99 lines
3.9 KiB
Swift
99 lines
3.9 KiB
Swift
import Foundation
|
|
import AVFoundation
|
|
import os
|
|
|
|
class LocalTranscriptionService: TranscriptionService {
|
|
|
|
private var whisperContext: WhisperContext?
|
|
private let logger = Logger(subsystem: "com.prakashjoshipax.voiceink", category: "LocalTranscriptionService")
|
|
private let modelsDirectory: URL
|
|
private weak var whisperState: WhisperState?
|
|
|
|
init(modelsDirectory: URL, whisperState: WhisperState? = nil) {
|
|
self.modelsDirectory = modelsDirectory
|
|
self.whisperState = whisperState
|
|
}
|
|
|
|
func transcribe(audioURL: URL, model: any TranscriptionModel) async throws -> String {
|
|
guard let localModel = model as? LocalModel else {
|
|
throw WhisperStateError.modelLoadFailed
|
|
}
|
|
|
|
logger.notice("Initiating local transcription for model: \(localModel.displayName)")
|
|
|
|
// Check if the required model is already loaded in WhisperState
|
|
if let whisperState = whisperState,
|
|
await whisperState.isModelLoaded,
|
|
let loadedContext = await whisperState.whisperContext,
|
|
let currentModel = await whisperState.currentTranscriptionModel,
|
|
currentModel.provider == .local,
|
|
currentModel.name == localModel.name {
|
|
|
|
logger.notice("✅ Using already loaded model: \(localModel.name)")
|
|
whisperContext = loadedContext
|
|
} else {
|
|
// Model not loaded or wrong model loaded, proceed with loading
|
|
let modelURL = modelsDirectory.appendingPathComponent(localModel.filename)
|
|
|
|
guard FileManager.default.fileExists(atPath: modelURL.path) else {
|
|
logger.error("Model file not found at path: \(modelURL.path)")
|
|
throw WhisperStateError.modelLoadFailed
|
|
}
|
|
|
|
logger.notice("Loading model: \(localModel.name)")
|
|
do {
|
|
whisperContext = try await WhisperContext.createContext(path: modelURL.path)
|
|
} catch {
|
|
logger.error("Failed to load model: \(localModel.name) - \(error.localizedDescription)")
|
|
throw WhisperStateError.modelLoadFailed
|
|
}
|
|
}
|
|
|
|
guard let whisperContext = whisperContext else {
|
|
logger.error("Cannot transcribe: Model could not be loaded")
|
|
throw WhisperStateError.modelLoadFailed
|
|
}
|
|
|
|
// Read audio data
|
|
let data = try readAudioSamples(audioURL)
|
|
|
|
// Set prompt
|
|
let currentPrompt = UserDefaults.standard.string(forKey: "TranscriptionPrompt") ?? ""
|
|
await whisperContext.setPrompt(currentPrompt)
|
|
|
|
// Transcribe
|
|
let success = await whisperContext.fullTranscribe(samples: data)
|
|
|
|
guard success else {
|
|
logger.error("Core transcription engine failed (whisper_full).")
|
|
throw WhisperStateError.whisperCoreFailed
|
|
}
|
|
|
|
var text = await whisperContext.getTranscription()
|
|
|
|
if UserDefaults.standard.object(forKey: "IsTextFormattingEnabled") as? Bool ?? true {
|
|
text = WhisperTextFormatter.format(text)
|
|
}
|
|
|
|
logger.notice("✅ Local transcription completed successfully.")
|
|
|
|
// Only release resources if we created a new context (not using the shared one)
|
|
if await whisperState?.whisperContext !== whisperContext {
|
|
await whisperContext.releaseResources()
|
|
self.whisperContext = nil
|
|
}
|
|
|
|
return text
|
|
}
|
|
|
|
private func readAudioSamples(_ url: URL) throws -> [Float] {
|
|
let data = try Data(contentsOf: url)
|
|
let floats = stride(from: 44, to: data.count, by: 2).map {
|
|
return data[$0..<$0 + 2].withUnsafeBytes {
|
|
let short = Int16(littleEndian: $0.load(as: Int16.self))
|
|
return max(-1.0, min(Float(short) / 32767.0, 1.0))
|
|
}
|
|
}
|
|
return floats
|
|
}
|
|
} |