Our first impressions of the GPT-Live-1 API for voice agents (thunderphone.com)

OpenAI has introduced the GPT-Live-1 API, a cutting-edge full-duplex voice model designed for AI phone agents. Preliminary testing shows that GPT-Live-1 offers a remarkably natural conversational experience, responding almost immediately to interruptions and engaging in genuine dialogue without distinct phases of transcription and synthesis. During extensive trials involving scripted conversations, the model demonstrated impressive adaptability, fluently handling various conversational dynamics and maintaining an engaging pace. However, it fell short in reliably following instructions, often misinterpreting clear affirmative responses or mechanically adhering to prompts rather than responding contextually. This limitation raises concerns about its applicability in settings requiring precise communication. Technical tests also revealed challenges with alphanumeric data and language accents. The model struggled with accurately spelling numbers and special characters, which could impede its usability in settings like customer support when precision is crucial. Additionally, while GPT-Live-1 performed admirably in English, it displayed a notable American accent in other languages, potentially affecting the user experience for non-English-speaking callers. Despite its significant advancements in natural language processing, the current shortcomings in instruction adherence and data handling highlight areas requiring improvement before widespread deployment in production environments. OpenAI plans to continue testing and iterating on this model, suggesting that further refinements are on the horizon.