Liquid Glass and Apple Music touches for the Spotify iOS app. A GPL-3.0 fork of spoti.pw 0.21.1.
C
fix: Apple Intelligence translates whole songs, each line in its place
My-Name-Is-Jeffb66ce15f2be0
diff --git a/tweak/Sources/Redesigned/Lyrics/SGRKaraokeView.m b/tweak/Sources/Redesigned/Lyrics/SGRKaraokeView.mindex 8ac414fb..19a21dd6 100644--- a/tweak/Sources/Redesigned/Lyrics/SGRKaraokeView.m+++ b/tweak/Sources/Redesigned/Lyrics/SGRKaraokeView.m@@ -1963,21 +1963,22 @@ static BOOL anyUntranslated(NSArray<SGKaraokeLine *> *lines) {SGLyricsSaveTranslation(track, language, lines);}, ^(NSArray<NSString *> *translations, NSString *error) {if (self->_translating == lines) self->_translating = nil;- if (!translations) {- [self offerExtras];- UIAlertController *alert = [UIAlertController alertControllerWithTitle:@"No translation" message:error- preferredStyle:UIAlertControllerStyleAlert];+ // Translations can come with a note, which names the lines left out. Those lines stay untranslated, and+ // asking again sends only them.+ if (translations) {+ [self takeTranslations:translations into:lines];+ SGLyricsSaveTranslation(track, language, lines);+ if (lines == self->_lines) self->_untranslated = anyUntranslated(lines);+ }+ [self offerExtras];+ if (error) {+ UIAlertController *alert = [UIAlertController alertControllerWithTitle:translations ? @"Partly translated" : @"No translation"+ message:error preferredStyle:UIAlertControllerStyleAlert];// Over the player, which is dark whatever the system's appearance.alert.overrideUserInterfaceStyle = UIUserInterfaceStyleDark;[alert addAction:[UIAlertAction actionWithTitle:@"OK" style:UIAlertActionStyleCancel handler:nil]];[SGTopController() presentViewController:alert animated:YES completion:nil];- return;}- [self takeTranslations:translations into:lines];- SGLyricsSaveTranslation(track, language, lines);- // A batch Apple Intelligence turned down stays untranslated, and asking again sends only its lines.- if (lines == self->_lines) self->_untranslated = anyUntranslated(lines);- [self offerExtras];});}diff --git a/tweak/Sources/Shared/LyricsTranslation/OnDeviceTranslation.swift b/tweak/Sources/Shared/LyricsTranslation/OnDeviceTranslation.swiftindex ff1e480d..19afe177 100644--- a/tweak/Sources/Shared/LyricsTranslation/OnDeviceTranslation.swift+++ b/tweak/Sources/Shared/LyricsTranslation/OnDeviceTranslation.swift@@ -119,7 +119,8 @@ public final class SGOnDeviceTranslation: NSObject {guard #available(iOS 26.0, *) else { return finish(done, nil, "Apple Intelligence needs iOS 26.") }let language = name(Locale.Language(identifier: languageTag))// Only the lines given text ("" for those translated already or with no words), 12 to a batch. A batch- // the model refuses or fails is left out and the rest go on, so an explicit verse costs its own lines only.+ // the model refuses is asked again in halves (translateSplitting), and one that fails is left out while+ // the rest go on, so an explicit line costs only itself.let wanted = lines.indices.filter { !lines[$0].isEmpty }Task {var out = [String](repeating: "", count: lines.count)@@ -132,10 +133,12 @@ public final class SGOnDeviceTranslation: NSObject {let chunkStarted = Date()do {let chunk = indices.map { lines[$0] }- let answer = try await translateChunk(chunk, into: language, song: song, before: before)- for (index, translation) in zip(indices, answer) { out[index] = translation }- before = Array(zip(chunk, answer))- translated += indices.count+ let answer = try await translateSplitting(chunk, into: language, song: song, before: before, failure: &failure)+ for (index, translation) in zip(indices, answer) where translation != nil {+ out[index] = translation!+ translated += 1+ }+ before = zip(chunk, answer).compactMap { line, translation in translation.map { (line, $0) } }log.notice("intelligence: \(first + indices.count) of \(wanted.count) lines in \(Date().timeIntervalSince(chunkStarted), format: .fixed(precision: 1)) s")} catch {failure = error@@ -146,10 +149,48 @@ public final class SGOnDeviceTranslation: NSObject {}log.notice("intelligence: \(translated) of \(wanted.count) lines into \(languageTag, privacy: .public) in \(Date().timeIntervalSince(started), format: .fixed(precision: 1)) s")if translated == 0, let failure { return finish(done, nil, problem(failure)) }- finish(done, out, nil)+ let left = wanted.count - translated+ finish(done, out, left == 0 ? nil+ : "Apple Intelligence left \(left) of the song's \(wanted.count) lines untranslated. Ask again to try those lines once more.")+ }+ }++ // A batch, or nil for each line the model refused even alone: a refused batch is asked again in halves, down+ // to single lines. Any other error ends the batch.+ @available(iOS 26.0, *)+ private static func translateSplitting(_ lines: [String], into language: String, song: String?, before: [(String, String)],+ failure: inout Error?) async throws -> [String?] {+ do {+ return try await translateChunk(lines, into: language, song: song, before: before)+ } catch where refused(error) {+ failure = error+ if lines.count == 1 { return [nil] }+ let half = lines.count / 2+ let head = try await translateSplitting(Array(lines[..<half]), into: language, song: song, before: before, failure: &failure)+ let tail = try await translateSplitting(Array(lines[half...]), into: language, song: song, before: before, failure: &failure)+ return head + tail}}+ @available(iOS 26.0, *)+ private static func refused(_ error: Error) -> Bool {+ #if compiler(>=6.4)+ if #available(iOS 27.0, *), let error = error as? LanguageModelError {+ switch error {+ case .guardrailViolation, .refusal: return true+ default: break+ }+ }+ #endif+ if let error = error as? LanguageModelSession.GenerationError {+ switch error {+ case .guardrailViolation, .refusal: return true+ default: break+ }+ }+ return false+ }+// iOS 27 throws LanguageModelError, iOS 26 the session's GenerationError. The first is only in the iOS 27 SDK// (Swift 6.4), and the release build has the iOS 26 one; GenerationOptions' sampling: label is in both.@available(iOS 26.0, *)@@ -183,21 +224,43 @@ public final class SGOnDeviceTranslation: NSObject {// and tone stay as they were; it is context only and gets no answer.@available(iOS 26.0, *)private static func translateChunk(_ lines: [String], into language: String, song: String?, before: [(String, String)]) async throws -> [String] {+ let json = { (value: Any) in+ String(decoding: (try? JSONSerialization.data(withJSONObject: value, options: .fragmentsAllowed)) ?? Data(), as: UTF8.self)+ }+ let context = before.isEmpty ? "" : "Just before, already translated: \(json(before.map { [$0.0, $0.1] }))\n\n"+ let options = GenerationOptions(sampling: .greedy, maximumResponseTokens: tokensPerLine * lines.count + 32)+ let follow = """+ Lines from just before may be given with their translations: they are not to be translated again, only \+ followed, so names, pronouns and tone carry on and a sentence that runs on reads as one.+ """+ // Each line under its own number, in the question and the answer, so a sentence the model joins or splits+ // across lines cannot move the translations after it onto the wrong lines (K-pop's mixed lines did).+ if #available(iOS 26.4, *) {+ let keys = lines.indices.map { String(format: "%02d", $0 + 1) }+ let session = LanguageModelSession(model: model, instructions: """+ You translate the lyrics of \(song ?? "a song") into \(language). You get a JSON object of numbered \+ lines and answer with the same numbers, each holding the translation of that line alone, natural as \+ sung, keeping the meaning of slang and idioms. A line that is empty, a sound or already in \(language) \+ is given back as it is. \(follow)+ """)+ let root = DynamicGenerationSchema(name: "Lines", properties: keys.map {+ DynamicGenerationSchema.Property(name: $0, schema: DynamicGenerationSchema(type: String.self))+ })+ let numbered = "{" + zip(keys, lines).map { "\(json($0)): \(json($1))" }.joined(separator: ", ") + "}"+ let content = try await session.respond(to: context + "Translate: " + numbered,+ schema: try GenerationSchema(root: root, dependencies: []), options: options).content+ return try keys.map { try content.value(String.self, forProperty: $0) }+ }let session = LanguageModelSession(model: model, instructions: """You translate the lyrics of \(song ?? "a song") into \(language). You get a JSON array of lines and answer \with an array of the same length: each line's translation at the same place, natural as sung, keeping the \meaning of slang and idioms. A line that is empty, a sound or already in \(language) is given back as it is. \- Lines from just before may be given with their translations: they are not to be translated again, only \- followed, so names, pronouns and tone carry on and a sentence that runs on reads as one.+ \(follow)""")let line = DynamicGenerationSchema(type: String.self)let schema = try GenerationSchema(root: DynamicGenerationSchema(arrayOf: line, minimumElements: lines.count, maximumElements: lines.count),dependencies: [])- let json = { (value: Any) in String(decoding: (try? JSONSerialization.data(withJSONObject: value)) ?? Data(), as: UTF8.self) }- let prompt = before.isEmpty ? json(lines)- : "Just before, already translated: \(json(before.map { [$0.0, $0.1] }))\n\nTranslate: \(json(lines))"- let options = GenerationOptions(sampling: .greedy, maximumResponseTokens: tokensPerLine * lines.count + 32)- let answer = try await session.respond(to: prompt, schema: schema, options: options).content.value([String].self)+ let answer = try await session.respond(to: context + "Translate: " + json(lines), schema: schema, options: options).content.value([String].self)return answer.count == lines.count ? answer : lines.indices.map { $0 < answer.count ? answer[$0] : "" }}}