feat: add ModelStorage and ModelCatalog; transcriber uses ModelStorage

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
2026-10-01 15:33:22 +02:00
parent fbc90cbd12
commit ae20008679
3 changed files with 149 additions and 34 deletions
@@ -0,0 +1,73 @@
+import Foundation
+
+/// A speech model the user can pick in Settings.
+public struct ModelOption: Identifiable, Equatable, Sendable {
+    public let id: String          // WhisperKit variant name
+    public let title: String
+    public let detail: String
+    public let approxSizeMB: Int
+
+    public init(id: String, title: String, detail: String, approxSizeMB: Int) {
+        self.id = id
+        self.title = title
+        self.detail = detail
+        self.approxSizeMB = approxSizeMB
+    }
+}
+
+public enum ModelCatalog {
+    public static let defaultVariant = "large-v3-v20240930_turbo"
+
+    /// Sizes measured from the Hugging Face repository `argmaxinc/whisperkit-coreml`.
+    public static let all: [ModelOption] = [
+        ModelOption(id: defaultVariant, title: "Whisper Large v3 Turbo", detail: "Aanbevolen · nauwkeurigst", approxSizeMB: 1638),
+        ModelOption(id: "large-v3-v20240930_turbo_632MB", title: "Whisper Large v3 Turbo (compact)", detail: "Kleiner, iets minder nauwkeurig", approxSizeMB: 646),
+        ModelOption(id: "small", title: "Whisper Small", detail: "Snel en licht, minder nauwkeurig", approxSizeMB: 486),
+    ]
+}
+
+/// Where models live on disk: `<base>/models/argmaxinc/whisperkit-coreml/openai_whisper-<variant>`.
+public struct ModelStorage: Sendable {
+    public static let repo = "argmaxinc/whisperkit-coreml"
+    public static let requiredModels = ["MelSpectrogram", "AudioEncoder", "TextDecoder"]
+
+    public let base: URL
+
+    public init(base: URL = ModelStorage.defaultBase) {
+        self.base = base
+    }
+
+    /// Application Support, so models are neither in ~/Documents (privacy prompt, iCloud) nor re-fetched.
+    public static var defaultBase: URL {
+        FileManager.default.urls(for: .applicationSupportDirectory, in: .userDomainMask)[0]
+            .appending(path: "Sst/models", directoryHint: .isDirectory)
+    }
+
+    var repoDirectory: URL {
+        base.appending(path: "models/\(Self.repo)", directoryHint: .isDirectory)
+    }
+
+    private func folders(for variant: String) -> [URL] {
+        let names = (try? FileManager.default.contentsOfDirectory(atPath: repoDirectory.path)) ?? []
+        return names
+            .filter { $0 == variant || $0 == "openai_whisper-\(variant)" }
+            .map { repoDirectory.appending(path: $0, directoryHint: .isDirectory) }
+    }
+
+    /// The folder of a completely downloaded variant, or nil.
+    public func folder(for variant: String) -> URL? {
+        folders(for: variant).first { folder in
+            Self.requiredModels.allSatisfy {
+                FileManager.default.fileExists(atPath: folder.appending(path: "\($0).mlmodelc/coremldata.bin").path)
+            }
+        }
+    }
+
+    public func isInstalled(_ variant: String) -> Bool {
+        folder(for: variant) != nil
+    }
+
+    public func delete(_ variant: String) throws {
+        for folder in folders(for: variant) { try FileManager.default.removeItem(at: folder) }
+    }
+}