OpenAI has just rolled out some exciting new voice models aimed at developers, enhancing its suite of real-time audio tools that can speak, translate, and transcribe conversations on the fly.
The company emphasizes that these new tools are crafted to make voice-driven applications more practical in everyday scenarios, particularly when users need software that can respond naturally while multitasking in real time.
At the heart of this launch is GPT-Realtime-2, a voice model that OpenAI claims can tackle more complex requests while keeping conversations smooth and engaging. Unlike its predecessors, this model leverages GPT-5-level reasoning to better manage interruptions, grasp context, and perform actions during discussions.
Additionally, OpenAI introduced GPT-Realtime-Translate, a live translation model capable of converting speech from over 70 languages into 13 different output languages.
According to the company, their system is designed to keep up with speakers during conversations, rather than waiting for pauses or completed sentences to translate. The third model, GPT-Realtime-Whisper, is all about live speech transcription, turning spoken words into text in real-time as someone talks. “With these models we’re launching, we’re shifting real-time audio from just simple back-and-forth exchanges to voice interfaces that can genuinely perform tasks: they can listen, reason, translate, transcribe, and take action as conversations happen,” OpenAI explained.
Voice technology has become a major focus for tech companies, especially as more users prefer interacting with software through speech rather than typing. OpenAI noted that developers are looking for systems that can handle tasks while conversations flow naturally. They highlighted areas like customer support, travel, education, and media as sectors that are likely to benefit from these new models.
OpenAI also pointed out three emerging trends in voice-based software. The first is “voice-to-action,” where users can speak naturally and the system takes care of tasks for them. For instance, OpenAI mentioned that the property platform Zillow is developing an assistant that helps users search for homes, avoid certain neighborhood issues, and book tours just by using voice commands.
Another interesting trend is the “systems-to-voice” approach, where software automatically delivers spoken updates. For instance, OpenAI highlighted how travel apps could notify passengers about flight delays, new boarding gates, or transfer routes without requiring any input from users.
Then there’s the “voice-to-voice” area, which is all about facilitating live multilingual conversations. OpenAI mentioned that Deutsche Telekom is working on customer support systems that can translate discussions in real-time, allowing both parties to speak in their preferred languages without missing a beat.
Additionally, travel company Priceline is developing voice-driven tools for managing trips, as noted by OpenAI. This means travelers might soon be able to book flights, modify hotel reservations, and get airport updates just by chatting.
Along with this wider rollout, OpenAI has introduced several new features to GPT-Realtime-2 to enhance live interactions. Developers can now incorporate brief phrases like “let me check that” or “one moment while I look into it” before the system wraps up a request. OpenAI explained that this provides users with clearer feedback while the model works behind the scenes.
The model is also capable of using multiple tools simultaneously and can vocalize those actions during conversations. For example, OpenAI mentioned that the system might say things like “checking your calendar” or “looking that up now” while it processes requests.
Another interesting trend is the “systems-to-voice” approach, where software automatically delivers spoken updates. For instance, OpenAI highlighted how travel apps could notify passengers about flight delays, new boarding gates, or transfer routes without requiring any input from users.
Then there’s the “voice-to-voice” area, which is all about facilitating live multilingual conversations. OpenAI mentioned that Deutsche Telekom is working on customer support systems that can translate discussions in real-time, allowing both parties to speak in their preferred languages without missing a beat.
Travel company Priceline is also diving into voice-driven trip management tools, as noted by OpenAI. This means travelers might soon be able to book flights, modify hotel reservations, and get airport updates just by chatting.
In addition to this wider rollout, OpenAI has introduced several new features to GPT-Realtime-2 to enhance live interactions. Developers can now incorporate brief phrases like “let me check that” or “one moment while I look into it” before the system wraps up a request. OpenAI explained that this provides users with clearer feedback while the model works behind the scenes.
The model is also capable of calling multiple tools simultaneously and can articulate those actions during conversations. For example, it might say things like “checking your calendar” or “looking that up now” as it processes requests.
Moreover, OpenAI pointed out that GPT-Realtime-2 has improved its ability to recover from errors or failed requests, avoiding abrupt conversation halts. It also boasts a larger context window, expanding from 32K to 128K, which allows for longer and more detailed discussions.
Lastly, OpenAI noted that the model has enhanced its understanding of specialized terminology, including healthcare jargon and proper nouns. Developers can also tweak the model’s reasoning capacity based on the complexity of the request.
According to the latest benchmark figures shared by the company, GPT-Realtime-2 has outperformed GPT-Realtime-1.5 in both audio intelligence and instruction-following tests. All three models can be accessed via OpenAI’s Realtime API.
The company mentioned that GPT-Realtime-Translate and GPT-Realtime-Whisper will be charged by the minute, while the pricing for GPT-Realtime-2 will vary based on token usage. OpenAI also highlighted that they’ve implemented safeguards to minimize misuse, including measures against spam, fraud, and harmful content. Additionally, they noted that conversations can be automatically halted if they violate safety guidelines.
Do you like the voice model from OpenAI
