Voice Design

Create custom voices using natural language descriptions

The VoiceDesign model (Qwen3-TTS-12Hz-1.7B-VoiceDesign) allows you to generate speech with custom voice characteristics by providing natural language descriptions. Describe the voice you want, and the model will synthesize speech matching your description.

Overview

Voice Design enables you to:

• Create voices from textual descriptions (gender, age, accent, emotion, etc.)
• Control speaking style, tone, and prosody through natural language
• Generate unique voices without reference audio
• Combine voice design with voice cloning for reusable custom speakers

Basic Usage

Generate speech with a voice description:

import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel

model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign",
    device_map="cuda:0",
    dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
)

# Generate with voice description
wavs, sr = model.generate_voice_design(
    text="哥哥，你回来啦，人家等了你好久好久了，要抱抱！",
    language="Chinese",
    instruct="体现撒娇稚嫩的萝莉女声，音调偏高且起伏明显，营造出黏人、做作又刻意卖萌的听觉效果。",
)

sf.write("output.wav", wavs[0], sr)

Voice Description Examples

Basic Characteristics

Describe fundamental voice attributes:

# Young male voice
wavs, sr = model.generate_voice_design(
    text="Hello, how can I help you today?",
    language="English",
    instruct="Male, 25 years old, friendly and professional tone",
)

# Elderly female voice
wavs, sr = model.generate_voice_design(
    text="Come here, dear, let me tell you a story.",
    language="English",
    instruct="Female, 70 years old, warm and grandmotherly voice with slight raspiness",
)

# Teen voice with nervousness
wavs, sr = model.generate_voice_design(
    text="H-hey! You dropped your notebook. I think it's yours?",
    language="English",
    instruct="Male, 17 years old, tenor range, gaining confidence - deeper breath support now, though vowels still tighten when nervous",
)

Emotional Expression

Control emotional characteristics through descriptions:

# Excited and energetic
wavs, sr = model.generate_voice_design(
    text="We did it! The project is finally complete!",
    language="English",
    instruct="Young female voice, extremely excited and energetic, speaking quickly with rising intonation",
)

# Worried and anxious
wavs, sr = model.generate_voice_design(
    text="I can't find my keys anywhere. What am I going to do?",
    language="English",
    instruct="Middle-aged male voice, worried and anxious tone, slightly breathless and speaking faster than normal",
)

# Incredulous with panic
wavs, sr = model.generate_voice_design(
    text="It's in the top drawer... wait, it's empty? No way, that's impossible! I'm sure I put it there!",
    language="English",
    instruct="Speak in an incredulous tone, but with a hint of panic beginning to creep into your voice.",
)

Accent and Regional Characteristics

# British accent
wavs, sr = model.generate_voice_design(
    text="Good afternoon. Shall we proceed with the meeting?",
    language="English",
    instruct="Male, 40s, British accent, formal and refined speaking style",
)

# Southern US accent
wavs, sr = model.generate_voice_design(
    text="Well, y'all come back now, you hear?",
    language="English",
    instruct="Female, Southern US accent, warm and hospitable tone",
)

Professional Settings

# News anchor
wavs, sr = model.generate_voice_design(
    text="Good evening. Here are tonight's top stories.",
    language="English",
    instruct="Male, professional news anchor voice, clear articulation, authoritative but approachable, medium paced",
)

# Customer service
wavs, sr = model.generate_voice_design(
    text="Thank you for calling. How may I assist you today?",
    language="English",
    instruct="Female, customer service representative, friendly and helpful tone, clear speech",
)

Batch Generation

Process multiple voice designs efficiently:

import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel

model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign",
    device_map="cuda:0",
    dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
)

wavs, sr = model.generate_voice_design(
    text=[
        "哥哥，你回来啦，人家等了你好久好久了，要抱抱！",
        "It's in the top drawer... wait, it's empty? No way, that's impossible! I'm sure I put it there!"
    ],
    language=["Chinese", "English"],
    instruct=[
        "体现撒娇稚嫩的萝莉女声，音调偏高且起伏明显，营造出黏人、做作又刻意卖萌的听觉效果。",
        "Speak in an incredulous tone, but with a hint of panic beginning to creep into your voice."
    ]
)

for i, wav in enumerate(wavs):
    sf.write(f"output_{i}.wav", wav, sr)

Voice Design Then Clone

A powerful workflow is to:

1. Design a voice using VoiceDesign
2. Create a reusable clone prompt from that voice
3. Use the clone prompt for consistent generation

This is perfect for creating consistent character voices:

import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel

# Step 1: Create reference audio with VoiceDesign
design_model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign",
    device_map="cuda:0",
    dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
)

ref_text = "H-hey! You dropped your... uh... calculus notebook? I mean, I think it's yours? Maybe?"
ref_instruct = "Male, 17 years old, tenor range, gaining confidence - deeper breath support now, though vowels still tighten when nervous"

ref_wavs, sr = design_model.generate_voice_design(
    text=ref_text,
    language="English",
    instruct=ref_instruct
)

sf.write("character_reference.wav", ref_wavs[0], sr)

# Step 2: Build a reusable clone prompt
clone_model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-Base",
    device_map="cuda:0",
    dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
)

voice_clone_prompt = clone_model.create_voice_clone_prompt(
    ref_audio=(ref_wavs[0], sr),
    ref_text=ref_text,
)

# Step 3: Use the clone prompt for new content
sentences = [
    "No problem! I actually... kinda finished those already? If you want to compare answers or something...",
    "What? No! I mean yes but not like... I just think you're... your titration technique is really precise!",
]

for i, sentence in enumerate(sentences):
    wavs, sr = clone_model.generate_voice_clone(
        text=sentence,
        language="English",
        voice_clone_prompt=voice_clone_prompt,
    )
    sf.write(f"character_line_{i}.wav", wavs[0], sr)

Generation Parameters

Fine-tune the generation process:

wavs, sr = model.generate_voice_design(
    text="Your text here",
    language="English",
    instruct="Your voice description",
    # Optional parameters
    max_new_tokens=2048,         # Maximum length
    temperature=0.9,             # Randomness (higher = more variation)
    top_k=50,                    # Top-k sampling
    top_p=1.0,                   # Nucleus sampling
    repetition_penalty=1.05,     # Reduce repetition
    non_streaming_mode=True,     # Use non-streaming generation
)

Complete Example

Here’s the official example from the repository:

examples/test_model_12hz_voice_design.py

import time
import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel

def main():
    device = "cuda:0"
    MODEL_PATH = "Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign/"

    tts = Qwen3TTSModel.from_pretrained(
        MODEL_PATH,
        device_map=device,
        dtype=torch.bfloat16,
        attn_implementation="flash_attention_2",
    )

    # Single inference
    torch.cuda.synchronize()
    t0 = time.time()

    wavs, sr = tts.generate_voice_design(
        text="哥哥，你回来啦，人家等了你好久好久了，要抱抱！",
        language="Chinese",
        instruct="体现撒娇稚嫩的萝莉女声，音调偏高且起伏明显，营造出黏人、做作又刻意卖萌的听觉效果。",
    )

    torch.cuda.synchronize()
    t1 = time.time()
    print(f"[VoiceDesign Single] time: {t1 - t0:.3f}s")

    sf.write("qwen3_tts_test_voice_design_single.wav", wavs[0], sr)

    # Batch inference
    texts = [
        "哥哥，你回来啦，人家等了你好久好久了，要抱抱！",
        "It's in the top drawer... wait, it's empty? No way, that's impossible! I'm sure I put it there!"
    ]
    languages = ["Chinese", "English"]
    instructs = [
        "体现撒娇稚嫩的萝莉女声，音调偏高且起伏明显，营造出黏人、做作又刻意卖萌的听觉效果。",
        "Speak in an incredulous tone, but with a hint of panic beginning to creep into your voice."
    ]

    torch.cuda.synchronize()
    t0 = time.time()

    wavs, sr = tts.generate_voice_design(
        text=texts,
        language=languages,
        instruct=instructs,
        max_new_tokens=2048,
    )

    torch.cuda.synchronize()
    t1 = time.time()
    print(f"[VoiceDesign Batch] time: {t1 - t0:.3f}s")

    for i, w in enumerate(wavs):
        sf.write(f"qwen3_tts_test_voice_design_batch_{i}.wav", w, sr)

if __name__ == "__main__":
    main()

Tips for Better Results

Be specific and detailed

More detailed descriptions produce better results. Instead of “young voice”, try “female, 22 years old, bright and energetic tone with clear articulation”.

Combine multiple attributes

You can combine age, gender, emotion, accent, speaking style, and other characteristics in one description.

Use emotional context

Describe not just how the voice sounds, but the emotional state: “anxious and speaking quickly” or “calm and reassuring”.

Specify prosody

Include details about rhythm, pace, intonation: “speaking slowly with rising intonation at the end of sentences”.

Next Steps

• Learn about Voice Cloning to replicate existing voices
• See Custom Voice for using predefined premium speakers
• Explore Batch Processing for efficient generation