AdaLN
An adaptive layer normalization technique that dynamically adjusts normalization parameters according to conditioning information. It is commonly used in generative models to strengthen conditional control.
Timeline (last 90 days)
SteerTTS replaces all LayerNorms with adaptive versions (AdaLN) and uses a masking mechanism to distinguish "unspecified" from "explicitly set to zero": when a control item is omitted the model infers a reasonable value from timbre, labels and text, while passing {"rate": 0} explicitly requests that timbre's neutral speaking rate.
Audio conditioning techniques typically rely on pretrained audio encoders (such as CLAP or EnCodec) to compress reference audio into embedding vectors, which are then injected into the generative network via cross-attention or AdaLN mechanisms.
The Transformer in diffusion language models typically employs bidirectional attention and needs to receive timestep information (injected via sinusoidal encoding or AdaLN), so standard pre-trained language model weights cannot be directly reused.
All Facts (3)
SteerTTS replaces all LayerNorms with adaptive versions (AdaLN) and uses a masking mechanism to distinguish "unspecified" from "explicitly set to zero": when a control item is omitted the model infers a reasonable value from timbre, labels and text, while passing {"rate": 0} explicitly requests that timbre's neutral speaking rate.
50%UnverifiedAudio conditioning techniques typically rely on pretrained audio encoders (such as CLAP or EnCodec) to compress reference audio into embedding vectors, which are then injected into the generative network via cross-attention or AdaLN mechanisms.
50%UnverifiedThe Transformer in diffusion language models typically employs bidirectional attention and needs to receive timestep information (injected via sinusoidal encoding or AdaLN), so standard pre-trained language model weights cannot be directly reused.
50%