trim.py

""" a simple gmm-based speech trimmer

Copyright © 2022 Brent M. Spell

Licensed under the MIT License (the "License"). You may not use this package 
except in compliance with the License. You may obtain a copy of the License at

https://opensource.org/licenses/MIT

Unless required by applicable law or agreed to in writing, software distributed 
under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR 
CONDITIONS OF ANY KIND, either express or implied. See the License for the 
specific language governing permissions and limitations under the License.

"""

import numpy as np
import torch

def trim(
    x, 
    fs, 
    block_length = 0.025, 
    block_overlap=0.6, 
    ref_db=-18.0,
    signal_estimate=-20.0, 
    noise_estimate=-60.0, 
    pad_length=0.25,
    edge_only=False,
):
    # frame the signal into overlapping windows and apply a hann window function
    frame_length = int(block_length * fs)
    frame_stride = int((1 - block_overlap) * frame_length)
    window = np.hanning(frame_length)
    
    frames = x.unfold(-1, size=frame_length, step=frame_stride)
    frames = frames * torch.from_numpy(window).to(frames.device)
    
    # compute the peak RMS energy of the frames and calculate normalized power
    peak_energy = frames.square().mean(-1).sqrt().max()
    ref_energy = 10.0 ** (ref_db / 20.0)
    gain = torch.nan_to_num(ref_energy / peak_energy)

    energy = (frames * gain).square().mean(-1).sqrt()
    power = 20.0 * (energy + 1e-5).log10()

    # fit the GMM to the power vector and extract the modes
    gmm = sklearn.mixture.GaussianMixture(
        n_components=2,
        init_params="random",
        means_init=np.array([noise_estimate, signal_estimate])[..., np.newaxis],
    ).fit(power.unsqueeze(-1).cpu().numpy())

    modes = gmm.means_[..., 0]

    # use the midpoint of the modes as the cutoff
    noise, signal = sorted(modes)
    cutoff = modes.mean()
    keep = power > cutoff

    if signal > -10:
        return x

    # apply a smoothing filter to the keep vector
    pad_frames = int(pad_length * fs / frame_stride) * 2
    pad_frames += 1 - pad_frames % 2
    pad_weight = 1.0 / pad_frames
    pad_filter = torch.full([1, 1, pad_frames], pad_weight).to(keep.device)
    keep = torch.nn.functional.conv1d(
        keep.type(torch.float32).view(1, 1, -1), 
        pad_filter, 
        padding="same",
    ).flatten() >= pad_weight
    
    # if we are only trimming edges, ignore interior silence
    if edge_only:
        start = keep.int().argmax()
        stop = len(keep) - keep.flip(-1).int().argmax()
        keep[start:stop] = True
    
    # discard any trimmed frames
    trimmed = frames[keep]

    # create a synthesis window for the overlap-and-add operation
    denom = window ** 2.0
    overlap = -(-frame_length // frame_stride)
    denom = np.pad(denom, [(0, overlap * frame_stride - frame_length)])
    denom = denom.reshape([overlap, frame_stride])
    denom = denom.sum(0, keepdims=True)
    denom = np.tile(denom, [overlap, 1])
    denom = denom.reshape([overlap * frame_stride])[:frame_length]
    window /= denom
    
    trimmed = trimmed * torch.from_numpy(window).to(trimmed.device)

    # reconstruct the trimmed audio signal
    y = torch.nn.functional.fold(
        trimmed.transpose(0, 1).unsqueeze(0), 
        output_size=(1, (trimmed.shape[0] - 1) * frame_stride + frame_length), 
        kernel_size=(1, frame_length), 
        stride=frame_stride,
    ).flatten()
    
    return y