OpenAI TTS-1

Apie OpenAI

Įkurta 2015 · San Francisco, California, USA

OpenAI was founded in December 2015 by Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever, Wojciech Zaremba and John Schulman as a non-profit AI research lab. It restructured into the capped-profit OpenAI LP in 2019 and has since received over $13B from Microsoft plus billions more from Thrive Capital, Sequoia, Khosla and SoftBank, at a 2024 valuation north of $157B and a planned 2025 round reportedly at $500B. The audio research line at OpenAI includes Jukebox (2020), Whisper (2022) and the TTS-1 family released in November 2023 alongside the ChatGPT voice feature. TTS-1 powers ChatGPT Voice Mode, the OpenAI API audio endpoint and the Realtime API used in voice agents.

Architektūra

Proprietary Transformer text-to-speech with neural codec

OpenAI TTS-1 is the standard-quality variant of OpenAI's text-to-speech model family, optimised for real-time streaming inside ChatGPT Voice and the Realtime API. It is a Transformer-based model that predicts neural-codec audio tokens conditioned on text and a fixed set of stock voices (alloy, echo, fable, onyx, nova, shimmer) that OpenAI recorded with professional voice talent. There is no voice cloning. The model supports six output formats (MP3, Opus, AAC, FLAC, WAV, PCM) and outputs at 24 kHz. TTS-1 is tuned for low latency and starts streaming audio within a few hundred milliseconds, making it suitable for live conversational agents, while the HD sibling trades latency for fidelity. Text input is capped at 4,096 characters per request. OpenAI has not published a technical paper but the system architecturally resembles published neural-codec TTS such as VALL-E.

Parametrai

Undisclosed

Kontekstas

4 096 žetonų

Galimybės

• Real-time streaming TTS with sub-second first-byte latency
• Six recorded voices: alloy, echo, fable, onyx, nova, shimmer
• Multilingual: 50+ languages including English, German, Spanish, French, Italian, Japanese, Mandarin
• Six output formats including streaming PCM for low latency
• Powers ChatGPT Voice and the OpenAI Realtime API
• Inexpensive at $0.015 per 1,000 characters (vs. $0.03 for HD)
• Best for: voice agents, IVR, interactive narration, accessibility

Mokymas ir licencija

Not disclosed. OpenAI states voices were recorded with paid professional voice actors and the model was trained on a mixture of recorded speech and licensed text-audio pairs.

Licencija: Proprietary commercial API. Generated audio may be used commercially under the OpenAI Usage Policy; the six stock voices cannot be impersonated outside the API.

Saugumo testai: OpenAI's Preparedness team red-teamed Voice Mode; outputs include watermark provenance metadata in the Realtime API where supported. Voice cloning is intentionally not exposed.