diff --git a/scripts/experiments/003_experimental_ringdown_fft.py b/scripts/experiments/003_experimental_ringdown_fft.py
index 237a261..db1c853 100644
--- a/scripts/experiments/003_experimental_ringdown_fft.py
+++ b/scripts/experiments/003_experimental_ringdown_fft.py
@@ -11,102 +11,104 @@ import networkx as nx
 from functools import reduce
 
 # %% load data
-path = "../../data/22_05_24/ringdown_with_hybridized_modes"
+path = "../../data/22_05_24/ringdown_try_2"
 scan = ScanData.from_dir(path)
 
 
-# %% Fourier
-
-# window = (0.027751370026589985, 0.027751370026589985 + 0.00001 / 2)
+# %% Set Window
+window = (0.027751370026589985, 0.027751370026589985 + 0.00001 / 2)
 window = tuple(
-    np.array([0.03075207891902308, 0.03075207891902308 + 0.00001]) - 1e-3 - 0.82e-6
-)
-freq, fft = fourier_transform(
-    scan.time, scan.output, window=window, low_cutoff=20e5, high_cutoff=100e6
+    np.array([0.03075207891902308, 0.03075207891902308 + 0.00001]) + 4e-3 - 0.87e-6
 )
 
+window = tuple(
+    np.array([0.016244684251065847, 0.016248626903395593 + 49e-5]) + 8e-3 - 12e-7
+)
 
-# %% plot
+# %% Plot Scan
 gc.collect()
-fig = plt.figure("interactive")
+fig = plt.figure("interactive", constrained_layout=True)
 fig.clf()
-ax, ax2 = fig.subplots(1, 2)
+(ax, ax2, ax_signal, ax_stft, ax_decay) = fig.subplot_mosaic("AB;CC;DE").values()
+ax.set_title("Fourier Spectrum")
+ax2.set_title("Reconstructed Spectrum")
+for spec_ax in [ax, ax2]:
+    spec_ax.set_xlabel("Frequency (MHz)")
+    spec_ax.set_ylabel("Power")
+ax3 = ax.twinx()
+ax3.set_ylabel("Phase (rad)")
+
+ax_stft.set_xlabel("Time (s)")
+ax_stft.set_ylabel("Frequency (Hz)")
+ax_stft.set_title("Short Time Fourier Transform")
+ax_decay.set_xlabel("Time (s)")
+ax_decay.set_ylabel("Power")
+
+# ax_signal.set_xlim(*window)
+plot_scan(scan, ax=ax_signal, smoothe_output=1e-8, linewidth=0.5)
+ax_signal.axvspan(*window, color="red", alpha=0.1)
+ax_signal.set_xlabel("Time (s)")
+ax_signal.set_ylabel("Signal (mV)")
+ax_signal.set_title("Raw Signal (Slighly Smoothened)")
+
+# %% Fourier
+freq, fft = fourier_transform(
+    scan.time, scan.output, window=window, low_cutoff=0.5e6, high_cutoff=90e6
+)
+
+freq *= 1e-6
+
 # ax.set_yscale("log")
 ax.plot(freq, np.abs(fft))
-ax3 = ax.twinx()
-ax.plot(freq, np.abs(fft.real), linewidth=1, color="red")
-
+# ax.plot(freq, np.abs(fft.real), linewidth=1, color="red")
 # ax.plot(freq, fft.imag, linewidth=1, color="green")
-# ax3.plot(
-#     freq,
-#     np.gradient(np.unwrap(np.angle(fft) + np.pi, period=2 * np.pi)),
-#     linestyle="--",
-# )
 
-ax2.set_xlim(*window)
-
-
-# plot_scan(scan, ax=ax2, linewidth=0.5, smoothe_output=1e-8)
-
-freq_step = freq[1] - freq[0]
-
-peaks, peak_info = scipy.signal.find_peaks(
-    np.abs(fft) ** 2, distance=2e6 / freq_step, prominence=3000
+ax3.plot(
+    freq[1:],
+    np.cumsum(np.angle(fft[1:] / fft[:-1])),
+    linestyle="--",
+    alpha=0.5,
+    linewidth=0.5,
+    zorder=10,
 )
 
+freq_step = freq[1] - freq[0]
+Ω_guess = 13
+δ_guess = 2.6
+
+peaks, peak_info = scipy.signal.find_peaks(
+    np.abs(fft) ** 2, distance=δ_guess / 2 / freq_step, prominence=1e-8
+)
+
+peak_freq = freq[peaks]
 anglegrad = np.gradient(np.unwrap(np.angle(fft) + np.pi, period=2 * np.pi))
 neg_peaks = peaks[anglegrad[peaks] < 0]
 pos_peaks = peaks[anglegrad[peaks] > 0]
-ax.plot(freq[neg_peaks], np.abs(fft[neg_peaks]), "x")
-ax.plot(freq[pos_peaks], np.abs(fft[pos_peaks]), "o")
+phase_detuning = np.angle(fft[peaks])
 
-# phase_detuning = np.angle(fft[peaks])
+ax.plot(peak_freq, np.abs(fft[peaks]), "*")
 
 
-def extract_peak(index, width, sign=1):
-    return sign * freq[index - width : index + width], np.abs(
-        fft[index - width : index + width]
+def extract_peak(index, width, sign, detuning):
+    begin = max(index - width, 0)
+    return sign * (freq[begin : index + width]) + detuning, np.abs(
+        fft[begin : index + width]
     )
 
 
-# for peak in neg_peaks:
-#     ax2.plot(*extract_peak(peak, 10, -1))
-# for peak in pos_peaks:
-#     ax2.plot(*extract_peak(peak, 10, 1))
-
-
-Ω_guess = 13e6
-δ_guess = 2.69e6
-
-N = np.arange(1, 3)
-possibilieties = np.concatenate([[2 * δ_guess], N * Ω_guess, N * Ω_guess - δ_guess])
-
-
-abs(
-    np.array(
-        [Ω_guess, 2 * δ_guess, Ω_guess - δ_guess, 2 * Ω_guess, 2 * Ω_guess - δ_guess]
-    )
-)
-
 mode_freqs = freq[peaks]
 
-final_freqs = [mode_freqs[0]]
-
-
-all_diffs = np.abs(
-    np.abs(mode_freqs[:, None] - mode_freqs[None, :])[:, :, None]
-    - possibilieties[None, None, :]
-)
+all_diffs = np.abs((mode_freqs[:, None] - mode_freqs[None, :])[:, :, None] - Ω_guess)
 
 all_diffs[all_diffs == 0] = np.inf
-all_diffs[all_diffs > δ_guess / 2] = np.inf
+all_diffs[all_diffs > 1] = np.inf
 matches = np.asarray(all_diffs < np.inf).nonzero()
 
 pairs = np.array(list(zip(*matches, all_diffs[matches])), dtype=int)
 
 relationships = nx.DiGraph()
 for node, peak in enumerate(peaks):
-    relationships.add_node(node, freq=freq[peak])
+    relationships.add_node(node, freqency=freq[peak])
 
 for left, right, relationship, diff in pairs:
     if freq[left] > freq[right]:
@@ -121,31 +123,90 @@ for left, right, relationship, diff in pairs:
         )
 
 
-pos = {}
-for node, peak in enumerate(peaks):
-    pos[node] = [freq[peak], abs(fft[peak])]
-# nx.draw(relationships, pos, with_labels=True)
-# nx.draw_networkx_edge_labels(relationships, pos)
+UG = relationships.to_undirected()
 
-# %%
-cycle = nx.find_cycle(relationships, orientation="ignore")
+# extract subgraphs
+neg, pos, *unmatched = [
+    list(sorted(i))
+    for i in sorted(list(nx.connected_components(UG)), key=lambda l: -len(l))
+]
 
-total = 0
-for s, t, direction in cycle:
-    difference = possibilieties[relationships[s][t]["type"]]
-    if direction == "reverse":
-        difference *= -1
-    total += difference
+ax.plot(mode_freqs[neg], np.abs(fft[peaks[neg]]), "x")
+ax.plot(mode_freqs[pos], np.abs(fft[peaks[pos]]), "o")
 
-# %%
-relationships.remove_nodes_from(list(nx.isolates(relationships)))
+# ax.plot(freq[pos_peaks], np.abs(fft[pos_peaks]), "o")
 
-spectrum = list(
-    sorted(nx.all_simple_paths(relationships, 1, 9), key=lambda x: -len(x))
-)[0]
+Ω = (np.diff(peak_freq[neg]).mean() + np.diff(peak_freq[pos]).mean()) / 2
+ΔΩ = np.sqrt((np.diff(peak_freq[neg]).var() + np.diff(peak_freq[pos]).var())) / 2
 
-for s, t in zip(spectrum[:-1], spectrum[1:]):
-    print(s, relationships[s][t])
 
-for node in spectrum:
-    plt.plot(freq[node], abs(fft[node]), "*")
+Δ_L = ((mode_freqs[pos] - mode_freqs[neg] - Ω) / 2).mean()
+
+
+ax2.cla()
+for peak in neg:
+    ax2.plot(*extract_peak(peaks[peak], 200, 1, Δ_L + Ω / 2), color="blue")
+for peak in pos:
+    ax2.plot(*extract_peak(peaks[peak], 200, -1, Δ_L - Ω / 2), color="blue")
+
+hybrid = []
+for peak, sign in zip(np.array(unmatched).flatten(), [1, -1]):
+    hybrid.append(sign * mode_freqs[peak] + Δ_L)
+    ax2.plot(*extract_peak(peaks[peak], 200, sign, Δ_L), color="green")
+
+δ = np.abs(np.diff(hybrid)[0] / 2)
+
+fig.suptitle(f"Ω = {Ω:.2f}MHz, ΔΩ = {ΔΩ:.2f}MHz, Δ_L = {Δ_L:.2f}MHz, δ = {δ:.2f}MHz")
+
+# %% Windowed Fourier
+windows = np.linspace(window[0], window[0] + (window[-1] - window[0]) * 0.1, 100)
+fiducial = peak_freq[neg[1]]
+size = int(300 * 1e-6 / fiducial / scan.timestep)
+w_fun = scipy.signal.windows.gaussian(size, std=0.1 * size, sym=True)
+# w_fun = scipy.signal.windows.boxcar(size)
+amps = []
+SFT = scipy.signal.ShortTimeFFT(
+    w_fun, hop=int(size * 0.1 / 5), fs=1 / scan.timestep, scale_to="magnitude"
+)
+
+t = scan.time[(window[1] > scan.time) & (scan.time > window[0])]
+ft = SFT.spectrogram(scan.output[(window[1] > scan.time) & (scan.time > window[0])])
+ft[ft > 1e-2] = 0
+ax_stft.imshow(
+    np.log((ft[:, :400])),
+    aspect="auto",
+    origin="lower",
+    cmap="magma",
+    extent=SFT.extent(len(t)),
+)
+ax_stft.set_ylim(0, 50 * 1e6)
+ax_stft.set_xlim(
+    2.8 * SFT.lower_border_end[0] * SFT.T, SFT.upper_border_begin(len(t))[0] * SFT.T
+)
+
+# %% Decay Plot
+index = np.argmin(np.abs(SFT.f - 1e6 * peak_freq[unmatched[0]])) + 1
+ax_stft.axhline(SFT.f[index])
+
+hy_mode = np.mean(ft[index - 3 : index + 3, :], axis=0)
+sft_t = SFT.t(len(t))
+
+mask = (sft_t > 1 * SFT.lower_border_end[0] * SFT.T) & (sft_t < np.max(sft_t) * 0.1)
+hy_mode = hy_mode[mask]
+sft_t = sft_t[mask]
+
+ax_decay.plot(sft_t, hy_mode)
+ax_decay.set_xscale("log")
+# plt.plot(sft_t, 3e-6 * np.exp(-.9e6 * (sft_t - 3*SFT.lower_border_end[0] * SFT.T)))
+
+
+def model(t, a, τ):
+    return a * np.exp(-τ * (t - SFT.lower_border_end[0] * SFT.T))
+
+
+p, cov = scipy.optimize.curve_fit(model, sft_t, hy_mode, p0=[hy_mode[0], 1e6])
+ax_decay.plot(sft_t, model(sft_t, *p))
+print(p[1] * 1e-6, np.sqrt(np.diag(cov))[1] * 1e-6)
+ax_decay.set_title(f"A Site decay γ = {p[1] * 1e-6:.2f}MHz")
+
+fig.savefig("/tmp/screen.png", dpi=500)
diff --git a/src/rabifun/analysis.py b/src/rabifun/analysis.py
index eb2b717..96fd4d1 100644
--- a/src/rabifun/analysis.py
+++ b/src/rabifun/analysis.py
@@ -1,5 +1,6 @@
 import numpy as np
 import scipy
+import os
 
 
 def fourier_transform(
@@ -23,8 +24,10 @@ def fourier_transform(
         t = t[mask]
         signal = signal[mask]  # * scipy.signal.windows.hamming(len(t))
 
-    freq = np.fft.rfftfreq(len(t), t[2] - t[1])
-    fft = np.fft.rfft(signal)
+        #
+
+    freq = scipy.fft.rfftfreq(len(t), t[2] - t[1])
+    fft = scipy.fft.rfft(signal, norm="forward", workers=os.cpu_count())
 
     mask = (freq > low_cutoff) & (freq < high_cutoff)
     return freq[mask], fft[mask]