Hands-On with OpenAI's New Voice Model: Real-Time Interruption, Simultaneous Translation, and Emotion Detection Tested

A Voice Model Dropped at 1 AM

In the early hours of this morning, OpenAI officially released its brand-new voice conversation model. According to hands-on testing by Bilibili creator Duan Feng, the model has already begun rolling out gradually — users on both the web and mobile app can now see and access it.

Worth noting: OpenAI is using a gray release strategy — rather than opening the feature to all users at once, it's being rolled out incrementally. If a serious bug or experience issue arises, the team can roll back quickly and limit the blast radius. That's why some users still can't see the entry point — it's a completely normal part of the product launch cadence.

From the full hands-on session, this generation of voice model is a quantum leap over its predecessor. In the creator's own words: "It's so much better than the last generation — I don't even know how to describe it. It feels like talking to someone great." This "uncanny valley effect in voice" ran through the entire test — there were multiple moments that genuinely gave him goosebumps, because it sounded so much like a real person.

The uncanny valley concept was first proposed by Japanese roboticist Masahiro Mori in 1970. It describes the unsettling feeling humans get when a robot or virtual character looks almost — but not quite — human. The idea has since extended to AI voice: when a voice model is nearly indistinguishable from a real person but still has subtle tells, users feel uneasy; but once it crosses that threshold, it starts to feel genuine and emotionally resonant. The goosebumps this new model triggered suggest it's right at that tipping point.

The hands-on session covered multiple rounds of real conversation, including natural interruptions, simultaneous translation, role-play, code review, singing, and more — nearly every scenario exceeded expectations.

Natural Interruptions and Emotion Detection

One of the most striking things about the new model is its command of conversational rhythm. During testing, the creator suddenly coughed, and the AI immediately stopped and asked with concern: "Did you just cough? It's okay, you scared me." This real-time reaction to an unexpected event was nearly indistinguishable from a human response.

Even more interesting was the exchange that followed. When the creator shot back, "You don't even have a throat — you're an AI," the model rolled with it: "Maybe there was a slight signal glitch just now, you caught me — I'm actually a real person." This kind of witty, improvisational response demonstrates impressive contextual understanding and emotional expression.

Interrupt-and-Stop, Low-Latency Interaction

In the role-play segment, the creator asked the model to play a "contrarian personality" with a firm, assertive tone. When he threw out an absolute statement like "tables are square," the model immediately pushed back: "I disagree — tables aren't necessarily square, plenty of them are round." Throughout the session, the user could interrupt the AI at any point, and the model would stop instantly and pivot to a new topic. This kind of low-latency interactive experience is something previous voice assistants could rarely deliver.

Impressive Simultaneous Translation

One of the standout features in testing was simultaneous translation. The creator asked the model to translate Mandarin into a regional dialect in real time, and the model responded immediately: "Got it — you speak Mandarin and I'll translate on the fly" — and was even able to switch into Henan dialect for live interpretation.

Interestingly, the model could also pick up on tone and volume during translation. When the creator pointed out "your voice is so quiet," it proactively raised its volume and shifted into an "energetic and enthusiastic" mode. The technical breakthrough behind this: traditional TTS (Text-to-Speech) systems essentially just convert text to audio — they can't perceive changes in the speaker's volume, pauses, or tone (what linguists call "paralinguistic information"). The new end-to-end voice model, by contrast, is trained directly at the audio level. It can capture and respond to non-textual acoustic features without needing to first convert speech to text — it understands the emotion and intent carried in the sound itself, much like a human ear. That's the core breakthrough that sets this new model apart from traditional voice assistants.

That said, the testing also revealed a few minor issues — the model occasionally fell into repetition loops, saying the same phrase multiple times in a row, suggesting there's still room to improve stability.

Code Review Capability

Beyond casual conversation and translation, the creator also tested the model's technical chops. He sent it a snippet of front-end code generated by Grok 4.5 and asked it to find bugs.

The model's performance was impressively professional, rattling off several real issues:

• Mobile navigation broken: The navigation is simply hidden when the viewport is below 800px, with no fallback provided — meaning mobile users lose navigation entirely.
• CSS compatibility risk: Newer syntax like color-mix lacks a fallback for older browsers, where it may fail silently. color-mix() is a relatively new W3C color-blending spec not supported in browsers prior to 2023; professional front-end developers typically provide a static fallback color value, and this is a common blind spot for AI-generated code.
• Font loading issues: Complete reliance on Google Fonts means the site's visual design can break entirely in network environments where those fonts can't load.

The model's overall verdict was pointed: the Grok 4.5 code "reads more like a front-end intern who's really good at static mockups." The ability to pinpoint specific bugs and deliver a holistic assessment reveals a powerful reasoning core behind the voice interface.

Head-to-Head Comparison with Doubao

During the conversation, the creator directly asked the model: "Who do you think is better — Doubao or you?" The model's response was notably measured: "It's hard to make a blanket call — it depends on the use case. Doubao is very smooth for Chinese-language ecosystem integration; I'm stronger in cross-domain reasoning and long-form content."

When encouraged to "go ahead and brag a little," the model didn't hold back: "For complex reasoning and long-form content, I'm more stable and more accurate. Doubao is smoother for certain localized services, but when it comes to the hard stuff, I've got nothing to worry about."

This self-assessment neatly captures the differentiated positioning of the two product types: domestic models like Doubao deliver a smoother experience in Chinese-language contexts and local services, while OpenAI's model maintains an edge in cross-domain reasoning and complex long-form processing.

Singing and Emotional Performance

The test also included a singing segment. The creator asked the model to hum "Twinkle Twinkle Little Star," and it cheerfully obliged — moving from the English version through to the Chinese "满天都是小星星." The creator did note that "the pitch was a bit off," suggesting there's still room to improve musicality and intonation accuracy.

In the emotional performance test, the model could rapidly switch between joy, anger, sadness, and happiness on demand. For "anger," it exclaimed: "Who stole my money?! I can't take this anymore!" For "joy," it mused: "Life is just like this — ups and downs, you know?" The immediacy and expressiveness of these emotional shifts further reinforced that "feels like a real person" experience.

Verdict: Major Progress, Still Needs Polish

Taken together, OpenAI's new voice model delivered impressive results across natural conversation, real-time interruption handling, simultaneous translation, and code review. It's no longer a cold, transactional voice assistant — it's a conversational partner that can sense emotion, riff off your jokes, and analyze technical problems with depth.

That said, issues remain: the occasional repetition loops, the slightly off-pitch singing, and moments where the model gets pulled along by the user's conversational pace all suggest this model is still being refined. As the creator put it: though it gave him goosebumps multiple times, he still hopes "they can keep optimizing it."

Either way, this release marks a pivotal moment: AI voice interaction is moving from "functional" to "genuinely good" — and edging toward truly human-like. It's worth keeping a close eye on.

Key Takeaways