Archive — US — techcrunch
Two undergrads built an AI speech model to rival NotebookLM
Two undergraduates developed an AI speech model named Dia, comparable to Google’s NotebookLM, enabling customizable podcast-style voice generation. Despite their inexperience, they leveraged Google's TPU Research Cloud for training, garnering attention in the growing voice AI market, which raised over $398 million in VC funding last year. However, concerns regarding copyright and model misuse persist.
Key points
- Undergraduates Toby Kim and a co-founder created Dia, an AI speech model similar to NotebookLM.
- They aimed to provide greater control over generated voices and scripts.
- Trained on Google's TPU Research Cloud, Dia boasts 1.6 billion parameters for advanced voice generation.
- Available on Hugging Face and GitHub, Dia can customize tones and nonverbal cues.
- Initial tests show Dia performs comparably to market leaders and features easy voice cloning.
- Concerns exist over the model's potential misuse for disinformation and copyright issues due to undisclosed training data.
Why it matters
- The rise of voice AI and the competitive landscape reveals historically strong investment trends in AI technologies.
- The ethical implications of AI training data reflect ongoing debates about copyright and fair use in the tech community.
- In 2025, the convergence of AI with social media platforms and digital content creation is expected to drive further innovation in this sector.
The development of Dia illustrates a significant leap in accessible AI voice technology, yet it brings to the forefront crucial ethical considerations concerning copyright and potential misuse that need addressing as the industry evolves.