| """Salience -> f0 for the exported pitch models (numpy only). |
| |
| CREPE and RMVPE share one 360-bin grid: 20 cents per bin from 1997.38 cents (32.70 Hz) re 10 Hz. |
| FCPE spaces its 360 bins evenly between 32.7 and 1975.5 Hz in cents (~19.8 cents per bin). |
| Decoding is the models' own: a salience-weighted mean of the 9 bins around the peak, voiced when |
| the peak clears a threshold. |
| """ |
| import numpy as np |
|
|
| CREPE_CENTS = 1997.3794084376191 + 20.0 * np.arange(360) |
| FCPE_CENTS = np.linspace(1200 * np.log2(32.7 / 10), 1200 * np.log2(1975.5 / 10), 360) |
| GENERIC_CENTS = (100.0 / 3) * np.arange(384) |
|
|
| CENTS = {"crepe": CREPE_CENTS, "rmvpe": CREPE_CENTS, "fcpe": FCPE_CENTS, "pesto-act": GENERIC_CENTS} |
| |
| |
| THRESHOLD = {"crepe": 0.21, "rmvpe": 0.03, "fcpe": 0.006, "pesto-act": 0.0} |
|
|
|
|
| def local_average_cents(sal: np.ndarray, cents: np.ndarray, width: int = 4) -> np.ndarray: |
| """sal [T, K] -> cents [T]: salience-weighted mean of the bins within `width` of the peak.""" |
| t, k = sal.shape |
| peak = sal.argmax(1) |
| idx = np.clip(peak[:, None] + np.arange(-width, width + 1)[None, :], 0, k - 1) |
| w = np.take_along_axis(sal, idx, 1) |
| w = np.maximum(w, 0) |
| return (w * cents[idx]).sum(1) / np.maximum(w.sum(1), 1e-12) |
|
|
|
|
| def decode(sal: np.ndarray, kind: str, threshold: float | None = None): |
| """sal [1, T, K] or [T, K] -> (f0 Hz [T], voiced bool [T]). Unvoiced frames get f0 0.""" |
| s = np.asarray(sal, np.float64) |
| if s.ndim == 3: |
| s = s[0] |
| th = THRESHOLD[kind] if threshold is None else threshold |
| cents = local_average_cents(s, CENTS[kind]) |
| voiced = s.max(1) > th |
| f0 = np.where(voiced, 10.0 * 2 ** (cents / 1200.0), 0.0) |
| return f0, voiced |
|
|
