KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Mel Spectrogram
Time-frequency representation of the audio signal on a Mel scale, which mimics human hearing perception and is commonly used as input for audio diffusion models.
Audio Noising Process
Forward step of the diffusion model where Gaussian noise is iteratively added to a clean audio signal over multiple time steps, until the signal becomes pure noise.
Audio Denoising Process
Generation phase where a neural network, often a U-Net, learns to predict and subtract the noise added at each step to reconstruct a coherent audio signal from noise.
Audio U-Net
Encoder-decoder neural network architecture with skip connections, adapted to process spectrograms and predict noise at each step of the audio denoising process.
Audio Conditioning
Technique to guide audio generation by providing the model with additional information, such as an instrument class, descriptive text, or reference melody.
Score Matching for Audio
Alternative training method where the model learns to estimate the gradient (the score) of the log probability distribution of audio data with respect to the noisy input.
Neural Vocoder
Neural network that converts an acoustic representation, such as a Mel spectrogram generated by a diffusion model, into a final audible audio waveform.
Latent Audio Diffusion
Approach where the diffusion process occurs in a compressed latent space of the audio, rather than directly on the signal or spectrogram, reducing computational costs.
Audio Fine-tuning
Process of adapting a pre-trained audio diffusion model on a specific dataset, such as a particular speaker's voice or a musical style, to specialize its generation.
Stable Audio
Latent diffusion point audio model capable of generating high-fidelity and long-duration audio samples conditioned by text.
AudioLDM
Family of audio diffusion models that use pre-trained text embeddings (such as CLAP) to condition the generation of sounds, music, or speech from text descriptions.
Waveform Diffusion
Variant of diffusion models that operates directly on the raw audio waveform in the time domain, thus avoiding information loss associated with spectrogram transformation.