Dabas Sound 500 misc Author: Kobalsky 0 0 He fo...
Créé le : 20 novembre 2026
Répondu en utilisant GPT-6 Thinking par Chat01
Créé le : 20 novembre 2026
Répondu en utilisant GPT-6 Thinking par Chat01
Dabas Sound
500
misc
Author: Kobalsky
0 0
He found this stupid font n°2 on Reddit, the unambiguousnumbers just below all chars, and now he can write with blocks in Minecraft. And for 3 chars he didn't even use that font and didn't even stick to the height the others have. He turned off my screen and bet me that I couldn't figure his writing in our superflat world out, this is sekuriti! Tonto...
But at least I could still look him in the eye as he straved to the side to start writing. And then he started writing with his 52+10+3 characters. I am still confident that this is doable. Gotta proof that caveza de huevo wrong.
Flag format: gctf{^[a-zA-Z0-9_]*$}
I’ll inspect the WAV for Minecraft movement and block-placement sounds, then use the timing and the font clue to reconstruct the writing. I’ll also check what “font n°2” refers to so the letters can be decoded accurately.
pythonimport soundfile as sf, numpy as np p="/mnt/data/dabas-sound.wav" a,sr=sf.read(p,dtype='float32') a.shape,sr,(a.min(),a.max()) ((4900550, 2), 44100, (np.float32(-0.21573311), np.float32(0.21966358)))
pythonfrom IPython.display import Audio, display display(Audio(a[:12*sr].T, rate=sr))
pythonimport matplotlib.pyplot as plt from scipy.signal import find_peaks sec= np.arange(0,len(a),sr//200) step=sr//200 n=len(a)//step chunks=a[:n*step].reshape(n,step,2) rms=np.sqrt(np.mean(chunks**2,axis=1)) tp=np.arange(n)/200 fig,ax=plt.subplots(figsize=(16,3)) ax.plot(tp, rms[:,0],linewidth=.7,label='L') ax.plot(tp,rms[:,1],linewidth=.7,label='R',alpha=.6) ax.set_xlim(0,111); ax.legend(); ax.set_title('Amplitude per 5ms'); plt.show()
pythonfig,axs=plt.subplots(2,1,figsize=(16,5),sharex=True) start=0; end=8 i=(tp>=start)&(tp<end) axs[0].plot(tp[i],rms[i,0],label='L');axs[0].plot(tp[i],rms[i,1],label='R') axs[0].legend() axs[1].imshow(np.log10(0.00001+np.abs(np.fft.rfft(chunks[:1600,::5,0],axis=1)).T),extent=[0,8,0,4410],aspect='auto',origin='lower') axs[1].set_ylim(0,4000);plt.show()
pythonfrom scipy.ndimage import gaussian_filter1d env = gaussian_filter1d(rms.mean(axis=1), 3) peaks,props=find_peaks(env,distance=0.3*200,prominence=.002) len(peaks), list(zip(np.round(tp[peaks[:30]],2),np.round(env[peaks[:30]],3))), (tp[peaks[-1]]) (143, [(np.float64(0.49), np.float32(0.017)), (np.float64(1.29), np.float32(0.019)), (np.float64(1.96), np.float32(0.019)), (np.float64(2.76), np.float32(0.021)), (np.float64(3.56), np.float32(0.02)), (np.float64(4.54), np.float32(0.018)), (np.float64(5.27), np.float32(0.019)), (np.float64(5.87), np.float32(0.022)), (np.float64(7.1), np.float32(0.02)), (np.float64(7.98), np.float32(0.022)), (np.float64(8.92), np.float32(0.023)), (np.float64(9.72), np.float32(0.022)), (np.float64(10.32), np.float32(0.023)), (np.float64(10.92), np.float32(0.024)), (np.float64(11.72), np.float32(0.024)), (np.float64(12.58), np.float32(0.024)), (np.float64(13.13), np.float32(0.026)), (np.float64(14.13), np.float32(0.028)), (np.float64(14.89), np.float32(0.027)), (np.float64(15.64), np.float32(0.029)), (np.float64(16.24), np.float32(0.028)), (np.float64(16.84), np.float32(0.027)), (np.float64(17.5), np.float32(0.027)), (np.float64(18.54), np.float32(0.029)), (np.float64(19.27), np.float32(0.03)), (np.float64(20.1), np.float32(0.032)), (np.float64(20.74), np.float32(0.031)), (np.float64(21.54), np.float32(0.033)), (np.float64(22.22), np.float32(0.034)), (np.float64(23.02), np.float32(0.035))], np.float64(107.6))
python# classify 143 signal events by top ~8 spectral peaks within first 200ms from scipy.signal import stft tseg = [(float(tp[ix]-0.04), float(tp[ix]+0.25)) for ix in peaks[:15]] fig,axs=plt.subplots(3,5,figsize=(16,6)) for j,ax in enumerate(axs.ravel()): s,e=tseg[j] sig=a[max(0,int(s*sr)):min(len(a),int(e*sr)),0] f=np.fft.rfftfreq(len(sig),1/sr) ft=np.abs(np.fft.rfft(sig)) ax.plot(f,ft) ax.set_xlim(50,1500);ax.set_title(f"{tp[peaks[j]]:.2f}") plt.tight_layout()
pythonde=np.diff(env,prepend=0) pts,_=find_peaks(de,distance=int(.28*200),prominence=.002) on=tp[pts] len(on), list(zip(np.round(on[:25],3), np.round(de[pts[:25]],3))), np.unique(np.round(np.diff(on),2),return_counts=True) (143, [(np.float64(0.46), np.float64(0.002)), (np.float64(1.26), np.float64(0.002)), (np.float64(1.93), np.float64(0.002)), (np.float64(2.735), np.float64(0.003)), (np.float64(3.535), np.float64(0.003)), (np.float64(4.51), np.float64(0.002)), (np.float64(5.24), np.float64(0.002)), (np.float64(5.84), np.float64(0.003)), (np.float64(7.075), np.float64(0.003)), (np.float64(7.945), np.float64(0.003)), (np.float64(8.89), np.float64(0.003)), (np.float64(9.69), np.float64(0.003)), (np.float64(10.295), np.float64(0.003)), (np.float64(10.895), np.float64(0.003)), (np.float64(11.695), np.float64(0.003)), (np.float64(12.555), np.float64(0.003)), (np.float64(13.1), np.float64(0.003)), (np.float64(14.1), np.float64(0.004)), (np.float64(14.86), np.float64(0.004)), (np.float64(15.605), np.float64(0.004)), (np.float64(16.205), np.float64(0.003)), (np.float64(16.805), np.float64(0.003)), (np.float64(17.47), np.float64(0.003)), (np.float64(18.51), np.float64(0.004)), (np.float64(19.245), np.float64(0.004))], (array([0.44, 0.47, 0.53, 0.54, 0.55, 0.57, 0.6 , 0.61, 0.63, 0.64, 0.66, 0.67, 0.68, 0.7 , 0.71, 0.73, 0.74, 0.75, 0.76, 0.77, 0.78, 0.8 , 0.81, 0.83, 0.84, 0.86, 0.87, 0.88, 0.9 , 0.91, 0.93, 0.95, 0.96, 0.97, 1. , 1.01, 1.03, 1.04, 1.2 , 1.24, 1.4 ]), array([ 1, 1, 1, 1, 2, 1, 12, 5, 4, 5, 2, 10, 1, 17, 3, 8, 1, 1, 3, 6, 1, 21, 1, 4, 3, 1, 3, 1, 6, 2, 1, 1, 1, 1, 3, 2, 1, 1, 1, 1, 1])))
python# avg rms in 0.05 - 0.2 from onset pans=[]; absv=[] for t in on: sel=(tp>=t+0.06)&(tp<t+0.14) b=np.mean(rms[sel],axis=0) pans.append(b[1]/(b[0]+b[1])) absv.append(b) pans=np.array(pans); absv=np.array(absv) fig,axs=plt.subplots(2,1,figsize=(16,5),sharex=True) axs[0].plot(on,absv[:,0],'.-',label='left') axs[0].plot(on,absv[:,1],'.-',label='right') axs[0].legend() axs[1].plot(on,pans,'.-',label='R fraction') axs[1].set_ylim(0,1) plt.show()
pythonfor i in range(0,len(on),10): print(' '.join(f"{j:3d}:{on[j]:5.1f} {absv[j,0]:.3f},{absv[j,1]:.3f} p{pans[j]:.2f}" for j in range(i,min(i+10,len(on)))))
pythonfrom scipy.signal import correlate for j in (0,15,35,55,65,68,75,78,81,92,130): s=int((on[j]+.07)*sr);x=a[s:s+int(.06*sr)] c=np.dot(x[:,0],x[:,1])/(np.linalg.norm(x[:,0])*np.linalg.norm(x[:,1])) ratio=np.sum(x[:,1]*x[:,0])/np.sum(x[:,0]*x[:,0]) xx=np.correlate(x[::20,0],x[::20,1],"full") lag=np.argmax(xx)-(len(x[::20])-1) print(j,round(on[j],2), 'corr',round(c,3),'reg_ratio',round(ratio,2),'delay~',lag*20/sr*1000)
pythonj=0;st=int((on[j]-.015)*sr);x=a[st:st+int(.07*sr)] fig,axs=plt.subplots(2,1,figsize=(14,5)) axs[0].plot(np.arange(len(x))/sr, x[:,0],lw=.5,label='L') axs[0].plot(np.arange(len(x))/sr, x[:,1],lw=.5,label='R') axs[0].legend();axs[0].set_xlim(.005,.055) axs[1].plot(np.arange(len(x))/sr, x[:,0],lw=.5,label='L') axs[1].plot(np.arange(len(x))/sr, x[:,1],lw=.5,label='R') axs[1].set_xlim(.010,.02) plt.show()
pythonfrom scipy.signal import correlate lagtab=[] cortab=[] for j,t in enumerate(on): s=int((t+.05)*sr);xx=a[s:s+int(.22*sr),:] xx=xx[::4,:] corr=correlate(xx[:,0],xx[:,1],mode='full',method='fft') lim=round(.006*sr/4) mid=len(xx)-1 cor=corr[mid-lim:mid+lim+1] ii=np.argmax(cor) lag=(ii-lim)*4/sr*1000 lagtab.append(lag) cortab.append(np.max(cor)/np.sqrt(np.sum(xx[:,0]**2)*np.sum(xx[:,1]**2))) fig,axs=plt.subplots(2,1,figsize=(15,5),sharex=True) axs[0].plot(on,lagtab,'.-');axs[0].set_ylabel('lag ms') axs[1].plot(on,cortab,'.-');axs[1].set_ylabel('coherence') plt.show()
pythonfrom scipy.signal import find_peaks fund=[] for j,t in enumerate(on): s=int((t+.02)*sr) xx=a[s:s+int(.35*sr),0] if len(xx)<int(.3*sr): continue f=np.fft.rfftfreq(131072,1/sr) ft=np.abs(np.fft.rfft(xx*np.hanning(len(xx)),n=131072)) k=np.argmax(ft[(f>=150)&(f<300)]) fund.append(f[np.where(f>=150)[0][0]+k]) print("freq",np.round(fund,3)) print("range",min(fund),max(fund))
The recording contains 143 distinct, regularly spaced tonal events. Their loudness and left/right balance change as though the sound source is moving past the listener, and a short group uses a different pitch. That gives several measurable clues beyond the sound alone.
pythonfor j in [0,20,53,65,68,72,78,95,120,142]: t=on[j]+0.08 x=a[int(t*sr):int((t+.25)*sr),:] f=np.fft.rfftfreq(65536,1/sr) ft=np.fft.rfft(x*np.hanning(len(x))[:,None], n=65536,axis=0) vals=[] for k in range(1,6): z=ft[np.argmin(abs(f-k*220))] vals.append(round(np.angle(z[1]/z[0]),2)) print(j, "phase differences at harmonics",vals)
python# set signed based on near-zero crossover two components: sL=np.ones(len(on));sL[78:]=-1 sR=np.ones(len(on));sR[:64]=-1 v=np.column_stack([absv[:,0]*sL,absv[:,1]*sR]) fig,axs=plt.subplots(1,3,figsize=(16,5)) for ax,p in zip(axs,[0.5,1,2]): mag=np.linalg.norm(v,axis=1) pos=v/mag[:,None]**(1+1/p) ax.plot(pos[:,0],pos[:,1],'-',alpha=.25) ax.scatter(pos[:,0],pos[:,1],c=range(len(pos)),cmap='viridis',s=11) ax.set_aspect('equal');ax.set_title(f'attn 1/r^{p}') plt.tight_layout()
pythonfig,ax=plt.subplots(figsize=(8,8)) ax.scatter(v[:,0],v[:,1],c=range(len(on)),cmap='viridis',s=25) ax.plot(v[:,0],v[:,1],lw=.3,alpha=.5) ax.set_aspect('equal');ax.grid() plt.show()
python# circle through origin: u*x + w*y = x^2+y^2 X=v; yy=np.sum(v*v,axis=1) c=np.linalg.lstsq(X,yy,rcond=None)[0] res=X@c-yy print('circle center',c/2,'radius',np.linalg.norm(c)/2,'res relative',np.std(res),np.mean(np.abs(res))) fig,ax=plt.subplots(figsize=(7,7)) ax.scatter(v[:,0],v[:,1],c=range(len(v)),cmap='viridis',s=25) theta=np.linspace(0,2*np.pi,500) ax.plot(c[0]/2+np.linalg.norm(c)/2*np.cos(theta),c[1]/2+np.linalg.norm(c)/2*np.sin(theta)) ax.set_aspect('equal');plt.show()
pythonfrom scipy.optimize import minimize_scalar def align_score(q,show=False): mag=np.linalg.norm(v,axis=1) xy=(v/mag[:,None])*mag[:,None]**(-q) ss=np.linalg.svd(xy-xy.mean(axis=0),full_matrices=False,compute_uv=False) return ss[1]/ss[0] for q in np.arange(-1,4.01,.25): print(round(q,2),round(align_score(q),3),end='; ') print() qopt=minimize_scalar(align_score,bounds=(-1,10),method='bounded') qopt.x,qopt.fun (np.float64(0.5012432760175421), np.float64(0.07496427175875309))
pythonfrom statsmodels.nonparametric.smoothers_lowess import lowess xx,yy=XY.T for frac in [.12,.2,.35,.5]: ypred=lowess(yy,xx,frac=frac,return_sorted=False,it=1) fig,ax=plt.subplots(figsize=(16,2.5)) ax.scatter(xx,yy-ypred,c=np.arange(len(xx)),cmap='viridis',s=20) ax.set_title(f'detrend frac {frac}');ax.grid() ax.set_ylim(-.5,.5) plt.show()
pythonfor q in np.arange(0,2.01,.2): xy1=v/mag[:,None]**(1+q) u,s,vt=np.linalg.svd(xy1-xy1.mean(axis=0),full_matrices=False) z=(xy1-xy1.mean(axis=0))@vt.T x=z[:,0] corr=abs(np.corrcoef(x,np.arange(len(x)))[0,1]) print(round(q,2),round(corr,3),round(s[1]/s[0],3),round(np.min(x),2),round(np.max(x),2))
pythonfrom scipy.optimize import minimize def score(params, opt='pca'): q,bet=params r=mag**(-q)-bet if r.min()<=0: return 100+abs(r.min()) xy1=v/mag[:,None]*r[:,None] u,s,vt=np.linalg.svd(xy1-xy1.mean(axis=0),full_matrices=False) xx1=(xy1-xy1.mean(axis=0))@vt[0] corr=np.corrcoef(xx1,np.arange(len(xx1)))[0,1] if opt=='pca':return s[1]/s[0] if opt=='both':return s[1]/s[0] +.03*(1-corr*corr) return 1-corr*corr for q in [.3,.5,.7,1,1.2,1.5]: beta_range=(mag.min()**(-q),mag.max()**(-q)) vals=[] for B in np.linspace(-5,mag.max()**(-q)*.9,50): vals.append(score((q,B))) k=np.argmin(vals);b=np.linspace(-5,mag.max()**(-q)*.9,50)[k] print(q,b,vals[k]) for x0 in [[.5,0],[1,0],[.7,0], [1,5],[.5,2]]: r=minimize(score,x0,method='Nelder-Mead',options={'maxiter':3000}) print("opt",x0,r.x,r.fun)
pythondef metric_c(c): r= c - np.log(mag) xy1=v/mag[:,None]*r[:,None] u,s,vt=np.linalg.svd(xy1-xy1.mean(axis=0),full_matrices=False) return s[1]/s[0],abs(np.corrcoef(((xy1-xy1.mean(axis=0))@vt[0]),np.arange(len(xy1)))[0,1]) for c in [-5,-2,0,2,4,6,8,10]: print(c, metric_c(c)) co=minimize_scalar(lambda c:metric_c(c)[0],bounds=(-5,10),method='bounded') co.x,co.fun (np.float64(-1.0116860086807886), np.float64(0.06530887559252528))
pythonang=np.unwrap(np.arctan2(v[:,1],v[:,0])) for i in range(0,50): print(f'{i:2d} t={on[i]:5.2f} θ={np.degrees(ang[i]):6.2f}° A={mag[i]:.5f}')
pythonypred=lowess(yy,xx,frac=.16,return_sorted=False,it=1) for i in range(38): print(f'{i:2d} x={xx[i]:+7.3f} y={yy[i]:+7.3f} residual={yy[i]-ypred[i]:+7.3f}')
pythonfor q in [0,.5,1,1.5,2]: pix=v/mag[:,None]**(1+q) o=np.linalg.svd(pix-pix.mean(axis=0),full_matrices=False)[-1] t=(pix-pix.mean(axis=0))@o.T ratio=(np.ptp(t[:9,0]),np.ptp(t[:9,1])) print(q,ratio,ratio[0]/ratio[1],np.ptp(t[:,0])/np.ptp(t[:,1]))
pythonfor i in range(0,40): print(f'{i:2d}: {v[i,0]:+0.6f}, {v[i,1]:+0.6f}')
I found a strong match for “font n°2”: a Reddit Minecraft block-lettering guide whose styles are labeled No1 and No2. I’m now checking the inferred block positions against that lettering system, including the nonstandard {, } and _ characters. (reddit.com)
pythonfrom scipy.optimize import brentq def make_xyz(alpha=1.5, h=0., mag=mag, v=v): p=2*alpha-1 amin=mag.max() k=amin*(1+h*h)**alpha target=mag/k # r near1 for max and greater for low def f(r,A): return r/((r*r+h*h)**alpha)-A r=np.array([brentq(f,1,10**4,args=(m,)) for m in target]) return v/mag[:,None]*r[:,None] def test_smooth(alpha,h): xy=make_xyz(alpha,h) u,s,vt=np.linalg.svd(xy-xy.mean(axis=0),full_matrices=False) z=(xy-xy.mean(axis=0))@vt.T return s[1]/s[0],abs(np.corrcoef(z[:,0],np.arange(len(z)))[0,1]) for alpha in [1.05,1.2,1.4,1.5,1.7,2.]: for h in [0,.25,.5,.8,1.]: print((alpha,h),tuple(round(c,3) for c in test_smooth(alpha,h)),end=" ") print()
pythonnp.where(np.array(fund)>230)[0] array([29, 30, 31, 32, 33, 34, 35])
pythonfrom scipy.optimize import least_squares def pixels(pattern, offset=0): return np.array([(x+offset,y) for y,row in enumerate(pattern) for x,v in enumerate(row) if v=='1'],dtype=float) templates=[ ['111','100','101','111'], # g 9 ['111','100','100','110'], # c7 ['111','010','010','010'], # t6 ['111','100','110','100'] # f7 ] template_q=[pixels(p,4*k) for k,p in enumerate(templates)] [Q.shape for Q in template_q] [(9, 2), (7, 2), (6, 2), (7, 2)]
pythonobs_groups=[np.sort(ang[a:b]) for a,b in [(0,9),(9,16),(16,22),(22,29)]] def resid_angs(par,return_pred=False): ox,oy,phase,spacing=par rp=[] for k,base in enumerate(template_q): q=base.copy() q[:,0]+=(spacing-4)*k pred=np.arctan2(q[:,1]-oy,q[:,0]-ox)+phase # wrap relative to mean observed; choose nearest branch pred=pred+2*np.pi*np.round((np.mean(obs_groups[k])-np.mean(pred))/(2*np.pi)) rp.append(np.sort(pred)) if return_pred:return rp return np.concatenate([(x-y) for x,y in zip(rp,obs_groups)]) for seed in [[32,12,2.2,4],[45,15,2.4,4],[15,5,2.4,4],[30,20,2.5,4], [70,20,2.5,4]]: res=least_squares(resid_angs,seed,bounds=([-100,-100,-10,3],[200,100,10,8]),max_nfev=4000) print('seed',seed,'fit',res.x,'rmsdeg',np.std(np.degrees(res.fun)))
pythonfrom scipy.ndimage import median_filter ii=np.arange(len(ang)) a_smooth=lowess(ang,ii,frac=.13,return_sorted=False,it=1) def r_geo(params): ox,oy,phase,scale=params pred=np.arctan2(2-oy,ii*scale-ox)+phase pred+=2*np.pi*np.round((a_smooth.mean()-pred.mean())/(2*np.pi)) return pred-a_smooth for guess in [[40,10,2.3,.55],[35,5,2.5,.5],[40,20,2.5,.6],[50,8,2.2,.55]]: r=least_squares(r_geo,guess,bounds=([-200,2.5,-10,.1],[200,200,10,1.5]),max_nfev=4000) print(r.x, np.std(np.degrees(r.fun)))
pythonr=least_squares(r_geo,[39,10.5,2.35,.55],bounds=([-200,2.5,-10,.1],[200,200,10,1.5])) pa=r.x for j in range(0,143,10): print(j, 'obs',round(np.degrees(a_smooth[j]),1),'pred',round(np.degrees(np.arctan2(2-pa[1], j*pa[3]-pa[0])+pa[2]),1))
pythondef fun_arctan(params): c,k,ic,h=params return c+k*np.arctan((ii-ic)/h)-a_smooth for initial in [[.8,.75,70,5],[.8,.75,70,10], [.8,1,70,7]]: res=least_squares(fun_arctan,initial,bounds=([-10,.1,0,.1],[10,3,143,100])) print(res.x, 'rmsdeg',np.std(np.degrees(res.fun))) fa=res.x plt.figure(figsize=(14,3));plt.plot(ii,np.degrees(a_smooth),label='data');plt.plot(ii,np.degrees(a_smooth+fun_arctan(fa)),label='fit') plt.legend();plt.show()
pythonth=(ang-np.pi/4)/(np.pi/4) u1=np.column_stack([np.sin(th),np.cos(th)]) for q in [.25,.5,.75,1.,1.25,1.5,2]: # z/r for each P=u1*mag[:,None]**(-q) eig=np.linalg.svd(P-P.mean(0),full_matrices=False)[1] print('q',q,'width',round(np.ptp(P[:,0]),1),'depth range',round(np.ptp(P[:,1]),1),'slenderness',round(eig[1]/eig[0],4)) fig,axs=plt.subplots(1,3,figsize=(16,4)) for ax,q in zip(axs,[.5,1,1.3]): P=u1*mag[:,None]**(-q) ax.scatter(P[:,0],P[:,1],c=np.arange(len(on)),cmap='viridis',s=20) ax.set_title(f'r=mag^-{q}');ax.grid();ax.set_aspect('equal') plt.tight_layout();plt.show()
pythoncosth=np.cos(th) cfit=np.cov(costh,costh/mag)[0,1]/np.var(costh,ddof=1) z0=(1/mag-cfit)*costh print('cfit',cfit,'z std',z0.std(),'z min max',z0.min(),z0.max(),'mean',z0.mean()) plt.figure(figsize=(16,4)) plt.scatter((1/mag-cfit)*np.sin(th),z0,c=np.arange(len(mag)),cmap='viridis',s=25) plt.ylim(-1,9);plt.grid();plt.show()
The stereo phase changes reveal an important detail: the channels appear to encode signed positional information, not just ordinary left/right loudness. After correcting for that, the events form a long, narrow strip consistent with block lettering. I’m refining the grid alignment to read the characters.
pythonpos=(1/mag-cfit)[:,None]*u1 for i in range(0,47): print(f'{i:2d} X={pos[i,0]:+7.3f} Z={pos[i,1]:+6.3f}')
pythonprint('variables:', [k for k in ['a','on','mag','ang','v','pos','fund','absv'] if k in globals()]) if 'on' in globals(): import numpy as np delta=np.diff(on) for p in [.5,.75,1,1.2]: xy=(mag**(-p))[:,None]*u1 d=np.linalg.norm(np.diff(xy,axis=0),axis=1) print(p, np.corrcoef(delta,d)[0,1],np.percentile(d,[0,25,50,75,100]),np.percentile(delta,[0,25,50,75,100]))
pythonimport matplotlib.pyplot as plt fig,ax=plt.subplots(2,1,figsize=(17,8)) ax[0].scatter(pos[:,0],pos[:,1],c=np.arange(len(pos)),cmap='turbo',s=25) for k in range(0,len(pos),5): ax[0].annotate(str(k),(pos[k,0],pos[k,1]),fontsize=8) ax[0].set_ylim(-1,9);ax[0].set_aspect('equal','box');ax[0].grid() ax[1].plot(np.arange(len(mag)),mag,'o-',ms=3) ax[1].set_ylim(0,.2);ax[1].grid() plt.show()
pythonfig,ax=plt.subplots(figsize=(18,4)) for i,(x,z) in enumerate(pos[:52]): ax.text(x,z,str(i),fontsize=9,color=('r' if i in range(29,36) else 'k'),ha='center',va='center',bbox={'boxstyle':'round,pad=.1','facecolor':'#f0f0f0','edgecolor':'none','alpha':.5}) ax.set_xlim(-44,-8);ax.set_ylim(0,5.5); ax.set_aspect('equal');ax.set_xticks(np.arange(-44,-8));ax.grid(True,alpha=.3) plt.show()
pythontheta=(ang-np.pi/4)/(np.pi/4) for d0 in [4,6.23,10,15,20]: x = d0*np.tan(theta) print('z',d0,'minmax',x.min(),x.max(),'range first29',np.min(x[:29]),np.max(x[:29]),'mean/std dx',np.mean(np.diff(x)),np.std(np.diff(x))) print('all first 42') print(np.round(np.column_stack([np.arange(42),theta[:42]*180/np.pi,6.23*np.tan(theta[:42]),mag[:42]]),3))
pythondef render_grid(xs,zs, dx=1, dz=1, xzero=0, zzero=0, annotate=False): xi=np.rint((xs-xzero)/dx).astype(int); zi=np.rint((zs-zzero)/dz).astype(int) arr=np.full((zi.max()-zi.min()+1,xi.max()-xi.min()+1),' ',dtype='<U1') for j,k in zip(zi,xi): arr[zi.max()-j,k-xi.min()]='#' if annotate: print('x',xi.min(),xi.max(),'z',zi.min(),zi.max(),'width',arr.shape[1], 'height',arr.shape[0]) for row in arr:print(''.join(row)) render_grid(pos[:,0],pos[:,1],annotate=True)
pythonprint(np.round((ang[:50]*180/np.pi),6)) print('unique rounded 0.0001 count', len(set(np.round(ang,4)))) print('dup at 0.1 deg', len(set(np.round(ang*180/np.pi,1))))
pythonfor i in range(64,89): print(i, 'X,Z',tuple(np.round(pos[i],2)),'amp',round(mag[i],4),'pan',round(th[i]*180/np.pi,1))
pythonfrom scipy.optimize import differential_evolution, minimize def obj_x(params): C,b,alpha,k=params theta=(ang-alpha)/k x=(C/mag-b)*np.sin(theta) return -abs(np.mean(np.exp(2j*np.pi*x))) res=differential_evolution(obj_x, [(0.9,1.15),(1,12),(.77,.80),(.77,.80)],popsize=15,maxiter=60,seed=7,polish=True,workers=1) print('best',res.x,'R',-res.fun) for scale in [1,2,3,4]: xx=pos[:,0]/scale print('current /',scale,abs(np.mean(np.exp(2j*np.pi*xx))))
The recording’s first section is consistent with the known gctf{ prefix, and a later group behaves differently from the surrounding letters—possibly one of the special characters mentioned in the clue. The difficult part is recovering the vertical block positions reliably; the stereo balance alone isn’t enough to read the flag yet.
pythonimport itertools def cand(i, xrange=3, zrange=3): xx,zz=pos[i] xs=np.arange(np.floor(xx)-xrange,np.ceil(xx)+xrange+1) zs=np.arange(max(0,np.floor(zz)-zrange),np.ceil(zz)+zrange+1) cc=[] for x,z in itertools.product(xs,zs): if z<.5:continue bearing=np.arctan2(x,z) e=np.abs((bearing-th[i]))*(180/np.pi) rms=((x-xx)**2+(z-zz)**2)**.5 cc.append((e,rms,x,z)) return sorted(cc)[:5] for i in list(range(0,18)) + [28,29,30,31,32,33,34,35,70,72,78,82,90,142]: print(i,[(round(e,3),round(d,1),int(x),int(z)) for e,d,x,z in cand(i)])
pythonfrom scipy.signal import find_peaks fig,ax=plt.subplots(1,1,figsize=(14,3)) for j in [0,1,14,29,36,70,80,100,140]: s=int(on[j]*sr) amp=np.linalg.norm(a[s:s+int(.5*sr)],axis=1) env1=np.sqrt(np.maximum(0,np.convolve(amp**2, np.ones(200)/200,'same'))) ax.plot(np.arange(0,len(env1),120)/sr,env1[::120]/np.max(env1),label=str(j),alpha=.6) ax.set_ylim(0,1.1);ax.legend(ncol=9);plt.show()
pythonfig,axs=plt.subplots(2,1,figsize=(13,5)) for jj,i in enumerate([0,1,30,80]): seg=a[max(0,int((on[i]-.1)*sr)):int((on[i]+.25)*sr)] step=256 envl=np.array([np.sqrt(np.mean(b*b)) for b in np.split(seg[:len(seg)//step*step,0],len(seg)//step)]) tt=np.linspace(-.1,.25,len(envl)) axs[0].plot(tt,envl,label=f'{i}') envr=np.array([np.sqrt(np.mean(b*b)) for b in np.split(seg[:len(seg)//step*step,1],len(seg)//step)]) axs[1].plot(tt,envr,label=f'{i}') axs[0].legend();axs[1].legend() axs[0].grid();axs[1].grid();plt.show()
pythonfrom scipy.ndimage import gaussian_filter1d t0,t1=56,67 ix=(tp>=t0)&(tp<t1) fig,ax=plt.subplots(figsize=(16,5)) ax.plot(tp[ix],rms[ix,0],label='L');ax.plot(tp[ix],rms[ix,1],label='R') for i,t in enumerate(on): if t0<t<t1: ax.axvline(t,color='k',alpha=.3);ax.text(t,.13,str(i),fontsize=9,rotation=90) ax.legend();ax.grid();plt.show()
pythonprint([(i,round(on[i],3)) for i in range(70,85)]) for i in [0,1,30,80,81]: print(i,'on',on[i], 'mean R t-.1',np.sqrt(np.mean(a[int((on[i]-.1)*sr):int((on[i]-.09)*sr),1]**2)))
pythonfor j in np.arange(62.9,63.32,.025): print(round(j,3),np.round(rms[int(j*200),:],3))
pythonprint('sr',sr,'shape',a.shape,'rms shape',rms.shape) t=63.16 print('raw',np.sqrt(np.mean(a[int(t*sr):int((t+.005)*sr)]**2,axis=0)),'rms',rms[int(t*200)]) print('a peak first second',a.max(),a.min()) print('rms t0.1',rms[20],np.sqrt(np.mean(a[int(.1*sr):int(.105*sr)]**2,axis=0)))
python# Correct long-term time drift: each RMS frame contains 220 samples, not 220.5. on_uncorrected=on.copy() on=on_uncorrected*(220/220.5) print('on old, corrected',[(i,round(on_uncorrected[i],4),round(on[i],4)) for i in [0,30,80,142]]) absv2=np.vstack([np.sqrt(np.mean(a[int((t+.025)*sr):int((t+.12)*sr)]**2,axis=0)) for t in on]) v2=absv2.copy() v2[:78,0]*=1 v2[78:,0]*=-1 v2[:64,1]*=-1 mag2=np.linalg.norm(v2,axis=1) ang2=np.unwrap(np.arctan2(v2[:,1],v2[:,0])) th2=(ang2-np.pi/4)/(np.pi/4) pos2=(1/mag2-6.34)[:,None]*np.column_stack([np.sin(th2),np.cos(th2)]) fig,ax=plt.subplots(figsize=(17,4)) ax.scatter(pos2[:,0],pos2[:,1],c=np.arange(143),s=40,cmap='turbo') ax.set_ylim(-2,11);ax.set_aspect('equal');ax.grid() plt.show() print('first14',np.round(np.column_stack([np.arange(14),pos2[:14],mag2[:14]]),2))
pythonfor cc in [0,3,4.5,5,5.5,6,6.34,7]: pp=(1/mag2-cc)[:,None]*np.column_stack([np.sin(th2),np.cos(th2)]) print('c=',cc, 'zpercent',np.round(np.percentile(pp[:,1],[0,10,50,90,100]),2),'z std',round(np.std(pp[:,1]),3),'xrange',round(np.ptp(pp[:,0]),1)) fig,ax=plt.subplots(1,1,figsize=(16,4)) for cc in [4.5,5,5.5,6,6.5,7]: zz=(1/mag2-cc)*np.cos(th2) ax.plot(np.arange(len(zz)),zz,label=f'c{cc}',alpha=.6) ax.legend();ax.grid();plt.show()
pythondt=220/44100 rms_t=np.arange(len(rms))*dt new_on=[] for t in on: start=max(3,round((t-.11)/dt));end=min(len(rms)-3,round((t+.04)/dt)) envt=gaussian_filter1d(rms[:,0]**2+rms[:,1]**2,1) deriv=np.diff(envt) k=np.argmax(deriv[start:end])+start new_on.append(rms_t[k]) new_on=np.asarray(new_on) print('on err stats',np.round(np.percentile(new_on-on,[0,10,25,50,75,90,100]),3)) print([(i,round(on[i],3),round(new_on[i],3)) for i in range(0,143,10)]) # Compare energy envelopes at detected onset for first few fig,ax=plt.subplots(figsize=(16,2)) for i in range(143): ax.scatter(i,new_on[i]-on[i],s=9) ax.grid();plt.show()
pythonfig,ax=plt.subplots(figsize=(15,5)) for i,(x,z) in enumerate(pos2[:37]): ax.text(x,z,str(i),fontsize=10,color=('r' if i>=29 else 'k'),ha='center',va='center',bbox={'boxstyle':'round,pad=.1','facecolor':'#f2f2f2','edgecolor':'none','alpha':.5}) ax.set_xlim(-40,-6);ax.set_ylim(1,5.2);ax.set_aspect('equal');ax.set_xticks(np.arange(-40,-6));ax.grid(True,alpha=.4) plt.show()
pythonfrom scipy.stats import linregress good=np.ones(143,bool);good[76:85]=False;good[29:36]=False sec=1/np.cos(th2) for target in [mag2,1/mag2, np.log(mag2), np.sqrt(1/mag2)]: fit=linregress(sec[good],target[good]);r2=fit.rvalue**2 print('target', [mag2,1/mag2, np.log(mag2), np.sqrt(1/mag2)].index(target) if False else '', 'slope',round(fit.slope,4),'intc',round(fit.intercept,4),'R2',round(r2,4),'residstd',round(np.std(target[good]-(fit.intercept+fit.slope*sec[good])),4)) fig,ax=plt.subplots(figsize=(8,4));ax.scatter(sec[good],mag2[good],s=10);ax.set_xlim(1,20);ax.grid();plt.show()
pythonfrom scipy.signal import savgol_filter x=pos2[:,0];z=pos2[:,1] zb=savgol_filter(z,29,2,mode='interp') z_res=z-zb fig,axs=plt.subplots(3,1,figsize=(17,10)) for ax,scale in zip(axs,[3,6,10]): ax.scatter(x,z_res*scale,s=32,c=np.arange(len(x)),cmap='turbo') ax.set_ylim(-10,12);ax.set_xlim(-40,40);ax.set_title('Z detrended factor '+str(scale));ax.set_aspect('equal');ax.grid(alpha=.3) plt.tight_layout();plt.show()
pythond=np.diff(on) from collections import Counter print('inter onset',np.round(np.percentile(d,[0,5,25,50,75,95,100]),3)) print('big > 1.0',[(i,round(d[i],3)) for i in range(len(d)) if d[i]>1.0]) fig,ax=plt.subplots(figsize=(16,2));ax.bar(np.arange(len(d)),d);ax.set_ylim(0,1.5);ax.grid();plt.show()
pythonfig,axes=plt.subplots(2,1,figsize=(15,7)) for ax,(i,j) in zip(axes,[(0,36),(36,100)]): p=pos2[i:j] ax.plot(p[:,0],p[:,1],'-o',ms=4,lw=.8) for k in range(i,j): ax.text(pos2[k,0],pos2[k,1]+.1,str(k),fontsize=7) ax.set_ylim(0,6);ax.set_aspect('equal');ax.grid() plt.show()
pythonq=np.tan(th2) for target in [q,1/q]: zs=np.linspace(.1,15,15000) R=np.abs(np.exp(2j*np.pi*np.outer(zs,target)).mean(axis=1)) inds=np.argsort(R)[-15:][::-1] print('best',[(round(zs[k],3),round(R[k],3)) for k in inds[:10]]) plt.figure(figsize=(14,3));plt.plot(zs,R);plt.grid();plt.ylim(0,1);plt.show()
pythonfrom scipy.signal import correlate # cross correlations 0.10s snippets of early audio, onset rounded to nearest sample, across notes with +/- 10ms shift win=int(.20*sr) st=int(.02*sr) ref=a[int(on[0]*sr)+st:int(on[0]*sr)+st+win,0] for i in [1,2,3,4,5,9,12,20,28,29,30,36,50,71,80,83,100,140]: si=int(on[i]*sr)+st sam=a[si:si+win, 0] cor=correlate(sam,ref,mode='full',method='fft') lags=np.arange(-len(ref)+1,len(sam)) ix=np.argmax(np.abs(cor[(lags>-800)&(lags<800)])) ix=np.where((lags>-800)&(lags<800))[0][ix] print(i,'corr',round(abs(cor[ix])/(np.linalg.norm(sam)*np.linalg.norm(ref)),3),'lag
python# Extract matched-filter gain ratios for each of the 143 original waveform events. from scipy.signal import fftconvolve win=int(.175*sr) delay=int(.045*sr) base=a[int(on[0]*sr)+delay:int(on[0]*sr)+delay+win,0].astype(float) base29=a[int(on[29]*sr)+delay:int(on[29]*sr)+delay+win,0].astype(float) matched=np.zeros((143,2)) lags_est=np.zeros(143,int) norms=np.zeros(143) for i in range(143): template=base29 if 29<=i<=35 else base pre=160 st=max(0,int(on[i]*sr)+delay-pre) chunk=a[st:st+win+2*pre].astype(float) sg=np.array([1 if i<78 else -1, -1 if i<64 else 1]) rough=absv2[i] dominant=int(np.argmax(rough)) cc=correlate(chunk[:,dominant],template,mode='valid',method='fft') ix=np.argmax(sg[dominant]*cc) sub=chunk[ix:ix+win,:] den=np.dot(template,template) matched[i]=template @ sub / den norms[i]=np.linalg.norm(sub[:,dominant])/np.linalg.norm(template) lags_est[i]=ix-pre print('gains first:',matched[:10].round(3)) print('corr difference',np.round(np.percentile((matched*np.where(np.arange(143)[:,None]<np.array([78,64]),np.array([1,-1]),np.array([-1,1])))[:,0] / absv2[:,0],[0,50,100]),2)) new_amp=matched.copy() mag3=np.linalg.norm(new_amp,axis=1) ang3=np.unwrap(np.arctan2(matched[:,1],matched[:,0])) th3=(ang3-np.pi/4)/(np.pi/4) new_coords=(1/mag3-6.34)[:,None]*np.column_stack([np.sin(th3),np.cos(th3)]) print('range',np.round([new_coords[:,0].min(),new_coords[:,0].max(),new_coords[:,1].min(),new_coords[:,1].max()],2)) print('lags:',np.percentile(lags_est,[0,10,50,90,100])) for j in [0,5,15,30,50,70,82,100,142]: print(j,'magold',round(mag2[j],4),'magnew',round(mag3[j],4),'angold',round(th2[j],4),'angnew',round(th3[j],4))
pythonfor j in range(72,87): print(j,np.round(matched[j],3),np.round(absv2[j],3),round(ang2[j]*180/np.pi,1),round(ang3[j]*180/np.pi,1),lags_est[j])
pythonsgnew=np.sign(matched) sgold=np.column_stack([np.where(np.arange(143)<78,1,-1),np.where(np.arange(143)<64,-1,1)]) mismatch=np.where(np.any(sgnew!=sgold,axis=1))[0] print('mismatch',mismatch) for i in mismatch: print(i,matched[i])
pythonscale=mag2[0]/mag3[0] rad=1/(scale*mag3) -6.34 pos3=rad[:,None]*np.column_stack([np.sin(th3),np.cos(th3)]) fig,ax=plt.subplots(figsize=(17,5)) ax.scatter(pos3[:,0],pos3[:,1],c=np.arange(143),s=40,cmap='turbo') for k in range(0,143,5): ax.text(pos3[k,0],pos3[k,1]+.12,str(k),fontsize=8) ax.set_ylim(-1,9);ax.set_aspect('equal');ax.grid() plt.show() print('coordinate near',np.round(pos3[72:86],2))
pythonfor i in range(0,29): print(f'{i:3} {pos3[i,0]:8.3f} {pos3[i,1]:6.3f} {mag3[i]*scale:6.4f} {th3[i]*180/np.pi:7.2f}')
pythonfrom scipy.optimize import differential_evolution not_special=np.ones(143,bool);not_special[29:36]=False mi=(mag3*scale)[not_special] ai=ang3[not_special] def opt_model(p): K,gamma,b,alpha,ks=p rr=K*(.05/mi)**gamma-b theta=(ai-alpha)/ks xx=rr*np.sin(theta);zz=rr*np.cos(theta) if (rr<=0).any():return 5 + np.mean(rr<=0)*10 sx=np.abs(np.mean(np.exp(2j*np.pi*xx))) sz=np.abs(np.mean(np.exp(2j*np.pi*zz))) return -(sx+sz) bounds=[(10,70),(.4,2.0),(0,15),(.75,.82),(.72,.84)] res=differential_evolution(opt_model,bounds,maxiter=100,popsize=16,seed=44,polish=True) print(res.fun,res.x) K,gamma,b,alpha,ks=res.x xx=(K*(.05/mi)**gamma-b)*np.sin((ai-alpha)/ks) zz=(K*(.05/mi)**gamma-b)*np.cos((ai-alpha)/ks) print('coherence xz',np.abs(np.mean(np.exp(2j*np.pi*xx))),np.abs(np.mean(np.exp(2j*np.pi*zz))), 'xrange',np.percentile(xx,[0,100]),'zrange',np.percentile(zz,[0,100]))
pythondef opt_model2(p): K,gamma,b,alpha,ks=p rr=K*(.05/mi)**gamma-b theta=(ai-alpha)/ks xx=rr*np.sin(theta);zz=rr*np.cos(theta) if (rr<=1).any():return 10+np.mean(rr<=1)*10 ranx=np.ptp(xx) if not 55<ranx<120:return 4+abs(ranx-80)/20 sx=np.abs(np.mean(np.exp(2j*np.pi*xx))) sz=np.abs(np.mean(np.exp(2j*np.pi*zz))) return -(sx+sz) res2=differential_evolution(opt_model2,[(12,45),(.6,1.8),(0,15),(.77,.80),(.76,.81)],maxiter=180,popsize=18,seed=12,polish=True) print('found',res2.fun,res2.x) K,gamma,b,alpha,ks=res2.x rr=K*(.05/mi)**gamma-b;theta=(ai-alpha)/ks xx=rr*np.sin(theta);zz=rr*np.cos(theta) print('coherence xz',np.abs(np.mean(np.exp(2j*np.pi*xx))),np.abs(np.mean(np.exp(2j*np.pi*zz))), 'xrange',np.percentile(xx,[0,100]),'zrange',np.percentile(zz,[0,100]))
python# Match measured stereo bearings against angles of a candidate INTEGER lattice. import scipy.spatial real_th=(ang3-np.pi/4)/(np.pi/4) xx0,zz0=np.meshgrid(np.arange(-80,81),np.arange(1,13)) ref_angles=np.arctan2(xx0.ravel(),zz0.ravel()) radii=np.hypot(xx0.ravel(),zz0.ravel()) angles_sort=np.argsort(ref_angles) sorted_th=ref_angles[angles_sort] j=np.searchsorted(sorted_th,real_th) j=np.clip(j,1,len(sorted_th)-1) err=np.minimum(np.abs(sorted_th[j]-real_th),np.abs(sorted_th[j-1]-real_th)) print('median errors degrees',np.percentile(err*180/np.pi,[0,25,50,75,90,100])) print('counts <.001deg',np.sum(err*180/np.pi<.001),'<.005',np.sum(err*180/np.pi<.005), '<.01',np.sum(err*180/np.pi<.01)) print('first30',np.round(err[:30]*180/np.pi,4))
pythonprint('i | x | z-res | rawz') for i in range(36): print(f'{i:2d} {pos3[i,0]:6.1f} {(pos3[i,1]-savgol_filter(pos3[:,1],29,2,mode="interp")[i])*7:6.1f} {pos3[i,1]:6.1f}')
pythonpp=pos3.copy() valid=np.r_[np.arange(29),np.arange(36,143)] for sz in [1.5,2,2.5,3,4]: print('\nZscale',sz) xs=pp[valid,0];zs=pp[valid,1]*sz xi=np.rint(xs).astype(int);zi=np.rint(zs).astype(int) ax0=-40; ax1=-10 zi0=0;zi1=18 arr=np.full((zi1-zi0+1,ax1-ax0+1),' ',dtype='<U1') for j,k in zip(zi,xi): if ax0<=k<=ax1 and zi0<=j<=zi1: arr[zi1-j,k-ax0]='#' print('\n'.join(f'{zi1-ii:02} '+''.join(row) for ii,row in enumerate(arr) if '#' in row))
pythonfig,ax=plt.subplots(figsize=(7,5)) for i in range(9): x,z=pos3[i]; ax.scatter(x,z,s=100,color='black');ax.text(x+.1,z+.03,str(i)) ax.set_xlim(-38,-24);ax.set_ylim(2,4.5);ax.set_aspect('equal');ax.grid() plt.show()
pythonprint('state?', [k for k in ['a','on','matched','mag2','ang3','pos3','th3','absv2','rms'] if k in globals()]) print((len(on),on[-1],len(a)) if 'on' in globals() else 'reset')
pythonimport matplotlib.pyplot as plt, numpy as np fig,axes=plt.subplots(2,1,figsize=(19,8)) axes[0].scatter(pos3[:,0],pos3[:,1],c=np.arange(len(pos3)),cmap='viridis',s=35) for j in range(len(pos3)): if j%5==0 or j in [28,29,35,36,75,76,84,85,135,136]: axes[0].text(*pos3[j],str(j),fontsize=7) axes[0].grid();axes[0].axis('equal');axes[0].set_ylim(-1,15) axes[1].plot(np.arange(len(mag2)), mag2,alpha=.8,label='mag2') axes[1].plot(np.arange(len(th3)),th3/15,label='bearing/15') axes[1].legend();axes[1].grid() plt.show()
pythonfor k in range(0,47): print(k,round(on[k],2), 'x',round(pos3[k,0],2),'z',round(pos3[k,1],2), 'm',round(mag2[k],4), 'th',round(th3[k],3))
pythondt=np.diff(on) print('Large gaps') for i,v in enumerate(dt): if v>0.95:print(i,'->',i+1,round(v,3)) print('gap percentiles',np.quantile(dt,[0,.1,.25,.5,.75,.9,.95,1]))
pythontheta=th3 pos=(1/mag2 - 6.34)[:,None]*np.column_stack([np.sin(theta),np.cos(theta)]) fig,ax=plt.subplots(figsize=(17,3.5));ax.scatter(pos[:,0],pos[:,1],c=np.arange(143),s=27);ax.axis('equal');ax.grid() for j in [0,7,8,14,15,22,23,28,29,35,36,46,54,55,71,72,76,84,85,106,135,136,142]: ax.text(*pos[j],str(j),fontsize=8) plt.show() print('29:35',np.round(pos[29:36],2))
pythonfrom scipy.signal import hilbert # compare raw RMS 10ms over first 2 sec and event gaps plt.figure(figsize=(14,2)) rms0=np.sqrt(np.mean(a[:int(2.4*sr),:].astype(float)**2,axis=1)) # huge numeric baseline int32 need scaling plt.plot(np.arange(0,len(rms0),441)/sr,rms0[::441]);plt.yscale('log');plt.ylim(1e-7,1e9);plt.grid();plt.show() print(a.dtype, a.min(),a.max())
pythonfrom scipy.signal import butter,sosfilt i0=int(on[0]*sr) i1=int(on[1]*sr) fig,ax=plt.subplots(2,1,figsize=(15,4)) N=900 ax[0].plot(np.arange(N)/sr,a[i0:i0+N,0],lw=.5);ax[0].set_title('first event first 20ms') t2=np.arange(i0,i1)/sr-on[0] ax[1].plot(t2[::100],a[i0:i1,0][::100],lw=.5);ax[1].set_title('first event waveform to next event') plt.tight_layout() plt.show() print('nonzero first',np.where(np.abs(a[:,0])>1e-8)[0][0],np.where(np.abs(a[:,1])>1e-8)[0][0])
pythonfrom scipy.spatial import cKDTree # nearest pairs of observed data (ang,mag) normalized X=np.column_stack([np.log(mag2)*8,th3]) tree=cKDTree(X) d,inds=tree.query(X,k=2) ix=np.argsort(d[:,1]) print([(int(i),int(inds[i,1]),round(d[i,1],4)) for i in ix[:25]])
I’ve narrowed down the 143 note events and found that their stereo direction and volume provide two different clues about position. The first few notes cluster into repeated columns, which supports a block-letter pattern. I’m testing how to recover the rows without distorting the lettering.
pythonfig,axs=plt.subplots(3,1,figsize=(18,9),sharex=True) for ax,K in zip(axs,[3.5,4.8,6.34]): p=(1/mag2 - K)[:,None]*np.column_stack([np.sin(th3),np.cos(th3)]) ax.scatter(p[:,0],p[:,1],c=np.arange(143),s=15,cmap='turbo') ax.set_ylim(1,6);ax.set_title(f'K={K}');ax.grid() for i in [0,7,8,14,15,22,23,28,29,35,36,84,85,135,136,142]: ax.text(*p[i],str(i),fontsize=8) plt.tight_layout();plt.show()
pythonfor i in range(72,89): print(i,round(on[i],2),round(th3[i],3),round(mag2[i],4),np.round(pos[i],2),round(np.linalg.norm(matched[i]),3))
pythonfig,ax=plt.subplots(figsize=(20,4)) p=(1/mag2 - 4.7)[:,None]*np.column_stack([np.sin(th3),np.cos(th3)]) y=p[:,1]-2.7 for i,(x,v) in enumerate(zip(p[:,0],y)): ax.add_patch(plt.Rectangle((x-0.65, v-0.14),1.3,0.28,color='black')) ax.set_xlim(-41,39); ax.set_ylim(-1.6,2); ax.set_aspect('auto');ax.grid();plt.show()
pythonfrom scipy.ndimage import gaussian_filter1d import matplotlib.pyplot as plt p=(1/mag2 - 4.7)[:,None]*np.column_stack([np.sin(th3),np.cos(th3)]) xp=20*np.arcsinh(p[:,0]/12) # polynomial fit lower envelope baseline vs x index for all notes except76..84 coef=np.polyfit(xp[np.r_[np.arange(0,76),np.arange(85,143)]],p[np.r_[np.arange(0,76),np.arange(85,143)],1],5) yl=p[:,1]-np.polyval(coef,xp) fig,axs=plt.subplots(2,1,figsize=(20,6),sharex=True) for ax,yy in zip(axs,[p[:,1],yl]): for i,(x,v) in enumerate(zip(xp,yy)): ax.add_patch(plt.Rectangle((x-0.25,v-0.13),.5,.26,color='black')) ax.set_ylim(-1.5,5.3
pythonfrom scipy.optimize import differential_evolution theta=(ang3-np.pi/4)/(np.pi/4) m=mag2 def fn(par,full=False): A,p,B,ang0,angs,xoff,zoff=par r=A/(m**p)+B t=(ang3-ang0)*angs xx=r*np.sin(t)+xoff zz=r*np.cos(t)+zoff # favor <=5 row levels err=(np.sin(np.pi*xx)**2+np.sin(np.pi*zz)**2)/2 penalty=np.maximum(0,3.0-(np.quantile(xx,.99)-np.quantile(xx,.01)))*.01 val=np.mean(err)+penalty return (val,xx,zz) if full else val init=[1,1,-6.34,.7854,1/.7854,0,0] print('current grid score',fn(init)) params=[(.45,2),(.75,1.25),(-20,3),(.65,.92),(1.0,1.6),(-.5,.5),(-.5,.5)] fit=differential_evolution(fn,params,maxiter=210,popsize=17,seed=58,polish=True
pythonval,xx,zz=fn(fit.x,True) print(np.min(xx),np.max(xx),np.min(zz),np.max(zz)) fig,ax=plt.subplots(figsize=(18,3)) ax.scatter(xx,zz,c=np.arange(143),s=35,cmap='rainbow') ax.set_xticks(np.arange(np.floor(min(xx))-2,np.ceil(max(xx))+2,2)) ax.set_yticks(np.arange(0,15,1)) ax.grid();ax.set_aspect('equal') plt.show() print([(i,round(xx[i],2),round(zz[i],2)) for i in range(16)])
pythondef fn_grid(par,full=False): A,p,B,ang0,angs,xoff,zoff=par r=A/m**p+B; t=(ang3-ang0)*angs xx=r*np.sin(t)+xoff;zz=r*np.cos(t)+zoff rxx=np.quantile(xx,[.02,.98]);rzz=np.quantile(zz,[.03,.97]) # an x range of 65-85 and y range of 2.5-4.5 for 4 rows xlen=rxx[1]-rxx[0];zlen=rzz[1]-rzz[0] pen=.15*(np.clip(65-xlen,0,None)/10)**2+.15*(np.clip(xlen-85,0,None)/10)**2 \ +.1*(np.clip(2.5-zlen,0,None)/2)**2+.1*(np.clip(zlen-5,0,None)/2)**2 err=np.sin(np.pi*xx)**2+np.sin(np.pi*zz)**2 value=err.mean()/2+pen return (value,xx,zz) if full else value params=[(.5,2),(.85,1.2),(-15,1),(.69,.88),(.95,1.8),(-.5,.5),(-.5,.5)] fit2=differential_evolution(fn_grid,params,maxiter=260,popsize=20,seed=69,tol=.0001,polish=True) print(fit2.fun,fit2.x) val,xx,zz=fn_grid(fit2.x,True) print(np.quantile(xx,[.02,.98]),np.quantile(zz,[.02,.98]))
pythonfor i in [0,30,64,80,85,120,142]: s=a[int((on[i]+.025)*sr):int((on[i]+.12)*sr)] corr=np.corrcoef(s.T)[0,1] print(i,'corrLR',round(corr,4),'R/L leastsq',round(np.dot(s[:,0],s[:,1])/np.dot(s[:,0],s[:,0]),3))
pythonscale=mag2[0]/np.linalg.norm(matched[0]) xy=matched*scale fig,ax=plt.subplots(figsize=(10,10)) ax.scatter(*xy.T,c=np.arange(143),cmap='turbo') for i in range(0,40): ax.annotate(str(i),xy[i],fontsize=8) ax.axis('equal');ax.grid();plt.show()
pythonfrom IPython.display import Audio, display display(Audio(a[:int(sr*12)],rate=sr))
pythonprint('sr=',sr, type(sr), 'a=',a.shape,a.dtype, 'audio len',12*44100) display(Audio(a[:12*44100],rate=44100))
pythonimport inspect print(Audio,inspect.signature(Audio)) print('rate',44100,'shape',a[:100].T.shape) display(Audio(data=a[:12*44100].T.astype(np.float64),rate=44100,normalize=False))
pythonp=(1/mag2 - 4.9)[:,None]*np.column_stack([np.sin(th3),np.cos(th3)]) groups=[(0,8,'g?'),(8,15,'c?'),(15,23,'t?'),(23,29,'f?'),(29,36,'{?')] fig,axs=plt.subplots(1,5,figsize=(15,3)) for ax,(s,e,name) in zip(axs,groups): sub=p[s:e] ax.scatter(sub[:,0]-np.min(sub[:,0]),sub[:,1]-np.min(sub[:,1]),s=100) for i in range(e-s):ax.text(sub[i,0]-sub[:,0].min(),sub[i,1]-sub[:,1].min(),str(i+s),fontsize=10) ax.set_title(name);ax.set_xlim(-1,11);ax.set_ylim(-.5,2.5);ax.set_xticks(range(0,12));ax.set_yticks(np.arange(0,2.6,.5));ax.grid() plt.tight_layout();plt.show()
The positional mapping still isn’t reliable enough to distinguish the letters. I can verify the sound events and the seven-note pitch change, but the plotted shapes do not consistently match the expected font. I’m saving the extracted measurements so the remaining decoding can be checked without repeating the audio analysis.
pythonimport numpy as np, pandas as pd, matplotlib.pyplot as plt from pathlib import Path out=Path('/mnt/data') onset_times=np.asarray(on) signed_rms=np.sign(matched)*absv2 df=pd.DataFrame({ 'event':np.arange(143), 'time_s':np.round(onset_times,5), 'pitch_hz':np.where((np.arange(143)>=29)&(np.arange(143)<=35),233.08,220.0), 'left_signed_rms':np.round(signed_rms[:,0],6), 'right_signed_rms':np.round(signed_rms[:,1],6), 'magnitude_rms':np.round(mag2,6), 'stereo_direction_radians':np.round(ang3,6), }) csv_path=out/'dabas_extracted_events.csv' df.to_csv(csv_path,index=False) fig,(ax1,ax2)=plt.subplots(2,1,figsize=(15,7),sharex=True,gridspec_kw={'height_ratios':[2,1]}) x=np.arange(143) colors=np.where((x>=29)&(x<=35),'#e66036','#397ac1') ax1.scatter(x,ang3,c=colors,s=22);ax1.plot(x,ang3,alpha=.22,color='black',lw=.8) ax1.set_ylabel('Stereo direction (radians)') ax1.set_title('Dabas Sound — extracted event features (143 events)') ax1.grid(alpha=.28) ax2.scatter(x,mag2,c=colors,s=22) ax2.plot(x,mag2,alpha=.22,color='black',lw=.8) ax2.set_ylabel('Relative sound strength (RMS)') ax2.set_xlabel('Event index (recording order)') ax2.grid(alpha=.28) for i in [29,35,84,135]: ax1.axvline(i,color='gray',ls='--',alpha=.4,lw=.7) ax2.axvline(i,color='gray',ls='--',alpha=.4,lw=.7) fig.tight_layout() png_path=out/'dabas_event_features.png' fig.savefig(png_path,dpi=180) plt.close(fig) print(csv_path,csv_path.exists(),csv_path.stat().st_size) print(png_path,png_path.exists(),png_path.stat().st_size)
I analyzed the attached dabas-sound.wav, but I haven't recovered a verified flag yet.
The recording contains 143 distinct sound events. Most are pitched around 220 Hz, while seven consecutive events (indices 29–35) are around 233 Hz. Their changing stereo direction and volume suggest that the sound encodes positions in the Minecraft world.
I also found a likely match for the Reddit reference: *For those who want to write with blocks*, which includes a font labeled No2. (www.reddit.com)
The remaining obstacle is converting the stereo measurements into the correct block coordinates. The current reconstruction doesn't produce consistently readable letters, so I don't want to invent a gctf{...} flag.
I saved the intermediate results:
Status: Audio events extracted; flag not yet decoded.