WEBVTT 00:00.480 --> 00:03.450 -: All right, this is how to use the Whisper API 00:03.450 --> 00:06.270 to transcribe a YouTube video. 00:06.270 --> 00:09.177 We're going to just run this. 00:09.177 --> 00:12.930 I'm just gonna hit command shift 00:12.930 --> 00:14.850 or control shift if you're on a PC. 00:14.850 --> 00:17.283 And this is gonna install OpenAI. 00:18.180 --> 00:19.860 That's one of the things we need. 00:19.860 --> 00:22.770 We're also going to run this as well. 00:22.770 --> 00:24.810 This is YouTube DLP. 00:24.810 --> 00:26.880 This is like YouTube DL, 00:26.880 --> 00:29.010 if you guys have used that before. 00:29.010 --> 00:32.160 But I found that it's more up to date, more frequently, 00:32.160 --> 00:35.850 and YouTube occasionally changes the way that they stream 00:35.850 --> 00:37.620 and that breaks YouTube DL. 00:37.620 --> 00:41.310 I found this works, but you can also try YouTube DL, 00:41.310 --> 00:43.470 works in a very similar way. 00:43.470 --> 00:45.300 So I'm just gonna download this YouTube video. 00:45.300 --> 00:46.740 This one is mine. 00:46.740 --> 00:47.970 I own the copyright for this. 00:47.970 --> 00:52.710 So I would say if you don't own the copyright for your video 00:52.710 --> 00:54.150 that you're downloading, just be careful, 00:54.150 --> 00:57.270 because you might be restricted in what you can do. 00:57.270 --> 01:00.870 But it depends on the use case using the transcripts for, 01:00.870 --> 01:03.450 in our case, what we're doing is 01:03.450 --> 01:05.820 we're taking the YouTube video that I have 01:05.820 --> 01:07.830 and we're just creating a transcript. 01:07.830 --> 01:10.230 And we can use that transcript to post 01:10.230 --> 01:12.990 and promote the YouTube video itself. 01:12.990 --> 01:14.460 So while this is downloading, 01:14.460 --> 01:16.110 this is basically just streaming 01:16.110 --> 01:18.210 and then capturing the fragments. 01:18.210 --> 01:22.290 I'm just gonna click here to go to OpenAI. 01:22.290 --> 01:24.420 We need to get an API key. 01:24.420 --> 01:26.820 So we're gonna create a new secret key. 01:26.820 --> 01:29.730 Just gonna call this whisper. 01:29.730 --> 01:31.950 I'm gonna delete this afterwards, 01:31.950 --> 01:35.220 because you shouldn't share your API keys. 01:35.220 --> 01:36.053 Cool. 01:36.053 --> 01:36.886 So, 01:39.330 --> 01:41.550 it looks like this is downloaded. 01:41.550 --> 01:43.650 And actually if we look in our file system, yeah, 01:43.650 --> 01:45.990 we can see this, the video here. 01:45.990 --> 01:49.710 So now let's set up our API key. 01:49.710 --> 01:53.190 Just gonna run this and then we can just paste it in there. 01:53.190 --> 01:56.400 And that kind of saves it in the open AI library. 01:56.400 --> 01:59.520 And now we need to, so we have this video, 01:59.520 --> 02:01.200 and it's WebM format. 02:01.200 --> 02:03.660 So we're gonna use ffmpeg, 02:03.660 --> 02:08.070 which is a library for basically just changing 02:08.070 --> 02:10.830 or manipulating video audio files. 02:10.830 --> 02:11.663 And we're gonna use this, 02:11.663 --> 02:13.770 we're gonna take in this video, 02:13.770 --> 02:16.200 and then output it as like an MP3. 02:16.200 --> 02:18.603 So you can get this file name. 02:19.770 --> 02:21.130 Just paste this in here 02:22.214 --> 02:23.943 and hopefully that runs. 02:25.690 --> 02:27.930 Okay. Yeah, so that's not running, I think, 02:27.930 --> 02:29.700 because this name is a little bit funny. 02:29.700 --> 02:33.753 So let's just call this input.webm, 02:34.680 --> 02:36.530 and then we'll just change this here. 02:39.042 --> 02:39.959 Input.webm. 02:44.367 --> 02:45.200 All right. 02:46.074 --> 02:49.560 And this is basically just converting the video format 02:49.560 --> 02:52.833 and stripping out just the MP3, like the, just the audio. 02:56.550 --> 02:57.480 All right, so that's done. 02:57.480 --> 02:59.460 And actually if we refresh we can see there's 02:59.460 --> 03:01.500 output.mp3 here. 03:01.500 --> 03:04.590 And now we're going to call the Whisper API. 03:04.590 --> 03:06.540 So this is the easy part. (laughs) 03:06.540 --> 03:08.730 I've asked it for the verbose JSON. 03:08.730 --> 03:12.030 By default, it just gives you back the transcript, 03:12.030 --> 03:13.800 but this will give you all the different 03:13.800 --> 03:15.240 segments of the transcript. 03:15.240 --> 03:17.823 So just gonna run this, you'll see what I mean. 03:21.948 --> 03:24.900 And there is a cost to this that's very cheap, though, 03:24.900 --> 03:28.380 negligible, like pennies, even for relatively long audio. 03:28.380 --> 03:30.330 It shouldn't be too much of an issue. 03:30.330 --> 03:32.580 But keep on top of your cost caps. (laughs) 03:32.580 --> 03:35.580 Make sure that you don't rack up too much 03:35.580 --> 03:37.530 of a fee if you're doing this for lots 03:37.530 --> 03:39.213 of different podcasts or audio. 03:40.620 --> 03:41.733 Wait for this to run. 03:47.458 --> 03:50.541 (instructor sniffs) 04:02.580 --> 04:03.510 Oh, here we go. 04:03.510 --> 04:04.650 Now it's finished. 04:04.650 --> 04:06.570 And look, it's a pretty big file. 04:06.570 --> 04:10.530 So you can see it's broken into these chunks, right? 04:10.530 --> 04:15.450 We have the start time, the end time, the actual text, 04:15.450 --> 04:19.440 and then even the tokens and the log probability. 04:19.440 --> 04:21.240 This is like a lot of information. 04:21.240 --> 04:23.340 We don't need all of this information. 04:23.340 --> 04:25.410 We can actually just get the text itself. 04:25.410 --> 04:27.840 So if we just try this. 04:27.840 --> 04:29.010 Yeah, here we go. 04:29.010 --> 04:30.570 So this is the full text 04:30.570 --> 04:32.250 and I can click there to expand it. 04:32.250 --> 04:33.540 It's the full text of the video. 04:33.540 --> 04:34.860 That's the full transcript, 04:34.860 --> 04:36.570 but this doesn't have any timestamps. 04:36.570 --> 04:39.180 What I wanted to show you is how to kinda 04:39.180 --> 04:41.280 split out the timestamps themselves. 04:41.280 --> 04:43.200 So I've gone through the segments. 04:43.200 --> 04:46.350 So it's four segment in transcript segments, 04:46.350 --> 04:49.110 and then I'm unbundling the start minutes and start seconds. 04:49.110 --> 04:50.520 I'm using this div mod. 04:50.520 --> 04:54.240 It's basically just divides the start number by 60 04:54.240 --> 04:55.890 to get the number of minutes 04:55.890 --> 04:58.650 and then the number of seconds afterwards. 04:58.650 --> 04:59.790 So that'd be how many seconds are 04:59.790 --> 05:02.040 left over after you divided. 05:02.040 --> 05:03.810 And then I'm creating a timestamp. 05:03.810 --> 05:06.120 So if I just run this, 05:06.120 --> 05:08.643 you'll be able to see, here we go. 05:11.160 --> 05:13.470 So the timestamp is formatted, 05:13.470 --> 05:15.960 so there's always like a prevailing zero, 05:15.960 --> 05:17.280 even if the minutes, 05:17.280 --> 05:20.010 even if it's less than one minute, for example. 05:20.010 --> 05:22.290 And same thing with the seconds as well. 05:22.290 --> 05:24.240 Just to make the formatting more consistent. 05:24.240 --> 05:25.073 And there we go. 05:25.073 --> 05:27.120 We have kind of line by line transcript 05:27.120 --> 05:29.223 with the seconds and the minutes. 05:31.500 --> 05:32.880 All right, there's a lot you can do with this. 05:32.880 --> 05:35.520 I think you can do real-time streaming as well. 05:35.520 --> 05:39.200 It is open source as well as available via API. 05:39.200 --> 05:42.240 So you could run this locally if you have this 05:42.240 --> 05:44.310 set up on your server with a GPU. 05:44.310 --> 05:45.990 But I think it's easier 05:45.990 --> 05:48.360 and relatively cheap to use the API. 05:48.360 --> 05:51.360 And now you can transcribe any audio you want. 05:51.360 --> 05:54.120 I would say that one thing I would watch out for is 05:54.120 --> 05:55.440 that it's not particularly good 05:55.440 --> 05:57.180 when you have more than one speaker. 05:57.180 --> 05:58.100 That's what I've found it. 05:58.100 --> 06:03.100 It tends to work best for like tutorials or podcasts, 06:03.390 --> 06:06.303 things where there's just one speaker at a time typically.