ChatGPT Voice
Overview
ChatGPT Voice (챗GPT 보이스) is a real-time voice conversation feature that OpenAI (오픈AI) has built into ChatGPT. When a user speaks natural language into a microphone, it recognizes the speech, understands the meaning, and immediately returns a voice response with human-like intonation, pace, and emotional expression. It first gained widespread popularity as the voice mode in the mobile app in 2023, and when it was redesigned in 2024 as "Advanced Voice Mode" based on GPT-4o, it became established as the de facto standard example of an AI voice interface.
Main Content
Background
Existing voice assistants (Siri, Alexa, Google Assistant) operated by processing fixed commands, and had limitations in maintaining conversational context or exchanging emotions. After ChatGPT's release in late 2022, OpenAI set the goal of extending the naturalness of text conversation into voice. In September 2023, voice mode was added to the iOS and Android apps; this was a pipeline structure that sequentially connected speech recognition (Whisper) → language model (GPT-3.5/4) → speech synthesis (TTS). As evaluations continued that response latency was long and emotional expression lacking, OpenAI began a redesign that changed the structure itself.
Advanced Voice Mode
GPT-4o, unveiled in May 2024, is an omni-modal model that directly processes text, images, and audio in a single model. Advanced Voice Mode does not go through the three-stage process of converting speech into text and then synthesizing it back into speech; instead, the model directly understands and generates audio tokens. Thanks to this, average response latency was reduced to the level of hundreds of milliseconds, and interruption handling, non-verbal reactions such as laughter and sighs, and fine adjustment of tone and emotion became possible. After an alpha rollout to Plus users in July 2024, it was officially launched in September, and was later expanded to free users and Team and Enterprise plans.
Voice Types and Characteristics
Early voice mode offered several voices including Breeze, Cove, Ember, Juniper, and Sky, but in May 2024 the "Sky" voice was discontinued after controversy that it resembled the voice of actress Scarlett Johansson from the film 《Her》. Since then, Advanced Voice Mode introduced a new voice set including Arbor, Maple, Sol, Spruce, and Vale. Each voice has a different tone and speaking rate, and users can freely switch between them in settings.
Key Features
- Real-time conversation: Supports interaction close to human conversation, such as interruption, silence, and asking back.
- Multilingual support: Can converse and translate in multiple languages including English, Korean, Japanese, Chinese, Spanish, and French.
- Context retention: Custom Instructions and memory are applied to voice conversations as well, so it remembers previous conversation content.
- Multimodal extension: By combining the camera and screen sharing, users can show objects and hear explanations or request problem-solving.
- Emotion recognition: Infers the speaker's emotional state from the intonation and pace of the voice and adjusts the response tone.
Use Cases
Representative examples include foreign-language conversation practice, information search while on the move, improved information accessibility for users who have difficulty accessing screens, such as the visually impaired, hands-free use in situations where both hands are not free such as driving or cooking, and prototyping customer service. For developers, it is opened up through a real-time API (gpt-realtime family) and is used for call center automation, AI tutors, companion apps, and more.
Limitations and Controversies
The risk of voice cloning and deepfake abuse is cited as a representative problem. Indeed, in 2024, a large-scale financial fraud using deepfake video calls occurred in Hong Kong. In addition, the problem of "attachment use," in which people become emotionally overly dependent on AI voices, and criticism of excessively agreeable responses (sycophancy) have also been raised. Entering 2025, OpenAI has published safety research related to emotional dependence and is strengthening watermarking and usage policies to prevent voice cloning. In the European Union, under the AI Act, transparency labeling obligations for AI-generated voices are being applied in stages.
Latest Trends
The changes of 2024–2025 fall broadly into three streams. First is the opening of models and APIs. After the real-time API was released in October 2024, low-latency speech-to-speech models such as gpt-realtime and gpt-4o-mini-realtime, telephone (SIP) integration, and image input were added in 2025, greatly lowering the barrier to entry for voice AI development. Second is intensifying competition. As Google's Gemini Live and Amazon's next-generation Alexa tout similar real-time voice conversation, voice interfaces have become a basic UX for generative AI. Third is expanded penetration into daily life. With integration into vehicle infotainment (Apple CarPlay and Android Auto), expanded access for free users, and the combination of camera and screen sharing, usage scenarios have broadened across mobility, education, and work overall. At the same time, discussions on regulation and safeguards regarding voice phishing, unauthorized voice cloning, and emotional dependence are also proceeding rapidly.
Related Topics
- [[OpenAI]]
- [[GPT-4o]]
- [[Generative AI]]
- [[Speech Recognition]]
- [[Gemini Live]]
- [[Deepfake]]