OpenAI TTS-1 HD

Про OpenAI

Засновано 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, restructured to capped-profit OpenAI LP in 2019. The company is best known for the GPT, DALL-E and Whisper model families and has raised more than $13B from Microsoft and billions more from Thrive Capital, Sequoia, Khosla and SoftBank, with a 2024 valuation above $157B. The TTS-1 HD model launched together with TTS-1 in November 2023 as the high-quality sibling, intended for offline audiobook, podcast and content-production workflows where fidelity matters more than latency. It powers the higher-quality option in the OpenAI audio endpoint and is exposed in tools like ChatGPT's Read Aloud feature on premium tiers.

Архітектура

Proprietary Transformer text-to-speech with neural codec (high-fidelity variant)

OpenAI TTS-1 HD is the high-fidelity variant of the TTS-1 family. It uses the same architectural family (Transformer-based neural-codec language-model TTS) and the same six voices (alloy, echo, fable, onyx, nova, shimmer) as TTS-1, but with a larger model and additional fine-tuning on long-form narration. The result is more natural prosody, clearer diction and better handling of long sentences, punctuation and emotional pacing, at roughly 2x the price ($0.030 per 1,000 characters) and noticeably higher latency. Output is 24 kHz in six formats (MP3, Opus, AAC, FLAC, WAV, PCM). Text input is capped at 4,096 characters per request, which is recommended to be split into paragraph chunks for very long narration. The model targets offline production: audiobooks, podcasts, training videos, accessibility narration. It does not support voice cloning.

Параметри

Undisclosed (larger than TTS-1)

Контекст

4 096 токенів

Можливості

• High-fidelity narration with natural prosody and emotional pacing
• Six recorded voices: alloy, echo, fable, onyx, nova, shimmer
• Multilingual: 50+ languages including English, German, Spanish, French, Italian, Japanese, Mandarin
• Six output formats including FLAC for lossless production
• Long-sentence handling tuned for audiobook-style narration
• Drop-in API-compatible with TTS-1 (just change the model name)
• Best for: audiobooks, podcasts, training content, voiceover, accessibility narration

Навчання та ліцензія

Not disclosed. OpenAI states voices were recorded with paid professional actors; training corpus is a curated mix of recorded speech and licensed text-audio pairs.

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

Тестування безпеки: Same Preparedness red-team coverage as TTS-1; voice cloning is intentionally not exposed; outputs may carry provenance metadata.