Add VoxCPM2 TTS model (2B params, 48kHz, voice cloning)

[sb1992]

Swift port of VoxCPM2, based on the Python implementation in Blaizzy/mlx-audio#641.

Autoregressive MiniCPM backbone + CFM diffusion + AudioVAE decoder.
Supports zero-shot generation and reference-audio voice cloning across 30+ languages.

Supported model

• mlx-community/VoxCPM2-4bit (4-bit quantized, 2.3 GB)

Example

import MLXAudioTTS
import MLXAudioCore

let model = try await VoxCPM2Model.fromPretrained("mlx-community/VoxCPM2-4bit")

// Voice cloning
let (_, refAudio) = try loadAudioArray(from: referenceAudioURL)
let audio = try await model.generate(
    text: "Hello, this is VoxCPM2.",
    voice: nil, refAudio: refAudio, refText: nil, language: nil,
    generationParameters: GenerateParameters(maxTokens: 100, temperature: 1.0)
)

Smoke test results (M5 Max, macOS 26.3.2)

Test	Audio	Duration	Peak Memory	Time
voxcpm2Generate()	92,160 samples @ 48kHz	1.92s	3.72 GB	6.96s
voxcpm2GenerateStream()	69,120 samples @ 48kHz	1.44s	—	2.03s

Both produce valid mono Float32 WAV at 48kHz.

Known limitations

• 4-bit stop predictor can be unreliable — callers should set maxTokens conservatively
• Short Chinese prompts may have text-following issues (consistent with Python, see mlx-audio#641)
• Streaming wraps the full generate() call rather than yielding incremental chunks
• Voice design and continuation modes from the Python PR are not yet ported

Checklist

• Conforms to SpeechGenerationModel
• Factory registration in TTSModel.swift
• Smoke tests for generate + stream
• README.md with usage and limitations
• swift build clean (zero warnings)
• Package.swift README exclude

[sb1992]

Pushed a small compatibility fix for the failed CI build.

The failure was from Xcode 26.3 rejecting Float * MLXArray in VoxCPMDiT.swift. I reproduced the failure locally with Xcode 26.3 (17C529), changed the scalar to MLXArray(scale), and verified xcodebuild build-for-testing now succeeds with the same Xcode version.