feat(audio): v1 signature producer, bit-exact with the pipeline

`AudioSignature` is the DSP — band table, periodic Hann, radix-2 FFT,
band-mean peak, energy class, packing — and `AudioSignatureService` the
decode, running the FFmpeg binary `IMediaEncoder.EncoderPath` names. The
plugin gained no dependency.

The server specification's prose does not determine a byte stream, so the
parameters it leaves open are pinned by the fixture shared with the
extraction repo and restated at the top of `AudioSignature`: double
throughout, whole frames only, periodic Hann, unnormalised FFT, band mean
rather than sum, argmax ties to the lowest index.

`fixtures/audio/` holds the extraction repo's three files byte-identically
and the computed signature equals the recorded vector exactly. The binding
check regenerates the fixture PCM from `make_fixture.py`'s arithmetic and
verifies it against the recorded decode checksums, so it runs on a host
with no codec at all and a decode divergence stays distinguishable from a
DSP one; the two tests that drive real FFmpeg self-skip without a binary.

The workflow named "Test Plugin" until now only compiled one. A test that
is built and never run is not evidence, and a golden vector shared across
two repos exists precisely so CI fails when they drift.

TRACES: JR-042, JR-043 | SR-003
This commit is contained in:
2026-07-31 16:24:11 +02:00
parent 19aecee646
commit 17be9bcc4e
9 changed files with 1359 additions and 0 deletions
@@ -0,0 +1,373 @@
using System;
using System.Collections.Generic;
namespace Jellyfin.Plugin.JRay.Services;
/// <summary>
/// The JRay v1 content-derived audio signature: spectral peak bins taken from
/// the centre of a media file, so a truth file is self-identifying.
/// </summary>
/// <remarks>
/// The construction is owned by the public server specification §3 and is
/// implemented a second time, in C++, by the extraction pipeline
/// (<c>src/audio_signature.*</c>, extraction <c>IR-004</c>). <b>The two must
/// agree byte for byte</b> — a signature that differs in any parameter simply
/// does not match, which defeats the entire point of having one. Every constant
/// below is therefore load-bearing, and any change to one is a change to the
/// <c>v1:</c> prefix as well.
/// <para>
/// The server specification's prose is not sufficient to reproduce a byte
/// stream, so the details it leaves open are pinned by the golden fixture
/// shared with the extraction repo (JR-043,
/// <c>fixtures/audio/jray_audio_v1_golden.json</c>), and restated here:
/// </para>
/// <list type="bullet">
/// <item>Arithmetic is IEEE-754 <c>double</c> throughout. <c>float</c> is not
/// sufficient: the fixture has frames whose two strongest bands are within 1.3%
/// of each other.</item>
/// <item>Samples arrive as FFmpeg's native <c>s16 -&gt; flt</c> conversion,
/// <c>x * (1/32768)</c>, widened to double here.</item>
/// <item>Whole frames only:
/// <c>n_frames = 1 + (n_samples - 4096) / 1024</c>, integer division.</item>
/// <item>Hann window, <b>periodic</b>: <c>0.5 * (1 - cos(2*pi*n/4096))</c>, not
/// the symmetric <c>N-1</c> variant.</item>
/// <item>Plain radix-2 FFT, no normalisation; magnitude is
/// <c>sqrt(re^2 + im^2)</c>.</item>
/// <item>A band's value is the <b>mean</b> of the linear magnitudes in it, so a
/// wide high band is not favoured over a narrow low one.</item>
/// <item>The peak is the <c>argmax</c> over the 32 bands, ties to the lowest
/// index. The specification's log is a monotone squash and so cannot change an
/// argmax; it is applied only where it is observable, in the energy class.</item>
/// </list>
/// <para>
/// The FFT is written out here rather than taken from a library for the same
/// reason the pipeline writes its own: it is a fixed, fully specified transform,
/// and a dependency whose version could change the numerics is a liability when
/// the output has to be identical across two languages.
/// </para>
/// </remarks>
// TRACES: JR-042 | SR-003
public static class AudioSignature
{
/// <summary>Sample rate the signature is computed at, in Hz.</summary>
public const int SampleRate = 11025;
/// <summary>STFT frame size, in samples.</summary>
public const int FrameSize = 4096;
/// <summary>STFT hop size, in samples (~93 ms).</summary>
public const int HopSize = 1024;
/// <summary>Number of logarithmically spaced bands.</summary>
public const int NumBands = 32;
/// <summary>Low edge of the analysed band, in Hz.</summary>
public const double BandLoHz = 300.0;
/// <summary>High edge of the analysed band, in Hz.</summary>
public const double BandHiHz = 3000.0;
/// <summary>Length of the analysed window, in seconds.</summary>
public const double WindowSec = 120.0;
/// <summary>
/// Length of the analysed window, in samples (120.000 s at 11025 Hz).
/// </summary>
/// <remarks>
/// The decoded window is truncated to exactly this, so the frame count is
/// the same for every input and does not wobble with seek granularity or a
/// resampler tail.
/// </remarks>
public const int WindowSamples = 1323000;
/// <summary>Frames a full window yields: <c>1 + (1323000 - 4096) / 1024</c>.</summary>
/// <remarks>
/// The server specification says "~1290" and accepts a tolerance; the exact
/// count follows from the framing rule and is 1288.
/// </remarks>
public const int ExpectedFrames = 1288;
/// <summary>Guard added to both sides of the energy ratio.</summary>
public const double EnergyEps = 1e-12;
/// <summary>
/// The signature's own version prefix, separate from <c>schema_version</c>.
/// </summary>
/// <remarks>
/// A future change to the DSP chain must be <i>detectable</i> rather than
/// silently producing signatures that no longer match (JR-045).
/// </remarks>
public const string VersionPrefix = "v1:";
// Class thresholds on log10(E_frame / E_median). They deliberately straddle
// r = 0 rather than sit on it, so the median frame itself is never on a
// boundary.
private static readonly double[] EnergyClassEdges = [-0.6, -0.2, 0.2];
private static readonly (int Low, int High)[] Bands = BuildBandTable();
private static readonly double[] Window = BuildHannWindow();
private static readonly int[] BitReversal = BuildBitReversal();
private static readonly double[][] TwiddleReal = BuildTwiddles(cosine: true);
private static readonly double[][] TwiddleImag = BuildTwiddles(cosine: false);
/// <summary>
/// Gets the half-open FFT bin range <c>[Low, High)</c> owned by each of the
/// 32 log-spaced bands.
/// </summary>
/// <remarks>
/// Exposed so the golden fixture can assert the table itself rather than
/// only the signature it produces: the band table is the part of the
/// construction most likely to drift between two implementations.
/// </remarks>
/// <returns>One range per band, contiguous and non-overlapping.</returns>
public static IReadOnlyList<(int Low, int High)> BandFftBins() => Bands;
/// <summary>
/// Packs one byte per whole STFT frame: a 5-bit peak band index and a 2-bit
/// energy class.
/// </summary>
/// <remarks>
/// The byte layout is <c>(band &lt;&lt; 2) | class</c>, so bit 7 is always
/// clear and an arbitrary byte is not a valid signature. That structural
/// constraint is what the server validates on upload, and what keeps the
/// field from being usable as a payload channel.
/// </remarks>
/// <param name="mono">Mono samples at <see cref="SampleRate"/>, in [-1, 1).</param>
/// <returns>One byte per frame; empty when not even one frame fits.</returns>
public static byte[] PackFrames(ReadOnlySpan<float> mono)
{
if (mono.Length < FrameSize)
{
return [];
}
var frames = 1 + ((mono.Length - FrameSize) / HopSize);
var binLow = Bands[0].Low;
var binHigh = Bands[NumBands - 1].High; // exclusive
double binCount = binHigh - binLow;
var re = new double[FrameSize];
var im = new double[FrameSize];
var peak = new byte[frames];
var energy = new double[frames];
for (var f = 0; f < frames; f++)
{
var src = mono.Slice(f * HopSize, FrameSize);
for (var n = 0; n < FrameSize; n++)
{
re[n] = src[n] * Window[n];
im[n] = 0.0;
}
Fft(re, im);
// Per-band mean magnitude. The bands tile 300-3000 Hz with no gaps
// and no overlaps, so the frame's band-limited energy is the sum of
// the band sums — accumulated in band order, because the order of a
// floating-point summation is part of the contract.
var best = -1.0;
var bestBand = 0;
var total = 0.0;
for (var b = 0; b < NumBands; b++)
{
var (low, high) = Bands[b];
var sum = 0.0;
for (var k = low; k < high; k++)
{
sum += Math.Sqrt((re[k] * re[k]) + (im[k] * im[k]));
}
total += sum;
var mean = sum / (high - low);
if (mean > best)
{
best = mean; // ties -> lowest index
bestBand = b;
}
}
peak[f] = (byte)bestBand;
energy[f] = total / binCount;
}
// The reference is the upper median of the frame energies: an actually
// observed value rather than the average of the two middle ones, so it
// is bit-reproducible. It is also gain-invariant — loudness
// normalisation must not change a signature — and barely moves when the
// window is trimmed.
var sorted = (double[])energy.Clone();
Array.Sort(sorted);
var reference = sorted[sorted.Length / 2];
var packed = new byte[frames];
for (var f = 0; f < frames; f++)
{
var r = Math.Log10((energy[f] + EnergyEps) / (reference + EnergyEps));
packed[f] = (byte)(((peak[f] & 0x1F) << 2) | (EnergyClass(r) & 0x03));
}
return packed;
}
/// <summary>
/// Computes the full signature string for a decoded centre window.
/// </summary>
/// <param name="mono">Mono samples at <see cref="SampleRate"/>, in [-1, 1).</param>
/// <returns>
/// <c>v1:</c> followed by the base64 of <see cref="PackFrames"/>, or
/// <c>null</c> when not even one frame fits.
/// </returns>
public static string? FromMonoSamples(ReadOnlySpan<float> mono)
{
var packed = PackFrames(mono);
if (packed.Length == 0)
{
return null;
}
// Standard alphabet with '=' padding, which is what the server and the
// pipeline both emit.
return VersionPrefix + Convert.ToBase64String(packed);
}
private static int EnergyClass(double ratio)
{
if (ratio < EnergyClassEdges[0])
{
return 0;
}
if (ratio < EnergyClassEdges[1])
{
return 1;
}
if (ratio < EnergyClassEdges[2])
{
return 2;
}
return 3;
}
// edge[b] = 300 * (3000/300)^(b/32); band b owns FFT bins
// [k_lo[b], k_lo[b+1]) with k_lo[b] = ceil(edge[b] / hz_per_bin). Taking the
// ceiling once, into an integer table, means membership is never decided by
// a float comparison per bin per frame — which is where two implementations
// would otherwise be free to disagree.
private static (int Low, int High)[] BuildBandTable()
{
var hzPerBin = (double)SampleRate / FrameSize;
var edges = new int[NumBands + 1];
for (var b = 0; b <= NumBands; b++)
{
var hz = BandLoHz * Math.Pow(BandHiHz / BandLoHz, (double)b / NumBands);
edges[b] = (int)Math.Ceiling(hz / hzPerBin);
}
var table = new (int Low, int High)[NumBands];
for (var b = 0; b < NumBands; b++)
{
table[b] = (edges[b], edges[b + 1]);
}
return table;
}
private static double[] BuildHannWindow()
{
var w = new double[FrameSize];
for (var n = 0; n < FrameSize; n++)
{
w[n] = 0.5 * (1.0 - Math.Cos(2.0 * Math.PI * n / FrameSize));
}
return w;
}
private static int[] BuildBitReversal()
{
var bits = 0;
while ((1 << bits) < FrameSize)
{
bits++;
}
var rev = new int[FrameSize];
for (var i = 0; i < FrameSize; i++)
{
var r = 0;
for (var b = 0; b < bits; b++)
{
if ((i & (1 << b)) != 0)
{
r |= 1 << (bits - 1 - b);
}
}
rev[i] = r;
}
return rev;
}
// Twiddles are precomputed per stage from cos/sin of -2*pi*j/len, so the
// angle is an exactly reproducible double in either language and only the
// library's own rounding of cos/sin (<= 1 ulp) can differ — orders of
// magnitude below the decision margins the golden fixture records.
private static double[][] BuildTwiddles(bool cosine)
{
var stages = new List<double[]>();
for (var len = 2; len <= FrameSize; len <<= 1)
{
var half = len / 2;
var stage = new double[half];
for (var j = 0; j < half; j++)
{
var angle = -2.0 * Math.PI * j / len;
stage[j] = cosine ? Math.Cos(angle) : Math.Sin(angle);
}
stages.Add(stage);
}
return [.. stages];
}
// Radix-2 decimation-in-time complex FFT, in place, no normalisation.
private static void Fft(double[] re, double[] im)
{
for (var i = 0; i < FrameSize; i++)
{
var j = BitReversal[i];
if (i < j)
{
(re[i], re[j]) = (re[j], re[i]);
(im[i], im[j]) = (im[j], im[i]);
}
}
var stage = 0;
for (var len = 2; len <= FrameSize; len <<= 1, stage++)
{
var half = len / 2;
var wr = TwiddleReal[stage];
var wi = TwiddleImag[stage];
for (var start = 0; start < FrameSize; start += len)
{
for (var j = 0; j < half; j++)
{
var a = start + j;
var b = a + half;
var tr = (re[b] * wr[j]) - (im[b] * wi[j]);
var ti = (re[b] * wi[j]) + (im[b] * wr[j]);
re[b] = re[a] - tr;
im[b] = im[a] - ti;
re[a] += tr;
im[a] += ti;
}
}
}
}
}