nexuml_library.layers.feature.lmbe¶
nexuml_library.layers.feature.lmbe
¶
Log Mel Band Energy (LMBE) feature extractor layer.
LMBE
¶
Bases: LayerDefinition
Compute Log Mel Band Energies from a raw waveform.
Uses torchaudio's MelSpectrogram + AmplitudeToDB pipeline. Optionally falls back to librosa for compatibility.
Attributes:
| Name | Type | Description |
|---|---|---|
n_mels |
int
|
Number of Mel filterbanks. |
n_fft |
int
|
FFT size. |
hop_length |
int
|
Number of samples between frames. |
win_length |
int | None
|
Window length in samples. |
power |
int
|
Spectrogram power (1=magnitude, 2=power). |
fmin |
int
|
Minimum frequency. |
fmax |
int
|
Maximum frequency. |
sr |
int
|
Sample rate. |
mel_scale |
str
|
Mel scale type ("slaney" or "htk"). |
pad_mode |
str
|
Padding mode for STFT. |
use_librosa |
bool
|
Use librosa backend instead of torchaudio. |