text speech 2.5 Key Features

The speech 2.5 turbo preview voice clone model can handle up to 10,000 characters of text per request. This allows for long-form content generation, though we recommend breaking extremely long text into chunks for optimal streaming performance. This ensures the text to audio conversion remains stable and responsive for the end user.

Yes. You only need a 3-6 second audio reference. The model uses this sample to clone the timbre and prosody of the target voice immediately without any additional training or fine-tuning, making it ideal for rapid deployment. The resulting speech maintains the accent and unique characteristics of the original speaker with high accuracy.

This 2.5 turbo model is engineered for real-time interaction. It achieves a Time-to-First-Audio (TTFA) of approximately 280ms. This sub-second latency ensures that your voice assistants respond to text inputs almost instantly, providing a fluid conversation experience that feels human and natural.

Absolutely. It excels at code-switching, meaning it can transition between two languages, such as English and Chinese, within a single text sentence without losing voice quality or creating audio glitches during the speech output. This makes it perfect for international applications and global users.

Yes, the 2.5 model supports 48kHz sampling rates. This high-definition output is suitable for professional media production, gaming, and any application where the audio quality of the speech must meet broadcast standards. You can specify the response format as flac or pcm to preserve this high-fidelity quality.

You can access the model via our OpenAI-compatible API. This allows for unified billing and no minimum monthly spend. Simply update your base URL and use your GPTProto key to start generating high-quality speech from your text. Our platform provides the lowest latency routing to ensure your application performs at its best.