346 lines
12 KiB
Go
346 lines
12 KiB
Go
// Package subsync fixes the timing of a subtitle by aligning it to one that is already
|
||
// right.
|
||
//
|
||
// The idea is borrowed from ffsubsync and alass, which AutoSubSync drives — but not the
|
||
// code, which is Python and Rust and needs FFmpeg to read a film's audio. This gateway's
|
||
// image is distroless with no ffmpeg and no shell, so the audio path is closed to it. What
|
||
// is open is the cheaper half of the same idea: both of those tools can align against a
|
||
// *reference subtitle* instead of audio, and a household's copy of a film usually carries
|
||
// a track that is already correct. Reduce both tracks to "somebody is speaking / nobody
|
||
// is" on a fixed time grid and the offset is the shift where the two agree most.
|
||
//
|
||
// Nothing here touches the network, Emby or the database, which is what lets the whole
|
||
// rule be tested against real subtitle text.
|
||
package subsync
|
||
|
||
import (
|
||
"fmt"
|
||
"sort"
|
||
"strings"
|
||
"time"
|
||
)
|
||
|
||
// Cue is one subtitle line: when it appears, when it goes, and what it says.
|
||
//
|
||
// The text is carried through untouched. This package changes *when* a line is shown and
|
||
// never what it says — a resynchronised subtitle that had also been reflowed or re-escaped
|
||
// would be impossible to tell from a corrupted one.
|
||
type Cue struct {
|
||
Start time.Duration
|
||
End time.Duration
|
||
Text string
|
||
}
|
||
|
||
// Duration is where the last cue ends, which stands in for the runtime the track was
|
||
// written against.
|
||
func Duration(cues []Cue) time.Duration {
|
||
var last time.Duration
|
||
for _, cue := range cues {
|
||
if cue.End > last {
|
||
last = cue.End
|
||
}
|
||
}
|
||
return last
|
||
}
|
||
|
||
// Shift moves every cue by the correction Align found.
|
||
//
|
||
// Scale is applied before the offset, in that order, because that is the order the damage
|
||
// happened in: a subtitle written for a 25fps transfer and played against a 23.976 one
|
||
// runs progressively further out, and whatever fixed offset remains sits on top of the
|
||
// stretch. A cue dragged before zero is clamped rather than dropped — it belongs to a line
|
||
// somebody is about to hear, and a track that silently lost its opening would look like a
|
||
// worse fault than the one being fixed.
|
||
func Shift(cues []Cue, offset time.Duration, scale float64) []Cue {
|
||
if scale <= 0 {
|
||
scale = 1
|
||
}
|
||
out := make([]Cue, 0, len(cues))
|
||
for _, cue := range cues {
|
||
start := scaleDuration(cue.Start, scale) + offset
|
||
end := scaleDuration(cue.End, scale) + offset
|
||
if end < 0 {
|
||
// The whole cue was pushed off the front. Keep it at zero rather than
|
||
// discarding it: losing a line is worse than showing one early.
|
||
start, end = 0, 0
|
||
} else if start < 0 {
|
||
start = 0
|
||
}
|
||
out = append(out, Cue{Start: start, End: end, Text: cue.Text})
|
||
}
|
||
sort.SliceStable(out, func(i, j int) bool { return out[i].Start < out[j].Start })
|
||
return out
|
||
}
|
||
|
||
func scaleDuration(d time.Duration, scale float64) time.Duration {
|
||
return time.Duration(float64(d) * scale)
|
||
}
|
||
|
||
// FrameRateScales are the stretch factors tried alongside a plain offset.
|
||
//
|
||
// They are the ratios between the frame rates films are actually delivered at, because
|
||
// the classic broken subtitle is not merely late — it is a track authored against one
|
||
// transfer and played against another, so it drifts, and no single offset can fix it. The
|
||
// list is deliberately short: every extra candidate is another chance for a wrong answer
|
||
// to win by luck, and these cover the transfers that exist.
|
||
var FrameRateScales = []float64{
|
||
1,
|
||
23.976 / 24, 24 / 23.976,
|
||
25.0 / 24, 24.0 / 25,
|
||
25 / 23.976, 23.976 / 25,
|
||
30 / 29.97, 29.97 / 30,
|
||
}
|
||
|
||
// Options tunes the search. The zero value is not usable; use DefaultOptions.
|
||
type Options struct {
|
||
// Bin is the width of one cell of the time grid. Fine enough that a correction is
|
||
// worth making at all, coarse enough that two tracks written by different people
|
||
// still land in the same cells.
|
||
Bin time.Duration
|
||
// MaxShift bounds the search either way. A subtitle for a different release can be a
|
||
// minute out; beyond that the two are far more likely to be different cuts of the
|
||
// film, where a confident answer would be a wrong one.
|
||
MaxShift time.Duration
|
||
// MinCues is how much evidence is needed on each side before an answer is offered.
|
||
MinCues int
|
||
// MinScore is the share of the shorter track's speech that must line up.
|
||
MinScore float64
|
||
// MinMargin is how far the winning shift must beat the best rival outside its own
|
||
// peak. This is what separates a real alignment from a track that correlates weakly
|
||
// with everything, which is what a wrong reference looks like.
|
||
MinMargin float64
|
||
// Scales are the stretch candidates tried. Empty means offset only.
|
||
Scales []float64
|
||
}
|
||
|
||
// DefaultOptions is the tuning the gateway uses.
|
||
func DefaultOptions() Options {
|
||
return Options{
|
||
Bin: 50 * time.Millisecond,
|
||
MaxShift: 60 * time.Second,
|
||
MinCues: 20,
|
||
MinScore: 0.45,
|
||
MinMargin: 0.10,
|
||
Scales: FrameRateScales,
|
||
}
|
||
}
|
||
|
||
// Result describes a correction.
|
||
type Result struct {
|
||
// Offset is what to add to every cue, after Scale.
|
||
Offset time.Duration
|
||
// Scale is the stretch applied first. 1 means the timing was merely displaced.
|
||
Scale float64
|
||
// Score is the share of the shorter track's speech that lines up once corrected —
|
||
// how much the two agree, not how confident we are.
|
||
Score float64
|
||
// Margin is how far ahead of the best rival alignment this one finished. A high
|
||
// score with no margin is a track that matches everything, which matches nothing.
|
||
Margin float64
|
||
}
|
||
|
||
// Correction reports whether this result actually changes anything a viewer would see.
|
||
func (r Result) Correction() bool {
|
||
return r.Offset.Abs() >= 100*time.Millisecond || r.Scale != 1
|
||
}
|
||
|
||
// String renders the correction the way the console and the log want it.
|
||
func (r Result) String() string {
|
||
seconds := r.Offset.Seconds()
|
||
out := fmt.Sprintf("%+.2fs", seconds)
|
||
if r.Scale != 1 {
|
||
out += fmt.Sprintf(" at %.4f×", r.Scale)
|
||
}
|
||
return out
|
||
}
|
||
|
||
// ErrNoAlignment is returned when the two tracks cannot be aligned with confidence.
|
||
//
|
||
// Most of this package is about producing this rather than a number. A subtitle nudged to
|
||
// the wrong place is worse than one left alone: the viewer asked for a fix, would be told
|
||
// it worked, and would have no way of knowing the file they now have is further out than
|
||
// the one they started with.
|
||
type ErrNoAlignment struct {
|
||
// Reason is safe to show a viewer.
|
||
Reason string
|
||
// Score and Margin are what the search actually found, for the log.
|
||
Score float64
|
||
Margin float64
|
||
}
|
||
|
||
func (e *ErrNoAlignment) Error() string {
|
||
return fmt.Sprintf("subsync: %s (score %.2f, margin %.2f)", e.Reason, e.Score, e.Margin)
|
||
}
|
||
|
||
// Align finds the correction that brings broken onto reference.
|
||
//
|
||
// The reference is assumed correct; nothing here checks that, because nothing could. The
|
||
// caller chooses it, and choosing badly is the one failure this package cannot detect —
|
||
// which is why the margin test exists, since a wrong reference tends to match everything
|
||
// equally rather than matching one shift particularly well.
|
||
func Align(broken, reference []Cue, opts Options) (Result, error) {
|
||
if opts.Bin <= 0 {
|
||
opts = DefaultOptions()
|
||
}
|
||
if len(broken) < opts.MinCues || len(reference) < opts.MinCues {
|
||
return Result{}, &ErrNoAlignment{
|
||
Reason: "there are not enough lines in one of these subtitles to compare them",
|
||
}
|
||
}
|
||
|
||
scales := opts.Scales
|
||
if len(scales) == 0 {
|
||
scales = []float64{1}
|
||
}
|
||
|
||
// The reference is rasterised once; the broken track is rasterised per stretch
|
||
// candidate, which is a handful of times rather than once per shift.
|
||
ref := rasterise(reference, opts.Bin, 1)
|
||
maxShiftBins := int(opts.MaxShift / opts.Bin)
|
||
|
||
best := Result{Scale: 1}
|
||
bestBins := 0
|
||
found := false
|
||
// Rivals holds the best score at each shift far enough from the winner to be a
|
||
// different answer rather than the same peak's shoulder.
|
||
var runnerUp float64
|
||
|
||
for _, scale := range scales {
|
||
signal := rasterise(broken, opts.Bin, scale)
|
||
if signal.on == 0 {
|
||
continue
|
||
}
|
||
floor := min(signal.on, ref.on)
|
||
if floor == 0 {
|
||
continue
|
||
}
|
||
for shift := -maxShiftBins; shift <= maxShiftBins; shift++ {
|
||
score := float64(signal.overlap(ref, shift)) / float64(floor)
|
||
switch {
|
||
case !found || score > best.Score:
|
||
// The old winner becomes a rival only if it is a different answer.
|
||
if found && farApart(bestBins, shift, best.Scale, scale, opts.Bin) {
|
||
runnerUp = max(runnerUp, best.Score)
|
||
}
|
||
found = true
|
||
best = Result{
|
||
Offset: time.Duration(shift) * opts.Bin,
|
||
Scale: scale,
|
||
Score: score,
|
||
}
|
||
bestBins = shift
|
||
case farApart(bestBins, shift, best.Scale, scale, opts.Bin):
|
||
runnerUp = max(runnerUp, score)
|
||
}
|
||
}
|
||
}
|
||
|
||
if !found {
|
||
return Result{}, &ErrNoAlignment{Reason: "these subtitles have no speech in common"}
|
||
}
|
||
best.Margin = best.Score - runnerUp
|
||
if best.Score < opts.MinScore {
|
||
return Result{}, &ErrNoAlignment{
|
||
Reason: "these two subtitles are too different to line up — they may be for " +
|
||
"different cuts of this title",
|
||
Score: best.Score, Margin: best.Margin,
|
||
}
|
||
}
|
||
if best.Margin < opts.MinMargin {
|
||
return Result{}, &ErrNoAlignment{
|
||
Reason: "no single timing fits better than the others, so the result would be " +
|
||
"a guess",
|
||
Score: best.Score, Margin: best.Margin,
|
||
}
|
||
}
|
||
return best, nil
|
||
}
|
||
|
||
// farApart says whether two candidates are different answers rather than two samples of
|
||
// one peak. Alignment produces a broad hill either side of the true shift, so the shifts
|
||
// immediately around the winner are not rivals — treating them as rivals would collapse
|
||
// every margin to nearly zero and refuse every correct answer.
|
||
func farApart(bestBins, shift int, bestScale, scale float64, bin time.Duration) bool {
|
||
if bestScale != scale {
|
||
return true
|
||
}
|
||
guard := int(time.Second / bin)
|
||
return abs(shift-bestBins) > guard
|
||
}
|
||
|
||
func abs(v int) int {
|
||
if v < 0 {
|
||
return -v
|
||
}
|
||
return v
|
||
}
|
||
|
||
// Reference picks which of the tracks on offer to align against.
|
||
//
|
||
// The rules are about what makes a *usable* yardstick, not what makes a good subtitle. A
|
||
// forced track carries only the lines that are foreign to the film's own audio, so it is
|
||
// mostly silence and would correlate with almost any shift; a track with very few cues is
|
||
// the same problem in a different shape. Longest wins because coverage is what the
|
||
// correlation is measured against. It returns the index so the caller can name the track
|
||
// it used, which is the one thing a viewer needs to judge the answer.
|
||
func Reference(candidates [][]Cue, exclude int, forced []bool, minCues int) (int, bool) {
|
||
best, bestCount := -1, 0
|
||
for i, cues := range candidates {
|
||
if i == exclude || len(cues) < minCues {
|
||
continue
|
||
}
|
||
if i < len(forced) && forced[i] {
|
||
continue
|
||
}
|
||
if len(cues) > bestCount {
|
||
best, bestCount = i, len(cues)
|
||
}
|
||
}
|
||
return best, best >= 0
|
||
}
|
||
|
||
// rasterise turns cues into one bit per time bin: is anybody speaking in this cell.
|
||
//
|
||
// The text is thrown away deliberately. Two subtitles for one film are in different
|
||
// languages as often as not, so the only thing they can be compared on is *when* lines
|
||
// happen — which turns out to be plenty, because dialogue rhythm is a property of the
|
||
// film rather than of the translation.
|
||
func rasterise(cues []Cue, bin time.Duration, scale float64) *signal {
|
||
end := time.Duration(0)
|
||
for _, cue := range cues {
|
||
if e := scaleDuration(cue.End, scale); e > end {
|
||
end = e
|
||
}
|
||
}
|
||
sig := newSignal(int(end/bin) + 2)
|
||
for _, cue := range cues {
|
||
start := scaleDuration(cue.Start, scale)
|
||
stop := scaleDuration(cue.End, scale)
|
||
if stop <= start {
|
||
// A zero-length or reversed cue still says a line happened here.
|
||
stop = start + bin
|
||
}
|
||
sig.set(int(start/bin), int(stop/bin))
|
||
}
|
||
return sig
|
||
}
|
||
|
||
// Normalise tidies a parsed track before it is compared or written back.
|
||
//
|
||
// Cues out of order, or overlapping, are ordinary in files people have edited by hand, and
|
||
// both would put speech in the wrong cell.
|
||
func Normalise(cues []Cue) []Cue {
|
||
out := make([]Cue, 0, len(cues))
|
||
for _, cue := range cues {
|
||
if strings.TrimSpace(cue.Text) == "" {
|
||
continue
|
||
}
|
||
if cue.End < cue.Start {
|
||
cue.Start, cue.End = cue.End, cue.Start
|
||
}
|
||
out = append(out, cue)
|
||
}
|
||
sort.SliceStable(out, func(i, j int) bool { return out[i].Start < out[j].Start })
|
||
return out
|
||
}
|