AI 용어집
인공지능 완전 사전
Audio Encoder
Module, often based on a VQ-VAE or autoencoder, that compresses a raw audio waveform into a lower-dimensional latent representation, better suited for processing by the diffusion process.
Audio Decoder
Neural network that reconstructs an audible audio waveform from the denoised latent representation produced by the diffusion model, finalizing the generation process.
Diffusion Timestep
Discrete variable representing a specific step in the noising or denoising process, controlling the level of noise applied and guiding the model through the transformation from pure noise to coherent audio signal.
Classifier-Guided Inference
Inference method that uses a pre-trained classifier to guide the denoising process toward an output belonging to a specific class (e.g., 'male voice', 'piano'), without modifying the diffusion model weights.
Consistency Model (Constant Throughput Diffusion Model)
Family of diffusion models designed to generate high-quality samples in a single step or very few steps, by learning to maintain consistency across different noise levels, thus drastically reducing inference time.
Velocity Scheduler
Scheduling strategy for the denoising process that determines the sequence of timesteps to use during inference, optimizing the trade-off between the quality of the generated sound and the number of computational steps required.
High-Resolution Audio
Goal of advanced audio diffusion models, aiming to generate waveforms with high sample rates (e.g., 48kHz) and large bit depth (e.g., 24-bit), approaching or exceeding professional recording quality.
Stochastic Diffusion Model
Diffusion approach where the denoising process includes a random component at each step, allowing greater diversity and creativity in audio generations, at the cost of lower reproducibility.
Deterministic Diffusion Model
Variant of the diffusion process where denoising follows a predictable and random-free trajectory, which promotes consistency and stability of results for the same input, often used for precise re-synthesis applications.
Speech Diffusion Model
Specialization of audio diffusion models trained exclusively on speech data, aiming to generate natural and expressive voices with fine control over speaker, intonation, and emotion.
Music Diffusion Model
Application of diffusion to music generation, where the model learns harmonic, rhythmic, and melodic structures to compose entire musical pieces or coherent instrumental samples.
Linear Sampling
Inference strategy where denoising time steps are spaced uniformly across the process timeline, a simple approach but sometimes suboptimal for final audio quality.
Log-Sampling
Inference strategy that concentrates denoising steps at the beginning of the process (when the signal is very noisy) and spaces them out towards the end, which has proven more effective for capturing low-frequency audio structures.