fix: CJK 转写分段拼接修复

Sources/Speech/TranscriptSegmentJoiner.swift

@@ -0,0 +1,39 @@
+import Foundation
+
+/// Joins the per-chunk texts WhisperKit returns when VAD chunking splits a
+/// recording. Latin scripts need a separating space; CJK scripts must not get
+/// one. When the language is auto-detected (`nil`), fall back to inspecting the
+/// boundary characters so CJK output still joins without spaces.
+enum TranscriptSegmentJoiner {
+    static func joined(_ segments: [String], language: String?) -> String {
+        let cleaned = segments
+            .map { $0.trimmingCharacters(in: .whitespacesAndNewlines) }
+            .filter { !$0.isEmpty }
+        guard var result = cleaned.first else { return "" }
+
+        for segment in cleaned.dropFirst() {
+            result += separator(previous: result, next: segment, language: language)
+            result += segment
+        }
+        return result
+    }
+
+    static func separator(previous: String, next: String, language: String?) -> String {
+        if let language, !language.isEmpty {
+            return usesNoSpaceScript(language) ? "" : " "
+        }
+        if let last = previous.last, let first = next.first,
+           isNoSpaceBoundary(last) || isNoSpaceBoundary(first) {
+            return ""
+        }
+        return " "
+    }
+
+    private static func usesNoSpaceScript(_ language: String) -> Bool {
+        ["zh", "yue", "ja", "ko"].contains(language.lowercased())
+    }
+
+    private static func isNoSpaceBoundary(_ character: Character) -> Bool {
+        character.unicodeScalars.contains(where: NoSpaceScript.contains)
+    }
+}

Sources/Speech/WhisperEngine.swift

@@ -117,6 +117,7 @@ final class WhisperEngine: SpeechEngine, @unchecked Sendable {
         )
         streamingSession = WhisperStreamingSession(
             whisperKit: whisperKit,
+            language: language,
             partialHandler: onPartialResult,
             optionsBuilder: { options }
         )
@@ -164,10 +165,10 @@ final class WhisperEngine: SpeechEngine, @unchecked Sendable {
             audioPath: url.path,
             decodeOptions: options
         )
-        let text = results
-            .compactMap { $0.text }
-            .joined(separator: " ")
-            .trimmingCharacters(in: .whitespacesAndNewlines)
+        let text = TranscriptSegmentJoiner.joined(
+            results.compactMap { $0.text },
+            language: language
+        )
 
         let elapsed = CFAbsoluteTimeGetCurrent() - t0
         Log.info("[WhisperEngine] transcribed \(text.count) chars in \(String(format: "%.1f", elapsed))s")

Sources/Speech/WhisperStreamingSession.swift

@@ -4,6 +4,7 @@ import WhisperKit
 
 final class WhisperStreamingSession: @unchecked Sendable {
     private let whisperKit: WhisperKit
+    private let language: String?
     private let partialHandler: @Sendable (String) -> Void
     private let optionsBuilder: () -> DecodingOptions
     private let queue = DispatchQueue(label: "opentype.whisper-stream")
@@ -23,10 +24,12 @@ final class WhisperStreamingSession: @unchecked Sendable {
 
     init(
         whisperKit: WhisperKit,
+        language: String? = nil,
         partialHandler: @escaping @Sendable (String) -> Void,
         optionsBuilder: @escaping () -> DecodingOptions
     ) {
         self.whisperKit = whisperKit
+        self.language = language
         self.partialHandler = partialHandler
         self.optionsBuilder = optionsBuilder
     }
@@ -134,10 +137,10 @@ final class WhisperStreamingSession: @unchecked Sendable {
                 audioArray: snapshot,
                 decodeOptions: optionsBuilder()
             )
-            let text = results
-                .compactMap(\.text)
-                .joined(separator: " ")
-                .trimmingCharacters(in: .whitespacesAndNewlines)
+            let text = TranscriptSegmentJoiner.joined(
+                results.compactMap(\.text),
+                language: self?.language
+            )
             await self?.finishPartialTask(text: text, submittedSampleCount: submittedSampleCount)
             return text
         }

Tests/OpenTypeTests/SpeechRecognitionQualityTests.swift

@@ -106,6 +106,21 @@ final class SpeechRecognitionQualityTests: XCTestCase {
         XCTAssertEqual(prompt, "Dictation terms: OpenType, MLX.")
     }
 
+    func testWhisperPromptTokensInjectTermsWhenLanguageIsAuto() {
+        let context = SpeechRecognitionContext(phrases: ["OpenType", "MLX"])
+
+        let tokens = context.whisperPromptTokens(
+            language: nil,
+            maximumCount: 160,
+            tokenize: { Array($0.utf8).map(Int.init) }
+        )
+        let prompt = tokens.map { String(decoding: $0.map(UInt8.init), as: UTF8.self) }
+
+        XCTAssertNotNil(tokens)
+        XCTAssertFalse(tokens?.isEmpty ?? true)
+        XCTAssertEqual(prompt, "Dictation terms: OpenType, MLX.")
+    }
+
     func testAppleCompatibleDictationPresetChangesAfterOneMinute() {
         XCTAssertEqual(
             AppleSpeechAnalyzer.dictationPreset(forDuration: 30),

Tests/OpenTypeTests/TranscriptSegmentJoinerTests.swift

@@ -0,0 +1,46 @@
+import XCTest
+@testable import OpenType
+
+final class TranscriptSegmentJoinerTests: XCTestCase {
+    func testCJKLanguageJoinsChunkSegmentsWithoutSpaces() {
+        XCTAssertEqual(
+            TranscriptSegmentJoiner.joined(["你好", "世界"], language: "zh"),
+            "你好世界"
+        )
+        XCTAssertEqual(
+            TranscriptSegmentJoiner.joined(["こんにちは", "世界"], language: "ja"),
+            "こんにちは世界"
+        )
+        XCTAssertEqual(
+            TranscriptSegmentJoiner.joined(["안녕하세요", "세계"], language: "ko"),
+            "안녕하세요세계"
+        )
+    }
+
+    func testLatinLanguageKeepsSpaceBetweenChunkSegments() {
+        XCTAssertEqual(
+            TranscriptSegmentJoiner.joined(["hello", "world"], language: "en"),
+            "hello world"
+        )
+    }
+
+    func testAutoLanguageFallsBackToBoundaryCharacters() {
+        XCTAssertEqual(
+            TranscriptSegmentJoiner.joined(["你好", "世界"], language: nil),
+            "你好世界"
+        )
+        XCTAssertEqual(
+            TranscriptSegmentJoiner.joined(["hello", "world"], language: nil),
+            "hello world"
+        )
+    }
+
+    func testTrimsSegmentsAndDropsEmpties() {
+        XCTAssertEqual(
+            TranscriptSegmentJoiner.joined(["  你好 ", "", " 世界  "], language: "zh"),
+            "你好世界"
+        )
+        XCTAssertEqual(TranscriptSegmentJoiner.joined([], language: "zh"), "")
+        XCTAssertEqual(TranscriptSegmentJoiner.joined(["", "   "], language: "en"), "")
+    }
+}