This is an isolated experiment derived from the running MCMamba-SSF setup. It does not modify or resume that run.
The audio path retains the four offline Bi-Mamba stages: full-band spatial, narrow-band spatial, sub-band spectral, and full-band spectral. Target DoA conditioning is changed as follows:
- The existing 180-way one-hot vector is mapped through a fixed circular Fourier basis with harmonics 1 through 4.
- A two-layer MLP encodes the resulting eight sine/cosine values.
- Separate zero-initialized FiLM heads generate scale and bias for the full-band and narrow-band spatial stages.
- FiLM is applied after LayerNorm as
(1 + scale) * normalized + bias. - Spectral stages remain unconditioned.
The circular basis makes the wrap-around relationship between -180° and
178° explicit and permits continuous/soft directional inputs. FiLM provides
stronger feature-wise conditioning than adding one repeated embedding before
normalization.
The dataset, STFT, CRM target, loss, optimizer, scheduler, Mamba dimensions, and effective global batch size are unchanged for a controlled comparison.