Search Results for "work"

...Built with a 169 million-parameter architecture that uses dilated convolutions and cross-attention layers instead of large Transformer stacks, it achieves relatively fast real-time performance even on CPUs (about a 0.25 real-time factor measured on an M3 base). The model is designed to work with a small set of dependencies and to be accessible for developers who want offline TTS with customizable voice style, including options for streaming or non-streaming generation modes. Users can install it with standard Python tools, run a demo server locally, and experiment with CLI or Python API usage for producing synthetic speech.

...It also includes a notion of speaker profiles: you can create a speaker from a short audio sample, save it as JSON, and reuse it for consistent voice identity across generations and sessions. For best quality, the model is designed to work with a reference speaker clip and will inherit emotion, style, and accent from that reference.

...Specifically, we train a neural codec language model (called VALL-E) using discrete codes derived from an off-the-shelf neural audio codec model, and regard TTS as a conditional language modeling task rather than continuous signal regression as in previous work. During the pre-training stage, we scale up the TTS training data to 60K hours of English speech which is hundreds of times larger than existing systems. VALL-E emerges in-context learning capabilities and can be used to synthesize high-quality personalized speech with only a 3-second enrolled recording of an unseen speaker as an acoustic prompt. ...