fix: MP3 recordings carried a hiss the WAV of the same audio did not

The operator split it in one test: WAV is fine, MP3 is not. The MP3 path is the
only one that resamples — 16 kHz up to 32 kHz, because the encoder emits broken
frames at MPEG-2 rates — and it did so by averaging neighbouring samples.

That is an upsampler in name only. Every component at f is mirrored to
16 kHz − f, and the test measures the mirror just 7 dB below the wanted signal:
a second copy of the whole spectrum. On speech it adds brightness; on receiver
noise, which is broadband, it lays a second noise floor over the top of the
band. Hence "a hiss that is not the QRM, louder than the voice" — and hence its
absence from the WAV, which resamples nothing.

Zero-stuffing plus a 63-tap windowed sinc at the old Nyquist puts the image
56.8 dB down instead of 7, at a cost measured in microseconds against the MP3
encode that follows. Two tests pin it: the image suppression, and that the level
is unchanged — a fix that quietly halved every recording would be a second
surprise on top of the one being repaired.
This commit is contained in:
2026-07-31 21:27:57 +02:00
parent da1f3eb2bd
commit fddb3c45c4
3 changed files with 151 additions and 9 deletions
+2 -2
View File
@@ -10,7 +10,7 @@
"The \"From radio\" and mic levels take effect as you drag the slider, and the RX level now applies to what you HEAR through OpsLog, not only to recordings.",
"The padlocks on frequency, band, mode and times no longer flicker under the pointer and refuse the click.",
"The callsign field is wider, taking the room from the RST pair.",
"QSO recordings are no longer swamped by hiss: the audio was resampled with the cheap converter, folding everything above 8 kHz back onto the voice."
"MP3 recordings no longer carry a hiss of their own: the 16→32 kHz conversion mirrored the whole spectrum only 7 dB down. It is now 57 dB down."
],
"fr": [
"Les filtres du cluster gagnent NEW POTA et NEW COUNTY, à côté des pastilles de statut existantes.",
@@ -20,7 +20,7 @@
"Les niveaux « From radio » et micro agissent pendant que vous déplacez le curseur, et le niveau RX s'applique désormais à ce que vous ENTENDEZ dans OpsLog, pas seulement aux enregistrements.",
"Les cadenas de fréquence, bande, mode et heures ne scintillent plus sous le curseur et acceptent le clic.",
"Le champ indicatif est plus large, la place venant de la paire RST.",
"Les enregistrements de QSO ne sont plus noyés par le souffle : l'audio était rééchantillonné avec le convertisseur économique, qui repliait tout ce qui dépasse 8 kHz sur la voix."
"Les enregistrements MP3 ne portent plus un souffle qui leur est propre : la conversion 16→32 kHz recopiait tout le spectre à seulement 7 dB en dessous. Il est désormais à 57 dB."
]
},
{
+70 -7
View File
@@ -3,6 +3,7 @@
package audio
import (
"math"
"os"
"github.com/braheezy/shine-mp3/pkg/mp3"
@@ -50,19 +51,81 @@ func writeMP3(path string, pcm []byte) error {
return enc.Write(f, stereo)
}
// upsample2 doubles the sample rate with linear interpolation (16 kHz → 32 kHz).
// upsample2 doubles the sample rate, 16 kHz → 32 kHz, WITH the interpolation
// filter that makes the operation legitimate.
//
// It used to interpolate linearly — each new sample the average of its
// neighbours. That is an upsampler in name only: every component at f is
// mirrored to 16 kHz f, measured just 7 dB down (see the test). On speech it
// adds brightness; on receiver noise, which is broadband, it lays a second
// noise floor across the top of the band. Operators heard it as a hiss louder
// than the voice, present in the MP3 and absent from the WAV of the very same
// audio.
//
// Zero-stuffing followed by a windowed-sinc low-pass at the old Nyquist is the
// textbook answer, and it puts the image below 40 dB. 63 taps is a few hundred
// microseconds of work on an eight-second recording — nothing, next to the MP3
// encode that follows.
func upsample2(in []int16) []int16 {
if len(in) == 0 {
return in
}
out := make([]int16, len(in)*2)
for i := range in {
out[2*i] = in[i]
if i+1 < len(in) {
out[2*i+1] = int16((int32(in[i]) + int32(in[i+1])) / 2)
} else {
out[2*i+1] = in[i]
half := len(upsampleFIR) / 2
for i := range out {
var acc float64
// The zero-stuffed signal is non-zero only at even indices, so only
// every other tap contributes — the loop skips the zeros rather than
// multiplying by them.
start := (i - half + 1) &^ 1 // first even index in the window
for j := start; j <= i+half; j += 2 {
k := i - j + half
if k < 0 || k >= len(upsampleFIR) {
continue
}
src := j / 2
if src < 0 || src >= len(in) {
continue
}
acc += float64(in[src]) * upsampleFIR[k]
}
// ×2 for the energy lost to zero-stuffing, then clamp.
v := acc * 2
if v > 32767 {
v = 32767
} else if v < -32768 {
v = -32768
}
out[i] = int16(v)
}
return out
}
// upsampleFIR is a 63-tap Hamming-windowed sinc, cut off at a quarter of the
// NEW rate (8 kHz at 32 kHz) — exactly the old Nyquist, which is where the
// images begin.
var upsampleFIR = func() []float64 {
const n = 63
const fc = 0.25 // cycles/sample at the new rate
h := make([]float64, n)
mid := (n - 1) / 2
var sum float64
for i := 0; i < n; i++ {
m := float64(i - mid)
var v float64
if m == 0 {
v = 2 * fc
} else {
v = math.Sin(2*math.Pi*fc*m) / (math.Pi * m)
}
// Hamming window: a rectangular one would ring and put the stopband
// only ~21 dB down, which is not enough to be worth the filter.
v *= 0.54 - 0.46*math.Cos(2*math.Pi*float64(i)/float64(n-1))
h[i] = v
sum += v
}
for i := range h {
h[i] /= sum // unity gain at DC
}
return h
}()
+79
View File
@@ -0,0 +1,79 @@
package audio
import (
"math"
"testing"
)
// goertzel returns the magnitude of one frequency bin, so a test can ask "how
// much energy is at 10 kHz" without pulling in an FFT.
func goertzel(x []int16, fs, freq float64) float64 {
w := 2 * math.Pi * freq / fs
c := 2 * math.Cos(w)
var s1, s2 float64
for _, v := range x {
s := float64(v) + c*s1 - s2
s2, s1 = s1, s
}
return math.Hypot(s1-s2*math.Cos(w), s2*math.Sin(w))
}
// Upsampling 16 kHz → 32 kHz must not create an audible mirror image.
//
// MP3 recordings came back with a hiss that is not on the air and is not in the
// WAV of the same audio — "louder than the voice", from a station that hears the
// voice well clear of the noise. The MP3 path is the only one that resamples,
// and it did so by linear interpolation: every component at f is mirrored to
// 16 kHz f, which for broadband receiver noise means a second noise floor
// spread across the top of the band.
func TestUpsampleSuppressesTheImage(t *testing.T) {
const fs = 16000.0
const tone = 6000.0 // its image lands at 16000 6000 = 10 kHz
in := make([]int16, 4096)
for i := range in {
in[i] = int16(12000 * math.Sin(2*math.Pi*tone*float64(i)/fs))
}
out := upsample2(in)
if len(out) != 2*len(in) {
t.Fatalf("upsample2 returned %d samples for %d", len(out), len(in))
}
wanted := goertzel(out, 2*fs, tone)
image := goertzel(out, 2*fs, 2*fs-tone-16000) // = 10 kHz
if wanted == 0 {
t.Fatal("the tone itself did not survive upsampling")
}
db := 20 * math.Log10(image/wanted)
t.Logf("image at 10 kHz is %.1f dB below the 6 kHz tone", db)
if db > -35 {
t.Errorf("image only %.1f dB down — it is audible as added hiss; want at least 35 dB", db)
}
}
// The filter must not change the level, or every existing recording would come
// back quieter (or clipped) after the fix — a second surprise on top of the one
// being repaired.
func TestUpsampleKeepsTheLevel(t *testing.T) {
const fs = 16000.0
in := make([]int16, 4096)
for i := range in {
in[i] = int16(10000 * math.Sin(2*math.Pi*1000*float64(i)/fs))
}
rms := func(x []int16) float64 {
var acc float64
for _, v := range x {
acc += float64(v) * float64(v)
}
return math.Sqrt(acc / float64(len(x)))
}
// Skip the filter's warm-up and tail, where the window is only partly fed.
out := upsample2(in)
got, want := rms(out[200:len(out)-200]), rms(in[100:len(in)-100])
ratio := got / want
t.Logf("level after upsampling: ×%.3f", ratio)
if ratio < 0.9 || ratio > 1.1 {
t.Errorf("level changed by ×%.3f — want within 10%%", ratio)
}
}