Home / Technology / OpenAI's New Voice AI Stumbles on Simple Tasks
OpenAI's New Voice AI Stumbles on Simple Tasks
10 Jul
Summary
- New voice model failed to count 'E's in 'seventeen'.
- Users report issues with basic counting and interruptions.
- OpenAI's flagship voice model struggles with simple prompts.
OpenAI recently launched its new generation voice model, GPT-Live-1, emphasizing natural human-AI interaction. However, the full-duplex model, designed to talk and listen simultaneously, has shown significant flaws. During a test, it incorrectly identified the number of 'E's in 'seventeen' and offered an awkward sign-off.
Further tests reveal broader issues, with users reporting failures on simple word-counting challenges, such as the 'Strawberry' test. A prominent TikToker known for testing AI voice models also highlighted the model's struggles.
Adding to the concerns, the model's full-duplex functionality, intended for smoother conversation with interjections like 'mhm,' has reportedly led to incessant interruptions. This has left users feeling the AI is incapable of simply listening. The performance has drawn criticism, with even OpenAI engineers expressing disappointment online.
