WEBVTT 00:00.480 --> 00:02.340 -: Hey, in this video, what we're gonna cover 00:02.340 --> 00:04.710 is how you can use OpenAI's package 00:04.710 --> 00:09.710 to convert text into real sounding voices and audio files. 00:09.840 --> 00:10.770 Now, there's a couple 00:10.770 --> 00:12.240 of different packages you're gonna need for this. 00:12.240 --> 00:14.340 You're gonna need langchain, langchain_openai 00:14.340 --> 00:16.050 and the openai package. 00:16.050 --> 00:18.150 You'll need to update your API key. 00:18.150 --> 00:21.990 And then what we can do is have a look at openai's package. 00:21.990 --> 00:24.420 Now, you can specify the file path 00:24.420 --> 00:26.790 and you'll see I've got this speech.mp3 00:26.790 --> 00:29.580 and you can also choose the model that you're interested in 00:29.580 --> 00:31.590 and there's lots of different voices. 00:31.590 --> 00:34.500 Now, the way this works is you put in your input 00:34.500 --> 00:38.370 and then you would use response.stream_to_file. 00:38.370 --> 00:39.510 And then after that, 00:39.510 --> 00:41.954 then what you'll see is on the left-hand side, 00:41.954 --> 00:45.690 we've got this speech.mp3 file that's been created, 00:45.690 --> 00:47.820 which is a couple of seconds long here. 00:47.820 --> 00:50.760 But what that will do is it will take the textual input 00:50.760 --> 00:52.770 that you're putting in and it'll convert that 00:52.770 --> 00:54.030 into an audio file. 00:54.030 --> 00:56.400 Now, let's take a little bit step further 00:56.400 --> 00:58.890 and figure out what is, you know, maybe a good use case 00:58.890 --> 01:01.380 for this and like how could you integrate this 01:01.380 --> 01:02.610 into a different flow? 01:02.610 --> 01:04.200 And so we're gonna use one 01:04.200 --> 01:06.420 of LangChain's community document loaders 01:06.420 --> 01:08.010 called a WebBaseLoader. 01:08.010 --> 01:10.230 And what you'll be able to do is you'll load a URL 01:10.230 --> 01:12.690 and we'll look at the BBC News article 01:12.690 --> 01:15.990 that recently came out in UK politics. 01:15.990 --> 01:18.150 And after this, what happens here 01:18.150 --> 01:21.150 is we've got this WebBaseLoader that can take in a URL. 01:21.150 --> 01:23.940 So I can take in, for example, this URL here. 01:23.940 --> 01:26.610 And then after that then, we get some data 01:26.610 --> 01:27.840 and we can have a look at this. 01:27.840 --> 01:29.937 And you can see we've got some documents 01:29.937 --> 01:32.490 and we can see how many of these documents that we have. 01:32.490 --> 01:34.530 So we've got one document in here 01:34.530 --> 01:36.960 and if we want to have a look at the page content 01:36.960 --> 01:40.080 of that document, you could just do the page_content 01:40.080 --> 01:41.940 and you can see this is all the information 01:41.940 --> 01:43.050 that we've got here. 01:43.050 --> 01:47.280 Now what I suggest doing is taking all of that data 01:47.280 --> 01:50.490 and then basically looping over all the documents 01:50.490 --> 01:52.740 and replacing any new line characters. 01:52.740 --> 01:53.910 And then you should end up with something 01:53.910 --> 01:54.990 that looks a little bit like this 01:54.990 --> 01:57.540 where you've got checking the character length. 01:57.540 --> 02:00.060 And we can see this is the number of characters 02:00.060 --> 02:02.820 and we could have a look at the webpage text if we wanted. 02:02.820 --> 02:05.280 But basically, you've got a lot of different webpage texts. 02:05.280 --> 02:07.950 Now, interestingly, I don't think we would want to put all 02:07.950 --> 02:10.230 of this webpage text directly in 02:10.230 --> 02:12.000 because you can see we've got other things 02:12.000 --> 02:16.050 like the navigation has somehow managed to make it in here, 02:16.050 --> 02:18.390 and that's not really what we're looking for. 02:18.390 --> 02:20.100 One thing that can be quite good to do, 02:20.100 --> 02:23.880 and especially as if we're over a certain character limit, 02:23.880 --> 02:25.680 we won't be able to convert text 02:25.680 --> 02:27.360 to speech using OpenAI's package. 02:27.360 --> 02:32.360 It does have a character limit of around 4,906 characters, 02:32.490 --> 02:35.610 but I've said it to 490 just to make sure. 02:35.610 --> 02:36.750 Now, I've decided 02:36.750 --> 02:38.910 to use something called the load_summarize_chain 02:38.910 --> 02:41.850 that allows us to take in a bunch of documents. 02:41.850 --> 02:43.110 And once we have those documents, 02:43.110 --> 02:46.710 then what we can do is we can basically map over them 02:46.710 --> 02:47.880 and reduce them all down 02:47.880 --> 02:52.080 and get some sort of final output from a bunch of documents. 02:52.080 --> 02:53.760 And what I've said, I want the prompters 02:53.760 --> 02:55.320 to say the following is a set of documents 02:55.320 --> 02:57.120 from a BBC News article. 02:57.120 --> 02:59.310 Based on this list of docs, please extract the story 02:59.310 --> 03:00.510 and make it sound engaging. 03:00.510 --> 03:02.670 We've set up a ChatPromptTemplate for that. 03:02.670 --> 03:05.220 And we've got our load_summarize_chain 03:05.220 --> 03:08.370 that's taking in the model, the map_reduce type of chain, 03:08.370 --> 03:10.140 and the map_prompt and the combine_prompt. 03:10.140 --> 03:13.200 And at the moment, I'm not really specifying what I want 03:13.200 --> 03:16.860 to happen there every single step as well as I could do, 03:16.860 --> 03:19.470 but I found this prompt seems to work fine for now. 03:19.470 --> 03:23.100 We invoke the chain, passing in the input_documents in here. 03:23.100 --> 03:25.260 You can see I've just made a new LangChain document 03:25.260 --> 03:26.670 with the web_text here. 03:26.670 --> 03:28.200 It's a single document. 03:28.200 --> 03:31.680 Now, what will happen is after this runs, 03:31.680 --> 03:34.950 if the character limit's above 4,090, 03:34.950 --> 03:36.030 we'll get the output_text 03:36.030 --> 03:38.430 and we'll save that back to the web_text. 03:38.430 --> 03:39.900 And the idea behind this 03:39.900 --> 03:41.430 is that what you really wanna be doing 03:41.430 --> 03:44.760 is not just scraping this data from the webpage, 03:44.760 --> 03:47.850 but actually, you want to be filtering it, 03:47.850 --> 03:49.200 maybe having some prompts 03:49.200 --> 03:51.300 that make it into a more interesting story. 03:51.300 --> 03:53.070 And you can see here we've got some information 03:53.070 --> 03:55.410 about the Labour deputy leader. 03:55.410 --> 03:57.090 Angela Rayner has accused the government 03:57.090 --> 03:58.740 of reneging on its promise. 03:58.740 --> 04:01.650 And so we've got a much better story here 04:01.650 --> 04:03.270 than the original webpage context. 04:03.270 --> 04:06.060 So we've really got the LLM to clean up the text 04:06.060 --> 04:08.910 before we're gonna expose that into downstream LLMs 04:08.910 --> 04:10.530 or downstream systems. 04:10.530 --> 04:13.290 And then again, we can just run the speech.mp3 04:13.290 --> 04:14.940 as the file_path. 04:14.940 --> 04:17.790 We're using the client.audio.speech.create, 04:17.790 --> 04:21.450 and then we're passing in a model, the voice 04:21.450 --> 04:24.623 and the input text, which is now the web_text. 04:24.623 --> 04:28.860 And then we're doing the response.stream_to_ file. 04:28.860 --> 04:31.350 Now, I know this looks like it does give a warning. 04:31.350 --> 04:32.640 It has recommended using 04:32.640 --> 04:35.310 this .with_streaming_response method, 04:35.310 --> 04:37.560 but I found that this doesn't actually work this method. 04:37.560 --> 04:39.450 So you can ignore this warning for now. 04:39.450 --> 04:42.300 And then that will save the audio directly 04:42.300 --> 04:44.190 into this speech.mp3 file. 04:44.190 --> 04:45.180 And you can see we've got 04:45.180 --> 04:48.420 about 1 minute and 46 seconds of speech. 04:48.420 --> 04:50.520 And so this can be a really great way for you 04:50.520 --> 04:54.060 to transform different types of content 04:54.060 --> 04:58.560 that were primarily textual based into an audio narrative. 04:58.560 --> 05:00.310 Cool, I'll see you in the next one.