待翻譯：Making Audio Stories Interactive with AI: The Story of Schmaudio

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

I have a somewhat complicated relationship with AI. As a software engineer, I find the technology fascinating. I use it regularly, and it has become a genuinely useful tool in my work. At the same time, I am cautious about some of the enthusiasm surrounding it. I don’t think we should automate everything simply because we can, and I am skeptical of the idea that handing a task to an AI and accepting whatever comes back necessarily represents progress. There are also still many open questions around how this technology will affect creative work, software development, and society more broadly. But there is another side to AI that I find much more exciting. Sometimes a new technology doesn’t merely make an existing process faster or cheaper. It makes something practical that previously wasn’t. For me, interactive audio stories are an example of that. And somewhat unexpectedly, they connect modern AI with an idea that fascinated me when I was a child. Choosing What Happens Next As a child, I was fascinated by pen-and-paper role-playing games. I also loved interactive adventure books. Instead of simply reading from the first page to the last, you would eventually have to make a decision. Perhaps you could enter a mysterious cave or continue along the path through the forest. If you entered the cave, the book might tell you to continue on page 74. If you chose the forest, you might jump to page 112. The story changed depending on what you decided. I loved the idea that I wasn’t simply waiting to find out what the character would do next. I had to make that decision myself. Of course, those books had an unavoidable limitation: every possible path had to be written beforehand. And branching grows very quickly. If every episode of a story leads to 3 completely different episodes, those 3 lead to 9, then 27, then 81. Continue for long enough and the number becomes absurd. A human author cannot write millions of episodes for a single story. Traditional interactive stories therefore have to work within that constraint. Branches can eventually reconnect, some decisions can have relatively small consequences, and the overall number of paths has to remain manageable. That doesn’t prevent great interactive stories from being written, but it puts a practical limit on just how far the idea can go. Generative AI changes that constraint. From My Childhood to My Children Many years later, audio stories became part of my own children’s lives. They listen to them regularly, and at some point this brought me back to the interactive stories I had enjoyed as a child. I started wondering what would happen if you combined the two ideas. What if an audio story stopped at an important moment and asked the listener what should happen next? And what if that decision didn’t simply cause a small variation before returning to the same story, but could actually send the adventure in a different direction? Modern generative AI suddenly made that idea practical, so I started experimenting. There wasn’t a business plan behind it. I simply wanted to create interactive audio adventures for my children and see whether the idea worked. A story would play for several minutes. Eventually, the characters would reach a situation requiring a decision. The listener would choose what they should do, and the next episode would continue from that decision. Then another choice would appear, and another. The more I experimented with it, the more I liked the idea. Eventually, that experiment became Schmaudio. AI Isn’t the Interesting Part It would be easy to describe Schmaudio as a platform for AI-generated audio stories. Technically, that’s true, but I don’t think that’s the interesting part. Ideally, when someone listens to a Schmaudio story, they shouldn’t be thinking about AI at all. They should be thinking about the story. The interesting part is what AI makes possible. A listener can make a decision, and that decision becomes part of the history of their adventure. Later episodes can build on what happened before. Characters can have experienced different things, relationships can develop differently, and one path might reveal places or characters that another listener never encounters. Over time, two people who started with exactly the same first episode can experience increasingly different stories. And when someone reaches the end, there is another interesting possibility: What would have happened if I had chosen differently? They can go back and find out. A Story That Grows as You Explore It There is still an obvious practical problem. Even if AI makes it possible to generate all those episodes, it wouldn’t make much sense to generate an enormous story tree before anyone listens to it. Most of those episodes might never be heard. Schmaudio therefore generates episodes lazily. The first episode of a story already exists. When someone listens to an episode for the first time, the possible episodes following it are generated. If there are three choices, three new paths into the story appear. But Schmaudio doesn’t yet need to know what happens ten decisions later. Those parts of the story can be created when someone actually gets there. I like this aspect of the system a lot. The complete story doesn’t have to exist before somebody begins listening to it. Instead, the world grows as people explore it. In theory, this allows stories to become much deeper than what Schmaudio currently offers. More Branches Don’t Automatically Make Better Stories There is an important caveat. Being able to create an enormous number of branches doesn’t mean that doing so automatically produces a better experience. The number of choices isn’t what makes an interactive story interesting. The choices themselves do. A good choice should make you hesitate. Do you trust someone who may be hiding something? Do you follow the strange light disappearing between the trees or help somebody who needs you? Do you investigate the mystery or take the safer route home? There shouldn’t always be an obvious correct answer. Different decisions should appeal to different personalities, priorities, or curiosities, and most importantly, the decision should matter to what happens afterward. AI makes deep branching possible. Good storytelling determines whether that branching is meaningful. How I Use AI This project also reflects how I currently prefer to use AI in my work. I use it as an assistant rather than expecting it to do the entire job for me. For Schmaudio, I currently use ChatGPT to generate the stories and individual episodes, and ElevenLabs to turn those scripts into narrated audio. I developed the web application myself, but AI assisted me there as well. I use ChatGPT to discuss architecture, write and review code, investigate bugs, explore alternatives, and question decisions. The creative process works similarly. Generating a Schmaudio episode isn’t simply a matter of asking an AI to “write the next part.” There is a framework around it. Stories have definitions that establish their worlds, characters, themes, atmosphere, and creative direction. Episodes have to respect everything that happened earlier on the current path. Characters should remember previous events. Choices should emerge naturally from the story and have meaningful consequences. The writing has to work as narration rather than simply as text on a page. I define and continuously refine that system, while AI works within it. In that sense, I don’t see AI as replacing my work on Schmaudio. It allows me to work at a scale that would otherwise be impossible for one person. And, appropriately enough, I also used AI as an assistant while writing this blog post. It Doesn’t Always Work None of this means that I think today’s generative AI is particularly close to perfect. It isn’t. Generating plausible text is remarkably easy now, but generating a consistently good story is much harder. Maintaining a coherent story across many episodes and many decisions is harder still. Characters need to remain consistent. Previous events need to matter. An episode needs to move the story forward rather than merely fill another ten minutes. Choices have to feel natural rather than being added simply because the product requires a decision. And then there is the audio. Text-to-speech technology has improved enormously, but it still has limitations. Pronunciation isn’t always correct, maintaining a consistent voice and delivery across many episodes can be difficult, and producing stories in several languages adds another layer of complexity. These are all areas where I think Schmaudio can become considerably better. I want to improve the quality of the scripts. I want decisions made much earlier in an adventure to influence later events more naturally. I want better long-term narrative coherence, and I want the voices and audio quality to continue improving. The interesting thing is that the idea already works. I can listen to an interactive story, make a decision, hear the adventure continue from that decision, and repeat the process. And this is probably the worst the underlying technology will ever be. I find that exciting. Where Does This Go? It’s difficult to predict what generative AI and text-to-speech will look like several years from now. Even looking only a year ahead, I expect the tools available for building something like Schmaudio to improve considerably. But improvements to the underlying models are only part of it. I am also learning more about interactive storytelling itself. How much should a previous choice affect later episodes? When should branches diverge dramatically, and when is a subtler consequence actually more satisfying? How do you maintain a coherent narrative when the listener can repeatedly push it in different directions? How long should a story be? How much should listeners be encouraged to replay it? Those are storytelling and product questions rather than AI questions, and I suspect answering them will ultimately matter more than simply having access to a better model. Schmaudio today therefore doesn’t explore the full potential of the idea. Stories could go considerably deeper. Choices could have consequences much further into the future. Different paths through the same story could diverge more dramatically while still remaining part of a coherent whole. There is a lot left to experiment with. The Same Idea, Without the Pages Despite all the technology involved, the part I find most interesting remains remarkably simple. A child is listening to a story. The characters reach a crossroads. Something has to happen next. But this time, the listener isn’t waiting to discover what the author decided. They get to decide, and the story continues from there. It’s fundamentally the same idea that fascinated me when I was jumping between numbered pages in adventure books as a child. Except now the number of pages no longer has to be decided in advance. And I think we’re only beginning to discover what that makes possible. Explore Schmaudio If you’d like to try the idea for yourself, you can find Schmaudio at schmaudio.net. Stories are available in multiple languages, and you can make your own choices as you listen. Here are three stories you can currently explore for free: The Alibi A mystery for adult listeners in which a seemingly straightforward story may not be quite what it appears. Listen to The Alibi The Map Awakens An interactive adventure that invites younger listeners into a world of exploration, discovery, and meaningful choices. Listen to The Map Awakens Who Comes Tonight? An interactive story for younger children, offering a gentler introduction to choosing where an audio adventure goes next. Listen to Who Comes Tonight? Tweet Share 0 Skype Reddit +1 Pocket Pinterest 0 LinkedIn 0 Email VKontakte