toolbox/torchaudio/models/dfnet3/features.py

#!/usr/bin/python3
# -*- coding: utf-8 -*-
import math

import numpy as np


def freq2erb(freq_hz: float) -> float:
    """
    https://www.cnblogs.com/LXP-Never/p/16011229.html
    1 / (24.7 * 9.265) = 0.00436976
    """
    return 9.265 * math.log(freq_hz / (24.7 * 9.265) + 1)


def erb2freq(n_erb: float) -> float:
    return 24.7 * 9.265 * (math.exp(n_erb / 9.265) - 1)


def get_erb_widths(sample_rate: int, fft_size: int, erb_bins: int, min_freq_bins_for_erb: int) -> np.ndarray:
    """
    https://github.com/Rikorose/DeepFilterNet/blob/main/libDF/src/lib.rs
    :param sample_rate:
    :param fft_size:
    :param erb_bins: erb (Equivalent Rectangular Bandwidth) 等效矩形带宽的通道数.
    :param min_freq_bins_for_erb: Minimum number of frequency bands per erb band
    :return:
    """
    nyq_freq = sample_rate / 2.
    freq_width: float = sample_rate / fft_size

    min_erb: float = freq2erb(0.)
    max_erb: float = freq2erb(nyq_freq)

    erb = [0] * erb_bins
    step = (max_erb - min_erb) / erb_bins

    prev_freq_bin = 0
    freq_over = 0
    for i in range(1, erb_bins + 1):
        f = erb2freq(min_erb + i * step)
        freq_bin = int(round(f / freq_width))
        freq_bins = freq_bin - prev_freq_bin - freq_over

        if freq_bins < min_freq_bins_for_erb:
            freq_over = min_freq_bins_for_erb - freq_bins
            freq_bins = min_freq_bins_for_erb
        else:
            freq_over = 0
        erb[i - 1] = freq_bins
        prev_freq_bin = freq_bin

    erb[erb_bins - 1] += 1
    too_large = sum(erb) - (fft_size / 2 + 1)
    if too_large > 0:
        erb[erb_bins - 1] -= too_large
    return np.array(erb, dtype=np.uint64)


def get_erb_filter_bank(erb_widths: np.ndarray,
                        sample_rate: int,
                        normalized: bool = True,
                        inverse: bool = False,
                        ):
    num_freq_bins = int(np.sum(erb_widths))
    num_erb_bins = len(erb_widths)

    fb: np.ndarray = np.zeros(shape=(num_freq_bins, num_erb_bins))

    points = np.cumsum([0] + erb_widths.tolist()).astype(int)[:-1]
    for i, (b, w) in enumerate(zip(points.tolist(), erb_widths.tolist())):
        fb[b: b + w, i] = 1

    if inverse:
        fb = fb.T
        if not normalized:
            fb /= np.sum(fb, axis=1, keepdims=True)
    else:
        if normalized:
            fb /= np.sum(fb, axis=0)
    return fb


def spec2erb(spec: np.ndarray, erb_fb: np.ndarray, db: bool = True):
    """
    ERB filterbank and transform to decibel scale.

    :param spec: Spectrum of shape [B, C, T, F].
    :param erb_fb: ERB filterbank array of shape [B] containing the ERB widths,
            where B are the number of ERB bins.
    :param db: Whether to transform the output into decibel scale. Defaults to `True`.
    :return:
    """
    # complex spec to power spec. (real * real + image * image)
    spec_ = np.abs(spec) ** 2

    # spec to erb feature.
    erb_feat = np.matmul(spec_, erb_fb)

    if db:
        erb_feat = 10 * np.log10(erb_feat + 1e-10)

    erb_feat = np.array(erb_feat, dtype=np.float32)
    return erb_feat


def _calculate_norm_alpha(sample_rate: int, hop_size: int, tau: float):
    """Exponential decay factor alpha for a given tau (decay window size [s])."""
    dt = hop_size / sample_rate
    result = math.exp(-dt / tau)
    return result


def get_norm_alpha(sample_rate: int, hop_size: int, norm_tau: float) -> float:
    a_ = _calculate_norm_alpha(sample_rate=sample_rate, hop_size=hop_size, tau=norm_tau)

    precision = 3
    a = 1.0
    while a >= 1.0:
        a = round(a_, precision)
        precision += 1

    return a


MEAN_NORM_INIT = [-60., -90.]


def make_erb_norm_state(erb_bins: int, channels: int) -> np.ndarray:
    state = np.linspace(MEAN_NORM_INIT[0], MEAN_NORM_INIT[1], erb_bins)
    state = np.expand_dims(state, axis=0)
    state = np.repeat(state, channels, axis=0)

    # state shape: (audio_channels, erb_bins)
    return state


def erb_normalize(erb_feat: np.ndarray, alpha: float, state: np.ndarray = None):
    erb_feat = np.copy(erb_feat)
    batch_size, time_steps, erb_bins = erb_feat.shape

    if state is None:
        state = make_erb_norm_state(erb_bins, erb_feat.shape[0])
        # state = np.linspace(MEAN_NORM_INIT[0], MEAN_NORM_INIT[1], erb_bins)
        # state = np.expand_dims(state, axis=0)
        # state = np.repeat(state, erb_feat.shape[0], axis=0)

    for i in range(batch_size):
        for j in range(time_steps):
            for k in range(erb_bins):
                x = erb_feat[i][j][k]
                s = state[i][k]

                state[i][k] = x * (1. - alpha) + s * alpha
                erb_feat[i][j][k] -= state[i][k]
                erb_feat[i][j][k] /= 40.

    return erb_feat


UNIT_NORM_INIT = [0.001, 0.0001]


def make_spec_norm_state(df_bins: int, channels: int) -> np.ndarray:
    state = np.linspace(UNIT_NORM_INIT[0], UNIT_NORM_INIT[1], df_bins)
    state = np.expand_dims(state, axis=0)
    state = np.repeat(state, channels, axis=0)

    # state shape: (audio_channels, df_bins)
    return state


def spec_normalize(spec_feat: np.ndarray, alpha: float, state: np.ndarray = None):
    spec_feat = np.copy(spec_feat)
    batch_size, time_steps, df_bins = spec_feat.shape

    if state is None:
        state = make_spec_norm_state(df_bins, spec_feat.shape[0])

    for i in range(batch_size):
        for j in range(time_steps):
            for k in range(df_bins):
                x = spec_feat[i][j][k]
                s = state[i][k]

                state[i][k] = np.abs(x) * (1. - alpha) + s * alpha
                spec_feat[i][j][k] /= np.sqrt(state[i][k])
    return spec_feat


if __name__ == '__main__':
    pass