增加 OpenAI 标准接口(audio/chat)的支持，vllm和OmniX-MLX测试正常

Shared/Sources/Shared/Audio/WAVEncoder.swift

@@ -0,0 +1,56 @@
+import Foundation
+
+/// 将原始 PCM 数据封装为 WAV 文件格式（内存中）
+/// WAV = 44 字节 RIFF 头 + 原始 PCM 数据，零转码开销
+public enum WAVEncoder {
+    /// 将 PCM 数据编码为 WAV
+    /// - Parameters:
+    ///   - pcmData: 原始 PCM 音频数据（例如 S16LE）
+    ///   - sampleRate: 采样率（例如 16000）
+    ///   - channels: 声道数（例如 1 = 单声道）
+    ///   - bitsPerSample: 每个采样的位数（例如 16）
+    /// - Returns: 包含 RIFF/WAV 头的完整 WAV 文件数据
+    public static func encode(pcmData: Data, sampleRate: Int, channels: Int, bitsPerSample: Int) -> Data {
+        let byteRate = sampleRate * channels * (bitsPerSample / 8)
+        let blockAlign = channels * (bitsPerSample / 8)
+        let dataSize = UInt32(pcmData.count)
+        let fileSize = UInt32(36) + dataSize
+
+        var header = Data(capacity: 44)
+
+        // RIFF chunk descriptor
+        header.append(contentsOf: [0x52, 0x49, 0x46, 0x46]) // "RIFF"
+        header.appendLittleEndian(fileSize)
+        header.append(contentsOf: [0x57, 0x41, 0x56, 0x45]) // "WAVE"
+
+        // fmt sub-chunk
+        header.append(contentsOf: [0x66, 0x6D, 0x74, 0x20]) // "fmt "
+        header.appendLittleEndian(UInt32(16))                 // sub-chunk size
+        header.appendLittleEndian(UInt16(1))                  // PCM format
+        header.appendLittleEndian(UInt16(channels))
+        header.appendLittleEndian(UInt32(sampleRate))
+        header.appendLittleEndian(UInt32(byteRate))
+        header.appendLittleEndian(UInt16(blockAlign))
+        header.appendLittleEndian(UInt16(bitsPerSample))
+
+        // data sub-chunk
+        header.append(contentsOf: [0x64, 0x61, 0x74, 0x61]) // "data"
+        header.appendLittleEndian(dataSize)
+
+        return header + pcmData
+    }
+}
+
+// MARK: - Data helpers for little-endian encoding
+
+private extension Data {
+    mutating func appendLittleEndian(_ value: UInt16) {
+        var v = value.littleEndian
+        Swift.withUnsafeBytes(of: &v) { append(contentsOf: $0) }
+    }
+
+    mutating func appendLittleEndian(_ value: UInt32) {
+        var v = value.littleEndian
+        Swift.withUnsafeBytes(of: &v) { append(contentsOf: $0) }
+    }
+}

Shared/Sources/Shared/Core/Storage/APIKeyStorage.swift

@@ -1,24 +1,81 @@
 import Foundation
 
-/// 存储火山引擎语音识别 API Key 和配置（使用 UserDefaults）
+/// vLLM API 调用模式
+public enum VLLMApiMode: String, CaseIterable, Identifiable, Sendable {
+    /// 标准 OpenAI Audio Transcriptions 接口 (POST /audio/transcriptions, multipart/form-data)
+    case audioTranscriptions = "audio_transcriptions"
+    /// Chat Completions 接口 (POST /chat/completions, base64 audio data URL)
+    case chatCompletions = "chat_completions"
+
+    public var id: String { rawValue }
+
+    public var displayName: String {
+        switch self {
+        case .audioTranscriptions: return "Audio Transcriptions (标准)"
+        case .chatCompletions: return "Chat Completions"
+        }
+    }
+}
+
+/// 语音识别服务提供商
+public enum SpeechProvider: String, CaseIterable, Identifiable, Sendable {
+    case volcEngine = "volcengine"
+    case vllm = "vllm"
+
+    public var id: String { rawValue }
+
+    public var displayName: String {
+        switch self {
+        case .volcEngine: return "火山引擎豆包"
+        case .vllm: return "vLLM (OpenAI 兼容)"
+        }
+    }
+}
+
+/// 存储语音识别 API 配置（使用 UserDefaults）
 @MainActor
 public final class APIKeyStorage: Sendable {
     public static let shared = APIKeyStorage()
-    
+
     private let defaults = UserDefaults.standard
-    
+
     private enum Keys {
+        // 通用
+        static let provider = "speech.provider"
+        // 火山引擎
         static let apiKey = "volcengine.api.key"
         static let resourceId = "volcengine.resource.id"
+        // vLLM
+        static let vllmBaseURL = "vllm.base.url"
+        static let vllmModelName = "vllm.model.name"
+        static let vllmAPIKey = "vllm.api.key"
+        static let vllmApiMode = "vllm.api.mode"
     }
-    
+
     public init() {}
-
+
+    // MARK: - Provider 选择
+
+    public var selectedProvider: SpeechProvider {
+        get {
+            guard let rawValue = defaults.string(forKey: Keys.provider),
+                  let value = SpeechProvider(rawValue: rawValue) else {
+                return .volcEngine
+            }
+            return value
+        }
+        set {
+            defaults.set(newValue.rawValue, forKey: Keys.provider)
+        }
+    }
+
+    // MARK: - 火山引擎配置
+
     public var apiKey: String? {
         get { defaults.string(forKey: Keys.apiKey) }
         set { defaults.set(newValue, forKey: Keys.apiKey) }
     }
-    
+
     /// 资源 ID（模型版本）
     public var resourceId: VolcEngineResourceId {
         get {
@@ -32,22 +89,81 @@ public final class APIKeyStorage: Sendable {
             defaults.set(newValue.rawValue, forKey: Keys.resourceId)
         }
     }
-
+
+    // MARK: - vLLM 配置
+
+    public var vllmBaseURL: String {
+        get { defaults.string(forKey: Keys.vllmBaseURL) ?? "http://localhost:8000/v1" }
+        set { defaults.set(newValue, forKey: Keys.vllmBaseURL) }
+    }
+
+    public var vllmModelName: String {
+        get { defaults.string(forKey: Keys.vllmModelName) ?? "openai/whisper-large-v3" }
+        set { defaults.set(newValue, forKey: Keys.vllmModelName) }
+    }
+
+    public var vllmAPIKey: String? {
+        get { defaults.string(forKey: Keys.vllmAPIKey) }
+        set { defaults.set(newValue, forKey: Keys.vllmAPIKey) }
+    }
+
+    public var vllmApiMode: VLLMApiMode {
+        get {
+            guard let rawValue = defaults.string(forKey: Keys.vllmApiMode),
+                  let value = VLLMApiMode(rawValue: rawValue) else {
+                return .audioTranscriptions
+            }
+            return value
+        }
+        set {
+            defaults.set(newValue.rawValue, forKey: Keys.vllmApiMode)
+        }
+    }
+
+    // MARK: - 状态检查
+
     public var isConfigured: Bool {
-        guard let apiKey else { return false }
-        return !apiKey.isEmpty
+        switch selectedProvider {
+        case .volcEngine:
+            guard let apiKey else { return false }
+            return !apiKey.isEmpty
+        case .vllm:
+            return !vllmBaseURL.isEmpty && !vllmModelName.isEmpty
+        }
     }
-
+
+    // MARK: - 保存
+
     public func save(apiKey: String) {
         self.apiKey = apiKey
     }
-    
+
     public func save(resourceId: VolcEngineResourceId) {
         self.resourceId = resourceId
     }
-
+
+    public func save(vllmBaseURL: String) {
+        self.vllmBaseURL = vllmBaseURL
+    }
+
+    public func save(vllmModelName: String) {
+        self.vllmModelName = vllmModelName
+    }
+
+    public func save(vllmAPIKey: String?) {
+        self.vllmAPIKey = vllmAPIKey
+    }
+
+    public func save(vllmApiMode: VLLMApiMode) {
+        self.vllmApiMode = vllmApiMode
+    }
+
     public func clear() {
         defaults.removeObject(forKey: Keys.apiKey)
         defaults.removeObject(forKey: Keys.resourceId)
+        defaults.removeObject(forKey: Keys.vllmBaseURL)
+        defaults.removeObject(forKey: Keys.vllmModelName)
+        defaults.removeObject(forKey: Keys.vllmAPIKey)
+        defaults.removeObject(forKey: Keys.vllmApiMode)
     }
 }

Shared/Sources/Shared/SpeechRecognition/Backend/MultipartFormData.swift

@@ -0,0 +1,40 @@
+import Foundation
+
+/// 轻量级 multipart/form-data 编码器
+/// 用于构建 HTTP 文件上传请求体，无需第三方依赖
+struct MultipartFormData {
+    private let boundary: String
+    private var body = Data()
+
+    /// Content-Type header 值（含 boundary）
+    var contentType: String {
+        "multipart/form-data; boundary=\(boundary)"
+    }
+
+    init(boundary: String = UUID().uuidString) {
+        self.boundary = boundary
+    }
+
+    /// 添加文本字段
+    mutating func addField(name: String, value: String) {
+        body.append("--\(boundary)\r\n".data(using: .utf8)!)
+        body.append("Content-Disposition: form-data; name=\"\(name)\"\r\n\r\n".data(using: .utf8)!)
+        body.append("\(value)\r\n".data(using: .utf8)!)
+    }
+
+    /// 添加文件字段
+    mutating func addFile(name: String, fileName: String, mimeType: String, data: Data) {
+        body.append("--\(boundary)\r\n".data(using: .utf8)!)
+        body.append("Content-Disposition: form-data; name=\"\(name)\"; filename=\"\(fileName)\"\r\n".data(using: .utf8)!)
+        body.append("Content-Type: \(mimeType)\r\n\r\n".data(using: .utf8)!)
+        body.append(data)
+        body.append("\r\n".data(using: .utf8)!)
+    }
+
+    /// 完成编码，返回最终的请求体数据
+    func finalize() -> Data {
+        var result = body
+        result.append("--\(boundary)--\r\n".data(using: .utf8)!)
+        return result
+    }
+}

Shared/Sources/Shared/SpeechRecognition/Backend/SpeechBackend.swift

@@ -0,0 +1,26 @@
+import Foundation
+
+/// 语音识别后端协议（内部使用）
+/// 每个后端负责将其原生响应转换为统一的 SpeechRecognitionResult。
+/// 后端实例在 @MainActor 的 SpeechRecognitionService 内创建和持有。
+@MainActor
+protocol SpeechBackend {
+    /// 开始识别会话，返回结果流
+    func startSession() async throws -> AsyncStream<SpeechRecognitionResult>
+
+    /// 发送音频数据
+    /// - 流式后端（火山引擎）：立即通过 WebSocket 发送
+    /// - 批量后端（vLLM）：缓存到内存 buffer
+    func sendAudioData(_ data: Data) async throws
+
+    /// 完成音频输入
+    /// - 流式后端：发送结束包
+    /// - 批量后端：组装 WAV，POST 到服务器，将结果 yield 到 stream
+    func finishAudio() async throws
+
+    /// 测试连接是否可用
+    func testConnection() async throws
+
+    /// 断开/清理
+    func disconnect() async
+}