pgefier

i wanted my mic to sound like the youtuber darklordpge, so i measured his voice off his uploads and wrote a tool that fits my mic's eq to match it. the eq runs as an effect on the mic itself, so discord, games, obs, and anything else that opens the mic all get the matched sound, with no virtual device to set up.

it only touches the eq, so it's still my voice, just with his tone.

measuring a voice

first it turns a recording into a curve. it ffts the audio in short frames, subtracts the stereo side channel so mostly the centered voice is left, keeps the loudest frames (the speech), and subtracts the quietest ones (the room). then it normalizes everything to the 1 khz band, so recordings at different volumes line up. i ran that over three of pge's videos to get his target curve, and the same function measures my mic.

def voice_curve(M, S, sr, bands):
    N,hop = 2048,512; win=np.hanning(N)
    idx,nf = frames(M,N,hop)
    PM = np.abs(np.fft.rfft(M[idx]*win,axis=1))**2
    PS = np.abs(np.fft.rfft(S[idx]*win,axis=1))**2
    PV = np.maximum(PM-PS,0.) if np.any(S) else PM
    f  = np.fft.rfftfreq(N,1/sr)
    sb = (f>=300)&(f<3400)
    e  = 10*np.log10(np.maximum(PV[:,sb].sum(axis=1),1e-20))
    act, bed = e>=np.percentile(e,75), e<=np.percentile(e,15)
    vs  = np.maximum(PV[act].mean(axis=0)-PV[bed].mean(axis=0),1e-20)
    ref = 10*np.log10(vs[(f>=900)&(f<1100)].mean())
    keys = sorted(bands,key=int); hi=[int(k) for k in keys]+[12500]
    cur = {k: float(10*np.log10(vs[(f>=int(k))&(f<hi[i+1])].mean())-ref)
           for i,k in enumerate(keys)}
    lv = 10*np.log10(np.maximum((M[idx]**2).mean(axis=1),1e-20))
    return cur, lv[act], act.sum()

his curve comes out about +11 db at 100-125 hz, falling to about -24 db by 6.3 khz. the bass mostly comes from him sitting 2-4 inches from the grille.

solving the eq

the eq software gives me three peaking bands and a high-pass, and each band's gain caps at +/-12 db. so it's a fitting problem. pick a frequency, a q, and a gain for each band so the total response lands as close to his curve as possible. each band picks from a small grid of frequencies and q values, with gains from -12 to +12.

GRID = {
 "low":  dict(freq=[125,150,175,200,225,250],                    q=[0.5,1,2,3,4.5]),
 "mid":  dict(freq=[400,450,500,600,700,800,900,1000,1150,1250], q=[0.5,1,2,2.5,3]),
 "high": dict(freq=[3000,4000,4650,5000,5650,6000,6250],          q=[0.5,1,2,3.5,5]),
}
GAINS = list(range(-12,13))

the fit is coordinate descent. hold two bands fixed, brute-force the third against whatever error is left, move on to the next band, and loop until a full pass stops improving (or six passes, whichever comes first).

def fit(desired, fc, w):
    cur = {k: dict(frequency=GRID[k]["freq"][0], gain=0, q=1.0) for k in GRID}
    def total(exclude=None):
        t = np.zeros_like(fc)
        for k,v in cur.items():
            if k!=exclude: t += peak_resp(fc, v["frequency"], v["gain"], v["q"])
        return t
    prev = None
    for _ in range(6):
        for k in ("low","high","mid"):
            resid = desired - total(exclude=k)
            bb, be = None, None
            for f0 in GRID[k]["freq"]:
                for q in GRID[k]["q"]:
                    for g in GAINS:
                        err = float(np.sum(w*(resid-peak_resp(fc,f0,g,q))**2))
                        if be is None or err < be: be, bb = err, dict(frequency=f0,gain=g,q=q)
            cur[k] = bb
        e = float(np.sum(w*(desired-total())**2))
        if prev is not None and abs(prev-e) < 1e-6: break
        prev = e
    return cur, total()

limits

his curve needs about -24 db at 6.3 khz and the bands stop at -12, so the de-esser and noise reduction do the rest of the treble cut. and since most of his low end is mic distance, the eq can only imitate it. sitting closer to the mic would do more than any of this.

download the code (8 kb zip)