Google's NotebookLM has long been praised for turning research documents into engaging podcast episodes, but a direct comparison with Anthropic's Claude Opus 4.8 reveals significant gaps in how the two AI models approach creative audio content generation. A journalist who has used NotebookLM since its early days discovered that when the same research material was processed through Claude instead, the resulting podcast was noticeably more detailed, varied, and engaging than anything NotebookLM had produced.

What Makes NotebookLM's Podcasts Sound Repetitive?

NotebookLM's Audio Overview feature, launched in September 2024, generates podcast episodes by having two AI hosts discuss your uploaded documents. The feature has become popular enough that users can now customize episode length, add follow-up questions, and choose from different formats like "Brief," "Debate," or "Critique." Yet despite these customization options, a consistent problem emerges when you listen to many episodes: they all sound remarkably similar.

The underlying issue traces back to the language model (LLM), which is an AI system trained to predict and generate text, powering the script generation. NotebookLM uses Google's Gemini 3.5 model, and like all LLMs, Gemini has its own distinctive "personality" reflected in how it writes, the phrases it favors, and the rhythms it naturally falls into. This means every podcast generated through NotebookLM exhibits the same structural patterns, verbal tics, and tonal consistency, regardless of the source material.

Users who listen closely notice recurring elements across episodes:

• Opening Pattern: Nearly every podcast begins with the same warm "so, let's get into it" introduction, creating a formulaic start.
• Call-and-Response Rhythm: The two hosts follow an identical back-and-forth pattern where one makes a point and the other affirms it with phrases like "wow, that's fascinating."
• Mid-Episode Pivot: At roughly the same point in every episode, one host suddenly shifts the conversation with "okay, but here's what really stood out to me," creating predictable structure.
• Tonal Uniformity: Every source, whether dry academic research or depressing subject matter, receives the same level of enthusiasm and positive framing.
• Closing Formula: Episodes wrap up with nearly identical "what a fascinating topic" send-offs, regardless of content.

How Does Claude Opus 4.8 Compare in a Real Test?

To test whether a different model could break this monotony, a journalist used Open Notebook, an open-source, self-hosted alternative to NotebookLM that allows users to choose which AI model generates the podcast script. Open Notebook was built by developer Luis Novo and released under the MIT license, giving users control over the language model, embedding model (which processes source material), and text-to-speech model at every stage.

For the experiment, the journalist uploaded an 80-page doctoral dissertation on Generation Z's attitudes toward mental health services and generated a podcast using Claude Opus 4.8 paired with OpenAI's text-embedding-3-small for processing sources and OpenAI's tts-1 for voice generation. The entire process took approximately 18 minutes from start to finish, with the script itself completed in about two minutes. The remaining time was spent rendering audio.

The results were striking. Claude Opus 4.8 produced a 20-minute episode with significantly more depth, variation in conversational flow, and nuanced discussion of the source material compared to NotebookLM's typical output. The script demonstrated greater storytelling flexibility, less reliance on formulaic transitions, and more natural integration of complex research concepts.

How to Generate Custom AI Podcasts With Different Models

If you want to experiment with different AI models for podcast generation beyond what NotebookLM offers, here are the key steps:

• Set Up Open Notebook: Download and run Open Notebook locally using Docker, an open-source containerization platform that lets you run software in isolated environments. This requires some technical setup but keeps all your data on your own machine rather than sending it to cloud servers.
• Obtain API Keys: Sign up for API access from your preferred AI providers, such as Anthropic (for Claude), OpenAI, Google, or xAI. Each provider offers different models with varying strengths in writing, creativity, and reasoning.
• Configure Speaker Profiles: Before generating a podcast, define how many hosts you want, what tone they should use, and which models should handle the outline, script, and voice generation. This gives you granular control over every aspect of the final episode.
• Select Your Models: Choose a language model for script writing, an embedding model for processing your source documents, and a text-to-speech model for voice generation. You can mix and match providers to optimize for your specific needs.
• Upload Sources and Generate: Add your documents, PDFs, links, YouTube videos, or PowerPoint slides, then hit generate. The system will create an outline, expand it into a full script, and render the audio in stages.

Why Model Choice Matters More Than You Might Think

The comparison between NotebookLM and Claude Opus 4.8 highlights a broader truth about AI-generated content: the underlying model fundamentally shapes the output in ways that customization settings alone cannot fix. NotebookLM's limitation isn't a bug or missing feature; it's a consequence of relying on a single model for script generation.

Claude Opus 4.8 has been widely praised for its creativity, storytelling ability, and nuanced writing style. When given the same research material, it produced a podcast that felt less like a template being filled in and more like a genuine conversation between two people who had deeply engaged with the source material. The difference suggests that for users who generate many research podcasts, the choice of underlying model can be as important as the tool's interface or customization options.

This finding also raises questions about NotebookLM's future direction. Google has been expanding the tool beyond podcasts, adding features like document summarization and insight generation. However, as long as the tool remains locked to Gemini models, users seeking greater variety in podcast tone and structure will face the same repetitive patterns, no matter which customization options they select.

For researchers, students, and professionals who rely on AI-generated podcasts as a learning or productivity tool, the message is clear: if monotony in your generated content is becoming a problem, exploring alternatives that let you swap in different AI models might be worth the technical effort required to set them up.