Original link: https://www.wikimoe.com/?post=279
Recently, a live game commentary program was created using the popular ChatGPT.
This program is based on the early years ” Create a virtual anchor of your own!” “The code is transformed.
The basic flow I envision looks something like this:
1. The user sets a profile for the AI.
2. The user describes a game scene.
3. Convert the description into text through the speech recognition function of the browser.
4. The description text is merged into the previously set character settings.
5. Let the AI explain a piece of game content according to the description through the ChatGPT API.
6. Convert the generated text to speech with Azure.
Of course, comment replies can also be made based on viewers’ comments.
Through the above settings, the following Bilibili live broadcast effect can be obtained
< click to enter >
Many problems can be seen in the live broadcast, such as:
The commentary is still very blunt, sometimes even the text is wrong, hahaha.
Although my prompt sentence mentioned adding meow at the end of the sentence, sometimes it is gurgling, or nya, which is very confusing.
Obviously said within 20 words, basically more than 20 words.
Because the return speed of the AI interface is sometimes very slow, it may not be able to explain fast-paced games.
As can be seen from the setting sentences, in order to prevent the AI from introducing itself or introducing the game every time, a lot of restrictive sentences have been added. (Of course, this may also be caused by not uploading complete records due to cost considerations)
In fact, the ideal state I envision is to obtain OBS screenshots every once in a while, identify the elements in the image through the image recognition interface and convert them into text and feed them to ChatGPT for explanation. Enter key information.
Generally speaking, this program can only be regarded as a toy, and even this kind of commentary feels a bit like taking off your pants and farting. After all, you have the energy to prompt the key words, why not just explain it manually, hahaha.
github address: < click to enter >
This article is transferred from: https://www.wikimoe.com/?post=279
This site is only for collection, and the copyright belongs to the original author.