Qwen3 TTS 12Hz 0.6B CustomVoice — MLX bf16

MLX bf16 (fp16) conversion of Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice for Apple Silicon inference. Weights are unquantized — the highest-fidelity CustomVoice variant.

Usage

Used by speech-swift Qwen3TTS module:

let model = try await Qwen3TTSModel.fromPretrained(
    modelId: TTSModelVariant.customVoice.rawValue
)
let audio = try model.synthesize("Hello!", speaker: "ryan")

audio speak "Hello!" --model custom-voice --speaker ryan -o output.wav

Model Details

• Architecture: Qwen3-TTS (Talker transformer + Code Predictor + Mimi speech tokenizer decoder)
• Parameters: 0.6B
• Quantization: None — bf16/fp16 weights (highest fidelity)
• Size: ~1.7 GB
• Sample rate: 24 kHz
• Codec rate: 12.5 Hz
• Voices: 9 preset speakers (serena, vivian, uncle_fu, ryan, aiden, ono_anna, sohee, eric, dylan) + instruction-based style control

---

• Guide: soniqo.audio/guides/speak
• Docs: soniqo.audio
• GitHub: soniqo/speech-swift