OpenAI introduced GPT-Live-1.1 and GPT-Live-1.1 Quip on October 15, alongside 13 new ChatGPT voices. The update promises more natural pronunciation across several languages and improved responses to spoken instructions. It also adds more varied announcements when the assistant searches for information. For multilingual users, the changes could make conversations easier to follow, although OpenAI has not quantified improvements in speech recognition or instruction accuracy.
The new voice options cover French and German, as well as Hindi and Japanese. Korean and Mexican Spanish also receive new options, while English gets a voice called Bubbie. Availability varies by account and region.
GPT-Live-1.1 replaces GPT-Live-1 as the default voice model for Plus and higher-tier users, including in Codex and Dots. Its smaller counterpart, Quip, replaces GPT-Live-1 mini for other users. At launch, OpenAI said API access to version 1.1 would follow.
New voices expand language options
More natural pronunciation could make extended conversations easier to understand, especially for users who regularly switch between languages. Someone practicing French or Japanese could rehearse conversations with the assistant, while multilingual employees might use it to prepare presentations or discuss documents aloud.
However, OpenAI has not published evidence that the update improves language-learning outcomes or the accuracy of spoken corrections. Better pronunciation also does not establish more accurate speech recognition. An assistant could speak Mexican Spanish naturally while still misunderstanding a regional expression or an unfamiliar name.
The launch announcement does not provide language-specific recognition error rates, so how consistently GPT-Live-1.1 understands speakers with different accents remains unanswered. Its advertised improvements primarily concern how the assistant sounds and responds during conversations.
GPT-Live-1.1 builds on continuous conversation
GPT-Live-1.1 builds on full-duplex interaction, which OpenAI introduced with the original GPT-Live on July 8. This capability allows the assistant to process incoming speech while generating a response, so users can interrupt or redirect it during an active conversation.
The original model also introduced delegation, which allows the assistant to pass demanding requests to another model while continuing the conversation. Users can ask follow-up questions or provide additional instructions while a delegated task proceeds.
With version 1.1, OpenAI says delegation feels more natural, and the model responds more readily to user directions, a capability the company calls steerability. The update also introduces more varied spoken status messages during searches, although OpenAI has not published comparative results showing how often the model follows revised instructions correctly.
When users change a request halfway through a conversation, the assistant needs to apply those instructions to the task already underway. For example, narrowing a research request or changing the intended audience should affect the final response, even if the assistant has already delegated the original task.
Voice commands can initiate work, with approval limits
Voice interactions can also initiate supported tasks through connected apps. According to OpenAI’s Voice documentation, Live can access apps available to an account, while Voice in Work supports document creation and browser tasks under the user’s permissions.
Actions requiring approval still need an on-screen response because spoken approval is unsupported. An assistant may understand a request and prepare an action, but completing it can still require confirmation through the interface.
For organizations evaluating voice-driven workflows, instruction accuracy becomes an important measure. A user might ask the assistant to research a topic, then revise the intended audience before requesting a document based on the findings. The completed work would need to reflect those changes and accurately represent the information retrieved.
Natural delivery may make these interactions easier to manage, but the quality of the final output depends on whether the assistant retains updated instructions and executes the requested tasks correctly.
Safety results show mixed findings
OpenAI’s GPT-Live-1.1 system card reports different findings across its safety evaluations. Tests using real-world audio prompts found no statistically significant regressions compared with predecessor models, while synthetic evaluations identified declines in several categories.
For GPT-Live-1.1, synthetic tests found statistically significant regressions involving sexual content and mental health, as well as emotional reliance and gore. OpenAI says further research is needed to understand these results.
The evaluations measure different conditions. Synthetic prompts deliberately examine difficult cases, while the real-world audio tests are not weighted to represent how frequently particular situations occur in everyday use. Consequently, neither set of results establishes how often users might encounter harmful responses during ordinary conversations.
More natural speech may also influence how users judge an assistant’s reliability. A familiar accent or responsive delivery could make an answer sound convincing even when its factual content is incorrect. OpenAI has not established whether GPT-Live-1.1 changes how readily users accept AI-generated information.
The Neuron’s earlier review of ChatGPT Voice documented misheard requests and factual errors alongside improvements in conversational speed. Although that review covered an older version, it shows why pronunciation and response speed should be assessed separately from factual accuracy.
GPT-Live-1.1 expands ChatGPT’s language options and introduces improvements to make spoken conversations easier to direct. However, OpenAI has not published enough comparative performance data to establish how much the update improves speech recognition or task accuracy across languages.
The safety evaluations also indicate that conversational improvements require separate assessment from the model’s responses to sensitive requests. For users who rely on voice to complete work, the relevant measures will be whether the assistant understands changing instructions and produces accurate results during extended conversations.
