vad/transforms.py

import numpy as np
import math

# Constants
P_REF = -93.9794


class Transform:
    def __init__(self, window_size, frames_per_second):
        self.window_size = window_size
        self.frames_per_second = frames_per_second

        # Find the next power of 2
        pow2_size = 1
        while pow2_size < window_size:
            pow2_size <<= 1
        self.points = pow2_size

        # Initialize arrays
        self.real = np.zeros(pow2_size, dtype=float)
        self.imaginary = np.zeros(pow2_size, dtype=float)
        self.power = np.zeros(pow2_size, dtype=float)
        self.sine = np.zeros(pow2_size // 2, dtype=float)
        self.cosine = np.zeros(pow2_size // 2, dtype=float)
        self.dbpower_buffer = np.zeros(frames_per_second, dtype=float)
        self.dbpower = 0

        # Precompute twiddle factors
        for i in range(pow2_size // 2):
            arg = (-2.0 * math.pi * i) / pow2_size
            self.cosine[i] = math.cos(arg)
            self.sine[i] = math.sin(arg)

        # Create Hanning window
        self.window = np.zeros(pow2_size, dtype=float)
        for i in range(window_size):
            # Hanning window
            self.window[i] = (
                1.0 - math.cos(2.0 * math.pi * (i + 1) / (window_size + 1))
            ) * 0.5


def new_transform(window_size, frames_per_second):
    """Create a new Transform object"""
    return Transform(window_size, frames_per_second)


def forward_fft(fft, real_input):
    """Perform forward FFT calculation"""
    k = fft.points
    fft.total_power = 0

    # Reset arrays
    fft.real = np.zeros(k, dtype=float)
    fft.imaginary = np.zeros(k, dtype=float)

    # Apply window function to input
    for i in range(fft.window_size):
        fft.real[i] = real_input[i] * fft.window[i]

    j = 0
    m = k // 2

    # Bit reversal
    for i in range(1, k - 1):
        L = m

        while j >= L:
            j = j - L
            L = L // 2

        j = j + L

        if i < j:
            temp_real = fft.real[i]
            temp_imaginary = fft.imaginary[i]
            fft.real[i] = fft.real[j]
            fft.imaginary[i] = fft.imaginary[j]
            fft.real[j] = temp_real
            fft.imaginary[j] = temp_imaginary

    L = 0
    m = 1
    n = k // 2

    # Computation
    i = k
    while i > 1:
        L = m
        m = 2 * m
        o = 0

        for j in range(L):
            cos = fft.cosine[o]
            sin = fft.sine[o]
            o = o + n

            for p in range(j, k, m):
                q = p + L

                xt = cos * fft.real[q] - sin * fft.imaginary[q]
                yt = sin * fft.real[q] + cos * fft.imaginary[q]
                fft.real[q] = fft.real[p] - xt
                fft.real[p] = fft.real[p] + xt
                fft.imaginary[q] = fft.imaginary[p] - yt
                fft.imaginary[p] = fft.imaginary[p] + yt

        n = n >> 1
        i = i >> 1

    # Calculate power spectrum
    fft.power = np.zeros(k, dtype=float)
    for i in range(k):
        fft.power[i] = math.sqrt(
            fft.real[i] * fft.real[i] + fft.imaginary[i] * fft.imaginary[i]
        )
        fft.total_power += fft.power[i] / k

    # Calculate dB SPL
    fft.dBSPL = 10 * math.log10(fft.total_power + 1e-6) - P_REF
    temp = fft.dBSPL

    # Update running average
    fft.dbpower = fft.dbpower + (temp - fft.dbpower_buffer[0]) / fft.frames_per_second
    fft.dbpower_buffer = np.roll(fft.dbpower_buffer, -1)
    fft.dbpower_buffer[-1] = temp


def inverse_fft(fft):
    """Perform inverse FFT calculation"""
    k = fft.points

    j = 0
    m = k // 2

    # Bit reversal
    for i in range(1, k - 1):
        L = m

        while j >= L:
            j = j - L
            L = L // 2

        j = j + L

        if i < j:
            temp_real = fft.real[i]
            temp_imaginary = fft.imaginary[i]
            fft.real[i] = fft.real[j]
            fft.imaginary[i] = fft.imaginary[j]
            fft.real[j] = temp_real
            fft.imaginary[j] = temp_imaginary

    L = 0
    m = 1
    n = k // 2

    # Computation (note negative sine for inverse)
    i = k
    while i > 1:
        L = m
        m = 2 * m
        o = 0

        for j in range(L):
            cos = fft.cosine[o]
            sin = -fft.sine[o]  # Negative for inverse
            o = o + n

            for p in range(j, k, m):
                q = p + L

                xt = cos * fft.real[q] - sin * fft.imaginary[q]
                yt = sin * fft.real[q] + cos * fft.imaginary[q]
                fft.real[q] = fft.real[p] - xt
                fft.real[p] = fft.real[p] + xt
                fft.imaginary[q] = fft.imaginary[p] - yt
                fft.imaginary[p] = fft.imaginary[p] + yt

        n = n >> 1
        i = i >> 1

    # Scale the result
    fft.real = fft.real / k


def destroy_transform(transform):
    """Clean up resources (not necessary in Python due to garbage collection)"""
    # In Python, we don't need to explicitly free memory
    # This function is included for API compatibility
    pass