📋TestingCatalog•較早收集於 18h
OpenAI 準備透過 GPT-Bidi-1 模型進行重大語音升級
#voice-ai#model-architecturegpt-bidi-1openaichatgptgpt-bidi-1
💡ChatGPT 即將迎來全新的雙向音訊模型，預計將大幅降低延遲並帶來更自然的語音互動體驗。
⚡ 30-Second TL;DR
有什麼變化
引入 GPT-Bidi-1 雙向音訊模型
為什麼重要
雙向音訊模型能降低延遲並改善對話流暢度，使語音 AI 代理在專業與客戶服務應用中更具可行性。
下一步行動
請測試目前的語音 API 並建立基準，以便在 GPT-Bidi-1 發布後評估其在降低延遲方面的效能提升。
誰應關注：Developers & AI Engineers
關鍵要點
- •引入 GPT-Bidi-1 雙向音訊模型
- •專注於增強即時語音對話能力
- •代表語音模式的重大架構升級
🧠 深度解析
背景與延伸：來自公開資料，非原文內容。引用 15 個來源。
🔑 增強重點摘要
- •GPT-Bidi-1 represents a shift from turn-based voice interactions to a continuous, bidirectional architecture, allowing the AI to listen and speak simultaneously, absorb interruptions, and adjust its response mid-sentence.
- •This new model aims to bridge the performance gap between OpenAI's highly advanced text models (like GPT-5.5) and its previous voice capabilities, aligning with OpenAI's strategic vision for speech to become the primary interface for AI.
- •The upcoming voice mode is expected to offer users different 'intelligence levels' (High, Medium, Instant), mirroring the existing tiers for text models, which will allow for a trade-off between response speed and conversational depth.
- •OpenAI has also introduced a suite of developer-focused 'GPT-Realtime' models, including GPT-Realtime-2 for live reasoning, GPT-Realtime-Translate for speech-to-speech translation across over 70 input languages, and GPT-Realtime-Whisper for streaming transcription.
- •The continuous, bidirectional sound channel enabled by GPT-Bidi-1 could unlock novel 'data-over-sound' applications, allowing inaudible ultrasonic signals to carry data for identity verification, proof of presence, or transaction authorization alongside voice.
📊 競品分析▸ Show
Competitor Analysis: Real-time Voice AI Models
| Feature / Platform | OpenAI (GPT-Bidi-1 / GPT-Realtime) | Krater.ai Voice Mode | Google Gemini Live | Inworld AI (Realtime TTS-2) | ElevenLabs | Cartesia Sonic 3.5 Turbo |
|---|---|---|---|---|---|---|
| Core Capability | Bidirectional, real-time voice, interruption handling, reasoning, translation, transcription | Real-time bidirectional voice, 350+ AI models access | Conversational AI voice, deeply integrated with Google ecosystem | Real-time TTS, full voice pipeline, LLM orchestration | High-quality TTS, voice cloning, multilingual, content creation | Industry-leading low latency TTS, high naturalness, accurate transcript following |
| Latency (End-to-End) | Aims for significantly reduced latency, moving from turn-based to continuous processing | Sub-second (real-time audio streaming) | Low | Sub-250ms P90 (Max model), sub-130ms (Mini model) | Not specified as primary focus for real-time, but strong quality | ~40ms TTFB (Time To First Byte) |
| Multimodal Support | Advanced Voice Mode with visual context (image recognition) | Yes – persistent photo attachment during voice sessions | Camera and screen sharing on mobile | Not explicitly detailed for multimodal input beyond voice | Primarily audio generation, some dubbing | Primarily audio generation |
| Model Access | OpenAI models only (GPT-Bidi-1, GPT-Realtime family) | 350+ AI models across text, image, video, code, audio | Gemini models only | Realtime Router across 200+ LLMs | Proprietary ElevenLabs models | Proprietary Cartesia models |
| Pricing | Included in Plus ($20/mo) and Pro plans for ChatGPT voice; API pricing for GPT-Realtime models (e.g., GPT-Realtime-2, Translate, Whisper) | 0.5 credits/second, transparent credit-based billing from $9/mo | Included in Google One AI Premium ($19.99/mo) | Competitive per-character pricing, significantly less expensive than alternatives at comparable quality | Usage-based, various tiers for instant/professional cloning, sound effects, dubbing | Not explicitly detailed, but noted for low latency |
| Language Support | 13 fully integrated languages for end-to-end voice, 70+ input languages for translation | Not specified | Not specified | Realtime TTS-2 adds cross-lingual voice identity across 100+ languages | 70+ languages | 42 languages including English, Hindi, Spanish, French, German, Japanese, Hebrew |
| Key Differentiator | Bidirectional architecture for human-like conversational flow, strategic bet on voice-first devices | Access to a vast array of AI models under one subscription | Deep integration with Google's ecosystem, camera/screen sharing | #1 ranked real-time TTS on Artificial Analysis, full voice pipeline | Extensive voice catalog, emotional range, advanced voice cloning, content creation focus | Lowest latency for TTS, strong naturalness for conversational AI |
🛠️ 技術深入
- Bidirectional (BiDi) Architecture: GPT-Bidi-1 moves away from the traditional 'turn-based' voice AI systems, where the model waits for a user to finish speaking before processing and generating a response. Instead, BiDi continuously processes the speaker's voice, allowing it to listen and speak simultaneously, absorb interruptions, and adjust its output mid-sentence.
- Latency Reduction: Traditional voice pipelines involve sequential steps (speech-to-text, LLM processing, text-to-speech), which introduce significant latency. Bidirectional audio models collapse parts of this pipeline by processing audio directly and continuously, drastically reducing the delay and enabling more fluid interactions.
- Interruption Handling: A key feature of the BiDi architecture is its ability to handle user interruptions (e.g., 'mm-hm' or mid-sentence changes) without stopping or appearing confused, allowing for a more natural and human-like conversational flow.
- Vocal Nuance and Emotion: By processing audio directly, these real-time native audio models can better pick up on vocal nuances, tone, and emotional cues, leading to more empathetic and contextually appropriate responses.
- GPT-Realtime Family: OpenAI has also released specialized models for developers:
- GPT-Realtime-2: A reasoning model specifically designed for real-time voice interactions and live agents.
- GPT-Realtime-Translate: A streaming speech-to-speech translation model capable of understanding over 70 input languages and speaking in 13 output languages, aiming to preserve meaning and handle accents.
- GPT-Realtime-Whisper: A streaming speech-to-text model for real-time transcription, useful for live captions, notes, and summaries.
🔮 前景展望AI analysis grounded in cited sources
GPT-Bidi-1 will accelerate the development and adoption of voice-first AI devices and interfaces.
OpenAI's focus on a natural, interruption-tolerant conversational model like BiDi is a strategic move to make sound the primary interface for AI, paving the way for more intuitive smart speakers and other voice-controlled hardware.
The model will significantly enhance the quality and empathy of AI interactions in customer service, education, and personal assistance.
By enabling real-time adjustments and seamless handling of interruptions, GPT-Bidi-1 can make AI conversations feel more natural, less rigid, and more capable of understanding complex, evolving user needs and emotional states.
Bidirectional audio models could enable new forms of 'data-over-sound' communication for secure, hands-free transactions and identity verification.
The continuous, bidirectional sound channel established by BiDi creates an opportunity to embed inaudible ultrasonic signals for data transmission, facilitating secure interactions in environments where traditional methods like GPS or screens are impractical.
⏳ 時間線
2024-09
OpenAI launches pilot of 'Advanced Voice Mode' for paid ChatGPT subscribers
2024-10
ChatGPT's Advanced Voice Mode enables emotionally nuanced conversations
2024-12
ChatGPT's Advanced Voice Mode integrates visual context capabilities for multimodal conversations
2025-01
User feedback highlights perceived issues and unnaturalness in Advanced Voice Mode compared to earlier versions
2026-05
OpenAI introduces developer-facing real-time audio models: GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper
2026-06
OpenAI prepares for the launch of GPT-Bidi-1, a next-generation bidirectional audio model for ChatGPT
📎 來源 (15)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： TestingCatalog ↗
每週 AI 簡報
每週一封，可隨時退訂。