swift-qwen3-tts

dependencies: [
    .package(url: "https://github.com/AtomGradient/swift-qwen3-tts.git", branch: "main"),
],
targets: [
    .target(
        name: "YourApp",
        dependencies: [
            .product(name: "Qwen3TTS", package: "swift-qwen3-tts"),
        ]
    ),
]

import Qwen3TTS

let model = try await Qwen3TTSModel.fromPretrained(
    "/path/to/Qwen3-TTS-12Hz-1.7B-VoiceDesign-bf16"
)

let audio = try await model.generate(
    text: "Hello, welcome to Qwen3 TTS!",
    instruct: "A warm female voice with a friendly tone",
    language: "english"
)

// audio is an MLXArray of Float samples at 24kHz
try saveAudioArray(audio, sampleRate: 24000, to: URL(fileURLWithPath: "output.wav"))

let model = try await Qwen3TTSModel.fromPretrained(
    "/path/to/Qwen3-TTS-12Hz-0.6B-CustomVoice-bf16"
)

let audio = try await model.generate(
    text: "This is exciting news!",
    speaker: "Aiden",
    instruct: "Happy and energetic",
    language: "english"
)

let model = try await Qwen3TTSModel.fromPretrained(
    "/path/to/Qwen3-TTS-12Hz-1.7B-Base-bf16"
)

// Load 3+ seconds of reference audio (24kHz)
let (sampleRate, refAudio) = try loadAudioArray(from: URL(fileURLWithPath: "reference.wav"))

let audio = try model.generateVoiceClone(
    text: "New content in the cloned voice.",
    referenceAudio: refAudio,
    referenceText: "Transcript of the reference audio.",
    language: "english"
)

for try await event in model.generateStream(
    text: "Streaming generation example.",
    instruct: "A calm narrator voice"
) {
    switch event {
    case .token(let id):
        break  // first-codebook codec token generated
    case .info(let info):
        print(info.summary)  // generation stats
    case .audio(let audio):
        // final audio MLXArray
        break
    }
}

public class Qwen3TTSModel: Module {
    // Load model from a local directory containing config.json and safetensors files
    public static func fromPretrained(_ modelPath: String) async throws -> Qwen3TTSModel

    // Unified generation (auto-routes by model type and parameters)
    public func generate(
        text: String,
        speaker: String? = nil,
        instruct: String? = nil,
        language: String = "auto",
        temperature: Float = 0.9,
        topK: Int = 50,
        topP: Float = 1.0,
        repetitionPenalty: Float = 1.05,
        maxTokens: Int = 2048
    ) async throws -> MLXArray

    // Streaming generation
    public func generateStream(/* same params */) -> AsyncThrowingStream<Qwen3TTSGeneration, Error>

    // Voice cloning (Base model only)
    public func generateVoiceClone(
        text: String,
        referenceAudio: MLXArray,
        referenceText: String,
        language: String = "auto",
        temperature: Float = 0.9,
        topK: Int = 50,
        topP: Float = 1.0,
        repetitionPenalty: Float = 1.5,
        maxTokens: Int = 2048
    ) throws -> MLXArray

    // Properties
    public var sampleRate: Int             // 24000
    public var ttsModelType: String        // "voice_design" / "custom_voice" / "base"
    public var supportedSpeakers: [String] // Available speaker names
    public var supportsVoiceCloning: Bool  // Whether model supports cloning
}

// Load audio file to MLXArray
public func loadAudioArray(from url: URL) throws -> (Int, MLXArray)  // (sampleRate, samples)

// Save MLXArray to WAV file
public func saveAudioArray(_ audio: MLXArray, sampleRate: Double, to url: URL) throws