decipher.featurise

Audio-to-feature extraction: mel-patch features, frequency-adaptive FFT parameters, SSL input preparation.

constant decipher.featurise.MEL_DIM

MEL_DIM: int   # default total dimension of mel_patch_feature output

Output dimension of mel_patch_feature at default parameters: flattened log-mel patch + auxiliary scalars.

function decipher.featurise.mel_dim

def mel_dim(n_mels: int = ..., patch_t: int = ...) -> int

Compute the output dimension of mel_patch_feature for non-default n_mels / patch_t.

function decipher.featurise.auto_mel_params

def auto_mel_params(
    sr: int,
    f_lo: float,
    f_hi: float,
    *,
    n_mels: int = 64,
) -> dict

Pick frequency-adaptive FFT parameters (n_fft, hop, fmin, fmax, n_mels) for a target frequency band. Use this for infrasonic blue-whale calls or very narrow-band high-frequency clicks.

function decipher.featurise.mel_patch_feature

def mel_patch_feature(
    samples: np.ndarray,
    sr: int,
    *,
    n_mels: int = ...,
    patch_t: int = ...,
    fmin: float | None = None,
    fmax: float | None = None,
    n_fft: int | None = None,
    hop: int | None = None,
) -> np.ndarray

Compute a fixed-length feature vector from a single clip: time-normalised log-mel patch (resampled to patch_t frames) flattened and concatenated with auxiliary scalars (duration, peak frequency, bandwidth).

function decipher.featurise.prepare_for_ssl

def prepare_for_ssl(
    samples: np.ndarray,
    sr: int,
    *,
    target_sr: int = 16000,
    clip_seconds: float = 1.0,
) -> np.ndarray

Resample to target_sr and center-pad/crop to a fixed-length clip. Use this to prepare units for AVESEncoder.embed_clip.