Phasmophobia Whenever I see Eberhardt Hall...I get...
Создано: 7 ноября 2026 г.
Отвечено с использованием GPT-6 Thinking от Chat01
Создано: 7 ноября 2026 г.
Отвечено с использованием GPT-6 Thinking от Chat01
Phasmophobia
Whenever I see Eberhardt Hall...I get quite the heebie jeebies.
Especially as Halloween draws near. That place just...feels like a haunted mansion. Matter of fact; I'm pretty sure it is! I went in recently, and some eerie, haunted house like music started playing. That caught me off guard; especially as I was pretty sure I heard someone...or something whispering. To make matters even worse...at the same time I think I even saw a portrait shift. It was pretty strange. Maybe. I wasn't sure. I was losing it; especially when I was certain I saw a ghost wearing a laurel from the corner of my eye." "BUT I did not think about it for even a moment. I scram skedaddled out of there. NOPE!" "Although...now that I think about it...I wonder what those whispers I heard were...
the flag format is spooky{}.
I’ll inspect the audio for hidden speech or patterns, then check whether the portrait and the laurel clue indicate a cipher. I’ll use those together to recover the spooky{...} flag.
pythonimport soundfile as sf, numpy as np, matplotlib.pyplot as plt, os p='/mnt/data/PhasmophobiaChallenge.wav' y,sr = sf.read(p) print(y.shape,sr) print('min-max',y.min(axis=0),y.max(axis=0),'RMS',np.sqrt((y*y).mean(axis=0))) print('channel correlation',np.corrcoef(y.T)) print('mono peak',np.max(np.abs(np.mean(y,axis=1)))) from IPython.display import Audio, display display(Audio(y.T, rate=sr))
pythonimport scipy.signal as sps fig,axes = plt.subplots(3,1,figsize=(14,9)) tt=np.arange(len(y))/sr idx=50 axes[0].plot(tt[::idx],y[::idx,0],alpha=.7,label='L') axes[0].plot(tt[::idx],y[::idx,1],alpha=.5,label='R') axes[0].set(xlabel='seconds',ylabel='Amplitude');axes[0].legend() for ax,x,label in zip(axes[1:],[m,d],['Mid','Side']): f,t,z=sps.stft(x,sr,nperseg=1024,noverlap=768) ax.pcolormesh(t,f,20*np.log10(np.abs(z)+1e-7),shading='auto',vmin=-100,vmax=-28,cmap='magma') ax.set_ylim(0,6000);ax.set(xlabel='seconds',ylabel='Hz',title=label) plt.tight_layout();plt.show()
pythonfig,axes=plt.subplots(2,1,figsize=(16,8)) for ax,x,label in zip(axes,[y[:,0],y[:,1]],['Left','Right']): f,t,z=sps.stft(x,sr,nperseg=2048,noverlap=1800) ax.pcolormesh(t,f,20*np.log10(np.abs(z)+1e-8),shading='auto',vmin=-110,vmax=-45,cmap='magma') ax.set(xlabel='sec',ylabel='Hz',title=label) plt.tight_layout()
pythonfrom matplotlib.colors import Normalize f,t,z=sps.stft(y[:,0],sr,nperseg=512,noverlap=460) S=20*np.log10(abs(z)+1e-9) plt.figure(figsize=(20,6)) plt.pcolormesh(t,f,S,shading='auto',vmin=-110,vmax=-58,cmap='gray_r') plt.xlim(7.2,14.8);plt.ylim(7800,11500) plt.xticks(np.arange(7,15,.25),rotation=90) plt.yticks(np.arange(8000,11501,250)) plt.grid(alpha=.1) plt.show()
pythonsel=(f>=8000)&(f<=12000);tsel=(t>=7)&(t<=15) w=S[np.ix_(sel,tsel)] print('stats:',np.percentile(w,[0,1,5,10,25,50,75,90,95,99,99.9,100])) for a,b in [(7,8),(8.5,9.5),(10,11.5),(13.2,14.7)]: v=S[np.ix_(sel,(t>=a)&(t<=b))] print(a, np.percentile(v,[50,80,90,95,99,99.9]))
pythonfig,axes=plt.subplots(3,1,figsize=(20,12)) for ax,N,contrast in zip(axes,[256,1024,4096],[(-95,-66),(-93,-65),(-102,-70)]): f2,t2,z2=sps.stft(y[:,1],sr,nperseg=N,noverlap=int(N*.93)) s2=20*np.log10(np.abs(z2)+1e-10) ax.pcolormesh(t2,f2,s2,shading='auto',vmin=contrast[0],vmax=contrast[1],cmap='magma') ax.set(xlim=(7.1,14.7),ylim=(7700,11500),title=f'STFT N={N}') ax.set_xticks(np.arange(7.25,14.75,.25), minor=True) ax.grid(which='minor',alpha=.1) plt.tight_layout();plt.show()
pythonN=512 F,T,Z=sps.stft(y[:,0],sr,nperseg=N,noverlap=N-70) DB=20*np.log10(abs(Z)+1e-9) fig,ax=plt.subplots(figsize=(24,6)) ax.pcolormesh(T,F,DB,shading='auto',vmin=-100,vmax=-70,cmap='gray_r') ax.set(xlim=(7.25,14.5),ylim=(8300,11000)) ax.set_xticks(np.arange(7.25,14.5,.25));ax.set_yticks(np.arange(8500,11001,500));ax.grid(alpha=.05) plt.tight_layout();plt.savefig('/mnt/data/spectrogram_reveal.png',dpi=180);plt.show()
The audio has a distinct high-frequency pattern around 8–11 kHz that looks like lettering in a spectrogram. I’m isolating it to read the message, and I’ll check whether the portrait supplies the Caesar-cipher shift.
pythonfrom scipy.ndimage import gaussian_filter # stft nfft=512, smooth horizontal a bit; each W shape holes fig,ax=plt.subplots(figsize=(18,5)) for fr in (8500,9000,9700,10300): idx=np.argmin(abs(F-fr)) arr=gaussian_filter(DB[idx],sigma=2) ax.plot(T,arr,label=str(round(F[idx])),alpha=.8) ax.set(xlim=(7.1,9.2),ylim=(-130,-50));ax.grid();ax.legend() <matplotlib.legend.Legend at 0x7f5a08a97120>
pythonfrom scipy.signal import butter, sosfiltfilt, hilbert x=y.mean(axis=1) sos=butter(4,[8100,11100],fs=sr,btype='bandpass',output='sos') flt=sosfiltfilt(sos,x) sub=flt*np.cos(2*np.pi*8200*tt)*2 sub=sosfiltfilt(butter(3,3000,fs=sr,output='sos'),sub) print(np.max(abs(sub)),np.sqrt((sub*sub).mean())) display(Audio(sub[int(7*sr):int(15*sr)],rate=sr))
pythonimport string s="WYVAOLIWWOOLWROLIR" for k in range(26): z="".join(chr((ord(c)-65-k)%26+65) for c in s) print(f'-{k:2d} {z}')
python# use mid signal STFT bands 8400-10500, smoothing selected=(F>=8400)&(F<=10400) arr=DB[selected,:] avg=np.mean(np.maximum(-110,arr),axis=0) p90=np.percentile(arr,90,axis=0) fig,ax=plt.subplots(figsize=(19,4)) ax.plot(T,gaussian_filter(avg,7),lw=1.5,label='average spect energy') ax.plot(T,gaussian_filter(p90,7),lw=1,label='p90 spect energy') ax.set(xlim=(7,14.6),ylim=(-130,-60)) ax.set_xticks(np.arange(7,14.7,0.1),minor=True);ax.grid(alpha=.3,which='both');ax.legend();plt.show()
pythonfor k in range(26): print(k, ''.join(chr((ord(x)-97+k)%26+65) for x in "spooky"))
The high-frequency pattern is present in both audio channels and absent from the opening music, so it appears deliberately embedded. The lettering is distorted by the soundtrack; I’m testing different spectral views before settling on a transcription.
pythonfrom scipy import signal f,P=signal.welch(x[7*sr:15*sr],sr,nperseg=16384) plt.figure(figsize=(15,4));plt.plot(f,10*np.log10(P+1e-20));plt.xlim(7000,12000);plt.ylim(-140,-50);plt.grid();plt.show()
pythonhigh = sosfiltfilt(butter(4,[8200,11100],fs=sr,btype='bandpass',output='sos'),y,axis=0) for a,b in [(7.4,7.8),(8.2,8.4),(8.4,8.7),(9.4,9.7),(11.1,11.4),(12,12.2),(12.85,13.1),(13.5,13.7)]: xx=high[int(a*sr):int(b*sr)] print(a, round(np.corrcoef(xx.T)[0,1],3),np.sqrt(np.mean((xx[:,0]-xx[:,1])**2)),np.sqrt(np.mean(((xx[:,0]+xx[:,1])/2)**2)))
pythonfig, ax=plt.subplots(figsize=(16,6)) f,t,z=sps.stft(x,sr,nperseg=2048,noverlap=1980) ax.pcolormesh(t,f,20*np.log10(abs(z)+1e-10),vmin=-92,vmax=-62,cmap='inferno',shading='auto') ax.set(xlim=(7.3,10.5),ylim=(8100,11200)) plt.show()
pythonimport numpy as np, soundfile as sf from scipy import signal from IPython.display import Audio,display x,sr=sf.read('/mnt/data/PhasmophobiaChallenge.wav') mid=x.mean(axis=1) hi=signal.sosfiltfilt(signal.butter(4,[8300,10800],fs=sr,btype='bandpass',output='sos'),mid) analytic=signal.hilbert(hi) t=np.arange(len(hi))/sr for fshift in [8000,8500,9000]: shifted=analytic*np.exp(-2j*np.pi*fshift*t) v=signal.sosfiltfilt(signal.butter(4,3200,fs=sr,output='sos'),shifted.real) print('shift',fshift) display(Audio(v[int(7.2*sr):int(14.4*sr)],rate=sr))
pythonimport soundfile as sf, numpy as np from scipy.signal import hilbert, butter, sosfiltfilt from IPython.display import Audio, display x,sr=sf.read('/mnt/data/PhasmophobiaChallenge.wav',start=round(7*44100),stop=round(15*44100)) z=x.mean(axis=1) h=sosfiltfilt(butter(4,[8300,10900],btype='bandpass',fs=sr,output='sos'),z) analytic=hilbert(h) t=np.arange(len(h))/sr v=np.real(analytic*np.exp(-2j*np.pi*8400*t)) v=sosfiltfilt(butter(3,3000,fs=sr,output='sos'),v) display(Audio(v,rate=sr))
pythondisplay(Audio(v[::-1],rate=sr))
The portrait and laurel clues still point strongly to Caesar shifting, but the recovered high-frequency shapes are not yet legible enough to trust as ciphertext. I’m checking the stereo separation and image data for a cleaner way to recover the exact characters rather than guessing a flag.
pythonimport numpy as np,soundfile as sf from scipy.signal import sosfiltfilt, butter, hilbert from IPython.display import Audio, display a,sr=sf.read('/mnt/data/PhasmophobiaChallenge.wav',start=int(7.2*44100),stop=int(14.5*44100)) z=a.mean(axis=1) h=sosfiltfilt(butter(4,[8300,10900],btype='bandpass',fs=sr,output='sos'),z) t=np.arange(len(h))/sr for shift in [10600,10800]: v=np.real(hilbert(h)*np.exp(-2j*np.pi*shift*t)) v=sosfiltfilt(butter(4,3200,fs=sr,output='sos'),v) print('fshift',shift) display(Audio(v,rate=sr))
python# playback twice: heterodyne at upper boundary, and at lower boundary; all in one note. import numpy as np, soundfile as sf from scipy.signal import butter,sosfiltfilt,hilbert from IPython.display import Audio,display dat,sr=sf.read('/mnt/data/PhasmophobiaChallenge.wav',start=round(7.2*44100),stop=round(14.4*44100)) v=dat.mean(axis=1); v=sosfiltfilt(butter(4,[8400,10500],fs=sr,btype='bandpass',output='sos'),v) an=hilbert(v); t=np.arange(len(v))/sr for f0 in (8350,10500): bb=(an*np.exp(-2j*np.pi*f0*t)).real bb=sosfiltfilt(butter(3,2600,fs=sr,output='sos'),bb) print('Frequency shift',f0) display(Audio(bb,rate=sr))
pythonimport numpy as np, soundfile as sf, scipy.signal as sig, scipy.ndimage as ndi from IPython.display import Audio, display x,sr=sf.read('/mnt/data/PhasmophobiaChallenge.wav') mono=x.mean(axis=1)[::2]; sr=sr//2 f,t,Z=sig.stft(mono,sr,nperseg=1024,noverlap=768) A=np.abs(Z) harm=ndi.median_filter(A,size=(1,17)) perc=ndi.median_filter(A,size=(17,1)) mask=(perc**2)/(perc**2+harm**2+1e-20) _,audio=sig.istft(Z*mask,sr,nperseg=1024,noverlap=768) print('HPSS audio',len(audio)/sr, np.sqrt(np.mean(audio*audio))) display(Audio(audio,rate=sr))
pythonimport soundfile as sf, scipy.signal as sig, numpy as np, matplotlib.pyplot as plt x,sr=sf.read('/mnt/data/PhasmophobiaChallenge.wav') f,t,Z=sig.stft(x[:,0],sr,nperseg=2048,noverlap=1800) for vmin,vmax in [(-145,-95),(-125,-75)]: fig,ax=plt.subplots(figsize=(14,4)) high=(f>11000)&(f<22050) a=20*np.log10(abs(Z[high])+1e-12) ax.pcolormesh(t,f[high],a,shading='auto',vmin=vmin,vmax=vmax,cmap='magma') ax.set(xlim=(0,19),ylim=(11000,22050),title=f'High frequency {vmin}/{vmax}') plt.tight_layout(); plt.show()
pythonfig,axes=plt.subplots(3,1,figsize=(15,10)) for ax,(lo,hi,vmi,vma) in zip(axes,[(0,2000,-95,-45),(2000,7000,-110,-65),(7000,11000,-100,-65)]): mask=(f>=lo)&(f<=hi) ax.pcolormesh(t,f[mask],20*np.log10(abs(Z[mask])+1e-12),shading='auto',vmin=vmi,vmax=vma,cmap='magma') ax.set(xlim=(0,19),ylim=(lo,hi),ylabel='Hz') plt.tight_layout();plt.show()
The portrait’s lowest bit planes don’t show an obvious hidden message. I’m checking whether the audio’s two channels can be separated to reveal a quieter whisper beneath the music.
pythonfrom IPython.display import display,Audio import soundfile as sf, numpy as np x,sr=sf.read('/mnt/data/isolated_diff_100.wav') display(Audio(x,rate=sr))
pythonx,sr=sf.read('/mnt/data/isolated_sum_100.wav') display(Audio(x,rate=sr))
The strongest cipher clue is the ghost wearing a laurel wreath, pointing to Julius Caesar and a Caesar-shift cipher. The portrait “shifting” reinforces that interpretation.
The WAV also contains a strong, unusual signal around 8.4–10.8 kHz, particularly between 7.3 and 14.3 seconds. That is the most promising area to investigate for the hidden whisper.
After isolating the suspicious frequencies and examining the image for steganography, I couldn't recover a reliable ciphertext. On closer inspection, part of the high-frequency pattern may be harmonics of the music rather than a hidden message.
I prepared three files for further inspection:
Flag status: Unsolved. The Caesar-cipher clue is strong, but I haven't recovered the whispered message needed to produce a verified spooky{...} flag. I don't want to give you a fabricated flag.