Stabilní latentní šíření zvuku

Fast Timing-Conditioned Latent Audio Diffusion describes a Stable Audio research system using a fully convolutional variational autoencoder to represent audio in a compressed continuous sequence. The diffusion model operates on that sequence, and the decoder reconstructs waveform samples. Its stated original components include a CLAP text encoder, timing embeddings, and a diffusion U-Net. This should not be blended into an unspecified CLAP-or-T 5, U-Net-or-transformer architecture for every product version. Timing provides signals about the start position and total duration. In the paper’s ninety-five-second output window, setting the start to zero and requested total duration to thirty seconds is expected to produce thirty seconds of signal followed by sixty-five seconds of silence. The method can trim that silent remainder to present a variable-length result. This behavior is evaluated in the paper, not a guarantee that every model always follows timing exactly. A text-only generation run begins with a noise latent and iteratively uses the trained denoiser with its conditions before waveform decoding. Audio-to-audio workflows additionally use an input recording and should not be described as the same initialization in every case. The April 2024 Stable Audio 2.0 announcement reports up to three-minute, 44.1-kilohertz stereo generation and audio-to-audio transformation. It also reports licensed AudioSparx training data. That stated dataset provenance is different from a universal claim about the rights or suitability of every output or uploaded recording. Version-specific capability descriptions remain necessary when studying newer Stable Audio models.