macOS 13 support

.github/workflows/unit-tests.yml

@@ -9,7 +9,15 @@ on:
 
 jobs:
  build-and-test:
- runs-on: macos-14
+ strategy:
+ matrix:
+ os: [macos-13-xlarge, macos-14]
+ include:
+ - os: macos-13-xlarge
+ ios-version: "16.1" # oldest available version
+ - os: macos-14
+ ios-version: "17.2" # latest available version
+ runs-on: ${{ matrix.os }}
  steps:
  - uses: actions/checkout@v4
  - uses: maxim-lobanov/setup-xcode@v1
@@ -40,14 +48,16 @@ jobs:
  run: |
  set -o pipefail
  xcodebuild clean build-for-testing -scheme whisperkit-Package -destination generic/platform=iOS | xcpretty
- xcodebuild test -only-testing WhisperKitTests/UnitTests -scheme whisperkit-Package -destination "platform=iOS Simulator,OS=17.2,name=iPhone 15" | xcpretty
+ xcodebuild test -only-testing WhisperKitTests/UnitTests -scheme whisperkit-Package -destination "platform=iOS Simulator,OS=${{ matrix.ios-version }},name=iPhone 15" | xcpretty
  - name: Build and Test - watchOS
+ if: matrix.os == 'macos-14'
  run: |
  set -o pipefail
  xcodebuild clean build-for-testing -scheme whisperkit-Package -destination generic/platform=watchOS | xcpretty
  xcodebuild test -only-testing WhisperKitTests/UnitTests -scheme whisperkit-Package -destination "platform=watchOS Simulator,OS=10.2,name=Apple Watch Ultra 2 (49mm)" | xcpretty
  - name: Build and Test - visionOS
+ if: matrix.os == 'macos-14'
  run: |
  set -o pipefail
  xcodebuild clean build-for-testing -scheme whisperkit-Package -destination generic/platform=visionOS | xcpretty
- xcodebuild test -only-testing WhisperKitTests/UnitTests -scheme whisperkit-Package -destination "platform=visionOS Simulator,name=Apple Vision Pro" | xcpretty
+ xcodebuild test -only-testing WhisperKitTests/UnitTests -scheme whisperkit-Package -destination "platform=visionOS Simulator,name=Apple Vision Pro" | xcpretty

Package.swift

@@ -6,10 +6,8 @@ import PackageDescription
 let package = Package(
  name: "whisperkit",
  platforms: [
- .iOS(.v17),
- .macOS(.v14),
- .watchOS(.v10),
- .visionOS(.v1)
+ .iOS(.v16),
+ .macOS(.v13),
  ],
  products: [
  .library(
@@ -18,7 +16,8 @@ let package = Package(
  ),
  .executable(
  name: "transcribe",
- targets: ["WhisperKitCLI"])
+ targets: ["WhisperKitCLI"]
+ ),
  ],
  dependencies: [
  .package(url: "https://github.com/huggingface/swift-transformers.git", exact: "0.1.2"),
@@ -35,7 +34,7 @@ let package = Package(
  name: "WhisperKitCLI",
  dependencies: [
  "WhisperKit",
- .product(name: "ArgumentParser", package: "swift-argument-parser")
+ .product(name: "ArgumentParser", package: "swift-argument-parser"),
  ]
  ),
  .testTarget(
@@ -51,11 +50,11 @@ let package = Package(
  "Makefile",
  "README.md",
  "LICENSE",
- "CONTRIBUTING.md"
+ "CONTRIBUTING.md",
  ],
  resources: [
  .process("Tests/WhisperKitTests/Resources"),
- .copy("Models/whisperkit-coreml")
+ .copy("Models/whisperkit-coreml"),
  ]
  ),
  ]

Sources/WhisperKit/Core/AudioEncoder.swift

@@ -14,7 +14,7 @@ public protocol AudioEncoding {
  func encodeFeatures(_ features: MLMultiArray) async throws -> MLMultiArray?
 }
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class AudioEncoder: AudioEncoding, WhisperMLModel {
  public var model: MLModel?
 

Sources/WhisperKit/Core/AudioProcessor.swift

@@ -131,7 +131,7 @@ public extension AudioProcessing {
  }
 }
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class AudioProcessor: NSObject, AudioProcessing {
  public var audioEngine: AVAudioEngine?
  public var audioSamples: ContiguousArray<Float> = []
@@ -302,7 +302,32 @@ public class AudioProcessor: NSObject, AudioProcessing {
  }
 
  public static func requestRecordPermission() async -> Bool {
- await AVAudioApplication.requestRecordPermission()
+ if #available(macOS 14, iOS 17, *) {
+ return await AVAudioApplication.requestRecordPermission()
+ } else {
+ #if os(watchOS)
+ // watchOS does not support AVCaptureDevice
+ return true
+ #else
+ let microphoneStatus = AVCaptureDevice.authorizationStatus(for: .audio)
+ switch microphoneStatus {
+ case .notDetermined:
+ return await withCheckedContinuation { continuation in
+ AVCaptureDevice.requestAccess(for: .audio) { granted in
+ continuation.resume(returning: granted)
+ }
+ }
+ case .restricted, .denied:
+ Logging.error("Microphone access denied")
+ return false
+ case .authorized:
+ return true
+ @unknown default:
+ Logging.error("Unknown authorization status")
+ return false
+ }
+ #endif
+ }
  }
 
  deinit {
@@ -312,7 +337,7 @@ public class AudioProcessor: NSObject, AudioProcessing {
 
 // MARK: - Streaming
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public extension AudioProcessor {
  /// We have a new buffer, process and store it.
  /// NOTE: Assumes audio is 16khz mono

Sources/WhisperKit/Core/AudioStreamTranscriber.swift

@@ -3,6 +3,7 @@
 
 import Foundation
 
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public extension AudioStreamTranscriber {
  struct State {
  public var isRecording: Bool = false
@@ -17,9 +18,11 @@ public extension AudioStreamTranscriber {
  }
 }
 
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public typealias AudioStreamTranscriberCallback = (AudioStreamTranscriber.State, AudioStreamTranscriber.State) -> Void
 
 /// Responsible for streaming audio from the microphone, processing it, and transcribing it in real-time.
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public actor AudioStreamTranscriber {
  private var state: AudioStreamTranscriber.State = .init() {
  didSet {

Sources/WhisperKit/Core/FeatureExtractor.swift

@@ -12,7 +12,7 @@ public protocol FeatureExtracting {
  func logMelSpectrogram(fromAudio inputAudio: MLMultiArray) async throws -> MLMultiArray?
 }
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class FeatureExtractor: FeatureExtracting, WhisperMLModel {
  public var model: MLModel?
 

Sources/WhisperKit/Core/LogitsFilter.swift

@@ -9,7 +9,7 @@ public protocol LogitsFiltering {
  func filterLogits(_ logits: MLMultiArray, withTokens tokens: [Int]) -> MLMultiArray
 }
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class SuppressTokensFilter: LogitsFiltering {
  let suppressTokens: [Int]
  private let suppressTokenIndexes: [[NSNumber]]
@@ -25,7 +25,7 @@ public class SuppressTokensFilter: LogitsFiltering {
  }
 }
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class SuppressBlankFilter: LogitsFiltering {
  let suppressBlankTokens: [Int]
  let sampleBegin: Int
@@ -46,7 +46,7 @@ public class SuppressBlankFilter: LogitsFiltering {
  }
 }
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class TimestampRulesFilter: LogitsFiltering {
  let tokenizer: Tokenizer
  let sampleBegin: Int

Sources/WhisperKit/Core/Models.swift

@@ -7,7 +7,7 @@ import NaturalLanguage
 import Tokenizers
 
 #if os(watchOS) || arch(arm64)
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public typealias FloatType = Float16
 #else
 public typealias FloatType = Float
@@ -200,7 +200,7 @@ public struct DecodingCache {
 /// - logProbThreshold: If the average log probability over sampled tokens is below this value, treat as failed.
 /// - noSpeechThreshold: If the no speech probability is higher than this value AND the average log
 /// probability over sampled tokens is below `logProbThreshold`, consider the segment as silent.
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public struct DecodingOptions {
  public var verbose: Bool
  public var task: DecodingTask
@@ -489,7 +489,7 @@ public class MelSpectrogramInput: MLFeatureProvider {
 }
 
 /// Model Prediction Output Type
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class MelSpectrogramOutput: MLFeatureProvider {
  /// Source provided by CoreML
  private let provider: MLFeatureProvider
@@ -526,7 +526,7 @@ public class MelSpectrogramOutput: MLFeatureProvider {
 // MARK: AudioEncoder
 
 /// Model Prediction Input Type
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class AudioEncoderInput: MLFeatureProvider {
  /// melspectrogram_features as 1 × {80,128} × 1 × 3000 4-dimensional array of floats
  public var melspectrogram_features: MLMultiArray
@@ -552,7 +552,7 @@ public class AudioEncoderInput: MLFeatureProvider {
 }
 
 /// Model Prediction Output Type
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class AudioEncoderOutput: MLFeatureProvider {
  /// Source provided by CoreML
  private let provider: MLFeatureProvider
@@ -589,7 +589,7 @@ public class AudioEncoderOutput: MLFeatureProvider {
 // MARK: TextDecoder
 
 /// Model Prediction Input Type
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class TextDecoderInput: MLFeatureProvider {
  /// input_ids as 1 element vector of 32-bit integers
  public var input_ids: MLMultiArray
@@ -657,7 +657,7 @@ public class TextDecoderInput: MLFeatureProvider {
 }
 
 /// Model Prediction Output Type
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class TextDecoderOutput: MLFeatureProvider {
  /// Source provided by CoreML
  private let provider: MLFeatureProvider
@@ -764,7 +764,7 @@ public class TextDecoderCachePrefillInput: MLFeatureProvider {
 }
 
 /// Model Prediction Output Type
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class TextDecoderCachePrefillOutput: MLFeatureProvider {
  /// Source provided by CoreML
  private let provider: MLFeatureProvider

Sources/WhisperKit/Core/SegmentSeeker.swift

@@ -6,7 +6,7 @@ import CoreML
 import Foundation
 import Tokenizers
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public protocol SegmentSeeking {
  func findSeekPointAndSegments(
  decodingResult: DecodingResult,
@@ -34,7 +34,7 @@ public protocol SegmentSeeking {
  ) throws -> [TranscriptionSegment]?
 }
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class SegmentSeeker: SegmentSeeking {
  public init() {}
 

Sources/WhisperKit/Core/TextDecoder.swift

@@ -5,7 +5,7 @@ import Accelerate
 import CoreML
 import Tokenizers
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public protocol TextDecoding {
  var tokenizer: Tokenizer? { get set }
  var prefillData: WhisperMLModel? { get set }
@@ -43,7 +43,7 @@ public protocol TextDecoding {
  )
 }
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public extension TextDecoding {
  func prepareDecoderInputs(withPrompt initialPrompt: [Int]) -> DecodingInputs? {
  let tokenShape = [NSNumber(value: 1), NSNumber(value: initialPrompt.count)]
@@ -234,7 +234,7 @@ public class TextDecoderContextPrefill: WhisperMLModel {
  public var model: MLModel?
 }
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class TextDecoder: TextDecoding, WhisperMLModel {
  public var model: MLModel?
  public var tokenizer: Tokenizer?

Sources/WhisperKit/Core/TokenSampler.swift

@@ -16,7 +16,7 @@ public struct SamplingResult {
  public var completed: Bool
 }
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class GreedyTokenSampler: TokenSampling {
  public var temperature: FloatType
  public var eotToken: Int

Sources/WhisperKit/Core/Utils.swift

@@ -39,6 +39,7 @@ extension MLMultiArray {
  }
 }
 
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 func initMLMultiArray(shape: [NSNumber], dataType: MLMultiArrayDataType, initialValue: Any) -> MLMultiArray {
  let multiArray = try! MLMultiArray(shape: shape, dataType: dataType)
 

Sources/WhisperKit/Core/WhisperKit.swift

@@ -9,12 +9,13 @@ import Hub
 import TensorUtils
 import Tokenizers
 
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public protocol Transcriber {
  func transcribe(audioPath: String, decodeOptions: DecodingOptions?, callback: TranscriptionCallback) async throws -> TranscriptionResult?
  func transcribe(audioArray: [Float], decodeOptions: DecodingOptions?, callback: TranscriptionCallback) async throws -> TranscriptionResult?
 }
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 public class WhisperKit: Transcriber {
  /// Models
  public var modelVariant: ModelVariant = .tiny

Sources/WhisperKitCLI/transcribe.swift

@@ -7,7 +7,7 @@ import Foundation
 
 import WhisperKit
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 @main
 struct WhisperKitCLI: AsyncParsableCommand {
  @Option(help: "Path to audio file")

Tests/WhisperKitTests/FunctionalTests.swift

@@ -5,7 +5,7 @@ import CoreML
 @testable import WhisperKit
 import XCTest
 
-@available(macOS 14, iOS 17, watchOS 10, visionOS 1, *)
+@available(macOS 13, iOS 16, watchOS 10, visionOS 1, *)
 final class FunctionalTests: XCTestCase {
  func testInitLarge() async {
  let modelPath = largev3ModelPath()