| """Light spectral-subtraction post filter for the FastWave output hiss (numpy). |
| |
| The diffusion model leaves a broadband noise floor (mostly above the input band). |
| We estimate the noise magnitude per frequency from the quietest frames of the |
| output itself and apply a smoothed Wiener-like gain. |
| """ |
| import numpy as np |
|
|
| from .audio import hann |
|
|
|
|
| def denoise(x, sr=48000, n_fft=1024, hop=256, alpha=1.2, floor=0.10, quiet_percentile=20, |
| smooth=3): |
| x = np.asarray(x, dtype=np.float32) |
| if len(x) < n_fft: |
| return x |
| win = hann(n_fft) |
| m = 1 + (len(x) - n_fft) // hop |
| idx = np.arange(n_fft)[None, :] + hop * np.arange(m)[:, None] |
| fr = x[idx] * win |
| X = np.fft.rfft(fr, axis=1) |
| mag = np.abs(X) |
|
|
| noise = np.percentile(mag, quiet_percentile, axis=0) |
| gain = 1.0 - alpha * noise[None, :] / np.maximum(mag, 1e-8) |
| gain = np.clip(gain, floor, 1.0) |
| if smooth > 1 and m >= smooth: |
| k = np.ones(smooth) / smooth |
| gain = np.apply_along_axis(lambda c: np.convolve(c, k, mode="same"), 0, gain) |
|
|
| Y = X * gain |
| frames = np.fft.irfft(Y, n=n_fft, axis=1) * win |
|
|
| out_len = (m - 1) * hop + n_fft |
| acc = np.zeros(out_len, dtype=np.float64) |
| wsum = np.zeros(out_len, dtype=np.float64) |
| for i in range(m): |
| acc[i * hop: i * hop + n_fft] += frames[i] |
| wsum[i * hop: i * hop + n_fft] += win ** 2 |
| acc /= np.maximum(wsum, 1e-8) |
| return acc[: len(x)].astype(np.float32) |
|
|
