OpenAI announced on Thursday (15) an update to ChatGPT’s Voice mode, with improvements to speech recognition, the understanding of different accents, and the naturalness of conversations. The update includes the GPT-Live-1.1 and GPT-Live-1.1 Quip models, a smaller version of the technology, as well as new voice options in different languages.
In Brazil, the update arrives as the use of voice and dictation features on the platform is expanding. According to information released by OpenAI, the country is among the five largest ChatGPT markets in this type of interaction, and the number of weekly Voice mode users doubled over the past year.
The company also emphasizes that the experience has been enhanced for everyday situations, such as asking questions, brainstorming ideas out loud, and asking for help organizing tasks. The aim is to make conversations more fluid, including when users mix languages or need to convey information that combines letters and numbers.
What changes in ChatGPT’s Voice mode?
The update brings together three main improvements: speech understanding, response fluency, and expanded voice options.
- Accent recognition: the models were improved to understand different forms of pronunciation in several languages, including in conversations that mix languages;
- More natural conversations: ChatGPT now varies the short expressions used while carrying out a request more effectively, avoiding repetitions such as “I’m checking this…” while performing a task;
- New voices: options were added in French, German, Hindi, Mexican Spanish, Korean, and Japanese, as well as the Bubbie voice, in English, inspired by the way grandmothers speak.
The update also seeks to improve the understanding of information that combines letters and numbers, which is useful for conveying codes, names, identifiers, and other data during a conversation.
TAG62
GPT-Live-1.1 and Quip: what is the difference?
OpenAI is making two updated versions of the models that power ChatGPT’s Voice mode available.
GPT-Live-1.1 is the successor to GPT-Live-1, while GPT-Live-1.1 Quip replaces GPT-Live-1 mini. According to the company’s technical documentation, the former will be the default voice model for Plus users and higher plans.Quip will be used by default by everyone else.
Both models were developed to enable continuous interactions, in which ChatGPT can listen and respond without waiting for each utterance to end in a perfectly defined way. This helps make interruptions and changes in cadence more natural.
The update also extends to Codex and Dots, OpenAI products that use artificial intelligence (AI) features to assist with tasks and workflows.
The rollout is gradual, and available features may vary depending on the plan and account.
Read more:
- What does artificial intelligence know about you?
- 8 US states sue OpenAI and accuse ChatGPT of putting young people at risk
- OpenAI warns the Federal Police about threats against Lula made on ChatGPT
Both models were developed to enable continuous interactions, in which ChatGPT can listen and respond without waiting for each utterance to end in a perfectly defined way – Image: Mamun sheikh K/Shutterstock
How do the voices sound in Brazilian Portuguese?
Although the update introduces new voices for other languages, OpenAI already offers options developed for Brazilian Portuguese: Viola and Rio. Viola has a light, spontaneous, and curious personality. Rio, meanwhile, was conceived with a warm, welcoming, and helpful style.
These options are part of the strategy of adapting the voice experience to the characteristics of different languages and audiences.The company did not announce Viola and Rio as new features in this particular update, but emphasized that Brazil is an important market for voice and dictation in ChatGPT.
To check the available options, users can open ChatGPT settings, go to the “Voice” section, and select the desired option. The list may vary depending on the region and account.
Voice gains traction in everyday tasks
OpenAI also wants to expand the use of Voice mode beyond quick questions. The tool can be used to explore ideas, organize activities, and develop plans through a conversation.
An example presented by the company is planning a family lunch. Users can specify which ingredients they have in the refrigerator, how many people will be attending the meal, and their dietary preferences. Based on this information, ChatGPT can suggest dishes, ingredient alternatives, and shopping lists.
The advantage of this format is that it allows users to add information and adjust the request naturally, without having to formulate a complete new request after every change.
Voice mode can also be used alongside text and images in the same chat, depending on the features available for the account.Thus, interaction is not limited to speech: users can complement the conversation by typing or sending images.
OpenAI discloses safety evaluations of new models
Alongside the update, OpenAI published technical documentation about GPT-Live-1.1 and GPT-Live-1.1 Quip, with information about training, evaluations, and safety mechanisms.
According to the company, the models generally maintain performance similar to that of their predecessors in safety evaluations. However, the documentation records statistically significant declines for GPT-Live-1.1 in some categories of synthetic tests, including sexual content, mental health, emotional dependence, and graphic violence. OpenAI says further research is needed to understand these differences and that it intends to improve these aspects in future versions.
The company emphasizes that these tests are designed to evaluate specific, challenging situations; therefore, they do not directly reflect the frequency with which such issues occur in everyday use.
The publication of the evaluations comes as voice features evolve and take on a more important role in interactions with ChatGPT. The stated goal is to make conversations more natural in different languages and expand the tool's usefulness in everyday activities.
Rodrigo Mozelli is a journalist graduated from São Paulo Methodist University (UMESP) and is currently a writer at Olhar Digital.
