it removes the 00:13:08.360 --> 00:13:10.799 very sound you were trying to practice. Gemini 00:13:10.799 --> 00:13:13.299 might say, you missed the the sound when actually 00:13:13.299 --> 00:13:14.899 you said it right and your phone just filtered 00:13:14.899 --> 00:13:16.899 it out. So don't get gaslighted by your hardware. 00:13:17.179 --> 00:13:19.779 Yes. Trust your ears or ask a human if you're 00:13:19.779 --> 00:13:22.720 really stuck. Don't let the AI destroy your confidence 00:13:22.720 --> 00:13:25.600 over a noise cancellation algorithm. So we have 00:13:25.600 --> 00:13:27.840 the tool, we have the prompt, we know the limits. 00:13:28.259 --> 00:13:30.700 The source material ends with a study plan, this 00:13:30.700 --> 00:13:33.600 record analyze fix routine. This is the practical 00:13:33.600 --> 00:13:36.379 application part. The author suggests a daily 00:13:36.379 --> 00:13:39.450 routine. And the key here is repetition. And 00:13:39.450 --> 00:13:41.509 this is a bit controversial. Don't record a new 00:13:41.509 --> 00:13:43.549 paragraph every day. Wait, really? I would have 00:13:43.549 --> 00:13:45.710 thought variety is better. I need to learn more 00:13:45.710 --> 00:13:48.110 words. No, because you can't track your progress 00:13:48.110 --> 00:13:51.110 if the target is always moving. The advice is 00:13:51.110 --> 00:13:53.330 to record the same paragraph every single day 00:13:53.330 --> 00:13:55.610 for a week. Ah, so you can A -test yourself. 00:13:55.929 --> 00:13:58.450 Exactly. You record on Monday, you get the feedback, 00:13:58.629 --> 00:14:00.850 you record on Tuesday, and you upload both files 00:14:00.850 --> 00:14:03.549 to Gemini. Then you ask, compared to my recording 00:14:03.549 --> 00:14:05.940 yesterday, What is better? You're building a 00:14:05.940 --> 00:14:08.519 feedback loop. A feedback loop, yeah. You master 00:14:08.519 --> 00:14:11.080 the rhythm of that specific paragraph. It's like 00:14:11.080 --> 00:14:13.100 learning a song on the piano. You play the same 00:14:13.100 --> 00:14:15.320 piece until it flows, then you move on. That 00:14:15.320 --> 00:14:17.779 makes so much sense. It's about depth, not breadth. 00:14:17.980 --> 00:14:20.480 But the author also suggests using other apps, 00:14:20.659 --> 00:14:23.710 the hybrid approach. Yes. This is acknowledging 00:14:23.710 --> 00:14:26.029 that Gemini is a big picture coach. It's great 00:14:26.029 --> 00:14:29.149 for flow, rhythm, intonation, the macro stuff. 00:14:29.730 --> 00:14:32.710 But for the nitty -gritty, like drilling a specific 00:14:32.710 --> 00:14:35.809 vowel sound over and over, it's not the most 00:14:35.809 --> 00:14:38.070 efficient tool. So you use apps like BoldVoice 00:14:38.070 --> 00:14:41.850 or Speechling for the drills? Right. Use Speechling 00:14:41.850 --> 00:14:45.029 to drill the word squirrel 50 times until your 00:14:45.029 --> 00:14:48.429 tongue stops tying itself in knots. Then you 00:14:48.429 --> 00:14:51.480 go to Gemini. and put squirrel into a full sentence 00:14:51.480 --> 00:14:53.580 to see if you can maintain that pronunciation 00:14:53.580 --> 00:14:56.279 while speaking naturally. I like that. Specialized 00:14:56.279 --> 00:14:59.700 tools for the bricks. Gemini for the house. That's 00:14:59.700 --> 00:15:01.279 a great way to put it. And the final piece of 00:15:01.279 --> 00:15:03.740 advice in the source was about input. You know, 00:15:04.059 --> 00:15:06.879 put correct sounds into your head. You can't 00:15:06.879 --> 00:15:09.419 output what you have in input. You need to listen 00:15:09.419 --> 00:15:12.509 to experts. The source mentions channels like 00:15:12.509 --> 00:15:15.250 Luke Pretty or Cloud English. You need to fill 00:15:15.250 --> 00:15:17.490 your brain with the target rhythm so that when 00:15:17.490 --> 00:15:19.549 you record, you actually have a reference point. 00:15:19.690 --> 00:15:21.929 It seems so obvious, but I think a lot of us 00:15:21.929 --> 00:15:25.169 skip that. We just start talking. We do. We want 00:15:25.169 --> 00:15:27.029 to perform before we've rehearsed. But listening 00:15:27.029 --> 00:15:30.169 is 50 % of speaking. So how do we get around 00:15:30.169 --> 00:15:33.529 the hardware erasing those small sounds? We rely 00:15:33.529 --> 00:15:36.409 on specialized apps for the details and use Gemini 00:15:36.409 --> 00:15:38.409 for the big picture flow. So let's bring this 00:15:38.409 --> 00:15:40.509 all together. What is the big takeaway for you 00:15:40.509 --> 00:15:42.429 here? Because for me, it's that we are finally 00:15:42.429 --> 00:15:45.049 moving away from passive learning. That is the 00:15:45.049 --> 00:15:48.169 core of it. For years, technology has kind of 00:15:48.169 --> 00:15:51.169 made us lazy learners. It auto -translates. It 00:15:51.169 --> 00:15:54.700 auto -corrects. This approach... using Gemini 00:15:54.700 --> 00:15:58.620 as a coach. It forces us to be active. We have 00:15:58.620 --> 00:16:00.519 to speak, we have to listen to the feedback, 00:16:00.679 --> 00:16:02.860 and we have to actually adjust. It's the difference 00:16:02.860 --> 00:16:05.700 between using a crutch and doing physical therapy. 00:16:05.799 --> 00:16:08.399 Yes. And the technology is finally ready to meet 00:16:08.399 --> 00:16:10.259 us there. It's not just matching text anymore. 00:16:10.679 --> 00:16:13.460 It's listening to the human element, the confidence, 00:16:13.639 --> 00:16:16.440 the hesitation, the rhythm. It turns the AI from 00:16:16.440 --> 00:16:19.240 a spell checker into a mentor. And a strict mentor, 00:16:19.440 --> 00:16:22.000 if you use the right prompt. A very strict mentor. 00:16:22.080 --> 00:16:25.100 No stereotypes, strict quotes. So... Here's our 00:16:25.100 --> 00:16:27.440 challenge to you, the listener. You have a phone. 00:16:27.500 --> 00:16:29.120 You probably have a Google account. It's free 00:16:29.120 --> 00:16:31.240 to try. Right now, when this deep dive ends, 00:16:31.320 --> 00:16:33.779 don't just think that's cool. Pick up your phone. 00:16:34.279 --> 00:16:36.299 Find a quiet closet if you have to. Maybe just 00:16:36.299 --> 00:16:38.919 a quiet room. Closets have bad acoustics. Fair 00:16:38.919 --> 00:16:41.340 point. Record 30 seconds about your day. Just 00:16:41.340 --> 00:16:44.299 30 seconds. Paste in that power prompt. Tell 00:16:44.299 --> 00:16:48.059 it to be a 20 -year veteran coach. And just see 00:16:48.059 --> 00:16:51.639 what it says. And specifically, look at the emotional 00:16:51.639 --> 00:16:54.919 feedback. Does it say you sound nervous? Does 00:16:54.919 --> 00:16:57.879 it say you sound robotic? That might be the most 00:16:57.879 --> 00:17:00.059 valuable thing you learned today. It's not about 00:17:00.059 --> 00:17:02.860 sounding like a native speaker perfectly. No. 00:17:03.200 --> 00:17:05.359 It's about being comfortable and being understood. 00:17:06.420 --> 00:17:08.440 That's it for this deep dive. Thanks for listening 00:17:08.440 --> 00:17:10.619 and good luck with the recording. Let us know 00:17:10.619 --> 00:17:12.660 what the AI hears. See you next time.