Ускорить локальную диктовку стабильным префиксом

Sources/WhisperKey/Audio.swift

@@ -6,6 +6,11 @@ struct RecordedChunk: Sendable {
     let continuesPhrase: Bool
 }
 
+enum AudioChunking: Sendable {
+    case stablePreviews
+    case committed
+}
+
 final class AudioRecorder: @unchecked Sendable {
     private var engine: AVAudioEngine?
     private let lock = NSLock()
@@ -21,8 +26,10 @@ final class AudioRecorder: @unchecked Sendable {
     private var onChunk: (@Sendable (RecordedChunk) -> Void)?
     private var onSilence: (@Sendable () -> Void)?
     private var recordingFormat: AVAudioFormat?
+    private var chunking = AudioChunking.committed
 
     func start(
+        chunking: AudioChunking,
         onChunk: @escaping @Sendable (RecordedChunk) -> Void,
         onSilence: @escaping @Sendable () -> Void
     ) throws {
@@ -33,6 +40,7 @@ final class AudioRecorder: @unchecked Sendable {
         let format = input.outputFormat(forBus: 0)
         self.onChunk = onChunk
         self.onSilence = onSilence
+        self.chunking = chunking
         recordingFormat = format
         try beginChunk(format: format)
         input.installTap(onBus: 0, bufferSize: 4096, format: format) { [weak self] buffer, _ in
@@ -130,15 +138,22 @@ final class AudioRecorder: @unchecked Sendable {
             return
         }
 
-        // Keep enough context for Whisper, but insert completed phrases while
-        // the user is still dictating. Prefer a short natural pause; cap a
-        // continuous phrase so the input never stays blank indefinitely.
-        let minimumChunk = AVAudioFramePosition(format.sampleRate * 2.5)
-        let pauseBoundary = AVAudioFramePosition(format.sampleRate * 1.0)
-        let maximumChunk = AVAudioFramePosition(format.sampleRate * 4.0)
-        let reachedPause =
-            currentFrames >= minimumChunk && chunkSilentFrames >= pauseBoundary
-        let reachedMaximum = currentFrames >= maximumChunk
+        let reachedPause: Bool
+        let reachedMaximum: Bool
+        switch chunking {
+        case .stablePreviews:
+            reachedPause = false
+            reachedMaximum = currentFrames >= AVAudioFramePosition(format.sampleRate * 1.25)
+        case .committed:
+            // Cloud dictation keeps the old larger chunks so it does not issue
+            // overlapping paid requests.
+            let minimumChunk = AVAudioFramePosition(format.sampleRate * 2.5)
+            let pauseBoundary = AVAudioFramePosition(format.sampleRate * 1.0)
+            let maximumChunk = AVAudioFramePosition(format.sampleRate * 4.0)
+            reachedPause =
+                currentFrames >= minimumChunk && chunkSilentFrames >= pauseBoundary
+            reachedMaximum = currentFrames >= maximumChunk
+        }
         let shouldRotate = currentChunkHasSpeech && (reachedPause || reachedMaximum)
 
         if shouldRotate, let finishedURL = currentURL {
@@ -174,6 +189,13 @@ final class AudioRecorder: @unchecked Sendable {
 
 enum AudioConverter {
     static func toWhisperWAV(_ input: URL) throws -> URL {
+        try toWhisperWAV([input])
+    }
+
+    static func toWhisperWAV(_ inputs: [URL]) throws -> URL {
+        guard !inputs.isEmpty else {
+            throw WhisperKeyError.processFailed("Нет аудиофрагментов для распознавания.")
+        }
         let output = FileManager.default.temporaryDirectory.appendingPathComponent("whisperkey-\(UUID().uuidString).wav")
         let ffmpegCandidates = ["/opt/homebrew/bin/ffmpeg", "/usr/local/bin/ffmpeg"]
         guard let ffmpeg = ffmpegCandidates.first(where: FileManager.default.isExecutableFile(atPath:)) else {
@@ -181,7 +203,26 @@ enum AudioConverter {
         }
         let process = Process()
         process.executableURL = URL(fileURLWithPath: ffmpeg)
-        process.arguments = ["-hide_banner", "-loglevel", "error", "-y", "-i", input.path, "-ar", "16000", "-ac", "1", "-c:a", "pcm_s16le", output.path]
+        var arguments = ["-hide_banner", "-loglevel", "error", "-y"]
+        for input in inputs {
+            arguments.append(contentsOf: ["-i", input.path])
+        }
+        if inputs.count > 1 {
+            let streams = inputs.indices.map { "[\($0):a:0]" }.joined()
+            arguments.append(contentsOf: [
+                "-filter_complex",
+                "\(streams)concat=n=\(inputs.count):v=0:a=1[out]",
+                "-map",
+                "[out]",
+            ])
+        }
+        arguments.append(contentsOf: [
+            "-ar", "16000",
+            "-ac", "1",
+            "-c:a", "pcm_s16le",
+            output.path,
+        ])
+        process.arguments = arguments
         let stderr = Pipe()
         process.standardError = stderr
         try process.run()