Skip to main content
The Base models (Qwen3-TTS-12Hz-1.7B-Base and Qwen3-TTS-12Hz-0.6B-Base) enable rapid voice cloning from just 3 seconds of reference audio. Clone any voice and generate new speech with the same timbre and characteristics.

​
Overview

Voice cloning allows you to:
  • Clone any voice from a short audio sample (3+ seconds recommended)
  • Generate new content in the cloned voice
  • Create reusable voice prompts for consistent generation
  • Choose between full cloning (ICL mode) or speaker embedding only

​
Basic Voice Cloning

Clone a voice and generate speech in one call: 
import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel

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

ref_audio = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-TTS-Repo/clone.wav"
ref_text = "Okay. Yeah. I resent you. I love you. I respect you. But you know what? You blew it! And thanks to you."

wavs, sr = model.generate_voice_clone(
    text="I am solving the equation: x = [-b ± √(b²-4ac)] / 2a? Nobody can — it's a disaster (◍•͈⌔•͈◍), very sad!",
    language="English",
    ref_audio=ref_audio,
    ref_text=ref_text,
)

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

​
Reference Audio Requirements

​
Audio Input Formats

The ref_audio parameter accepts multiple formats: 
# Local file path
ref_audio = "/path/to/audio.wav"

# URL
ref_audio = "https://example.com/audio.wav"

# Base64 encoded string
ref_audio = "data:audio/wav;base64,UklGRiQAAABXQVZFZm10..."

# NumPy array with sample rate tuple
import numpy as np
waveform = np.array([...])  # Your audio data
ref_audio = (waveform, 24000)  # (audio, sample_rate)

​
Quality Guidelines

Duration

3+ seconds recommended for best results. Longer samples may improve quality.

Clean Audio

Use clear audio without background noise, music, or multiple speakers.

Single Speaker

Reference audio should contain only the target speaker’s voice.

Natural Speech

Normal speaking pace and intonation work best. Avoid shouting or whispering.

​
Reusable Voice Prompts

For better performance when generating multiple times with the same voice, create a reusable prompt: 
import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel

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

ref_audio = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-TTS-Repo/clone.wav"
ref_text = "Okay. Yeah. I resent you. I love you. I respect you. But you know what? You blew it! And thanks to you."

# Step 1: Create the voice clone prompt once
prompt_items = model.create_voice_clone_prompt(
    ref_audio=ref_audio,
    ref_text=ref_text,
    x_vector_only_mode=False,
)

# Step 2: Reuse the prompt for multiple generations
sentences = [
    "Sentence A: This is the first test.",
    "Sentence B: Here's another example.",
    "Sentence C: And one more for good measure.",
]

for i, text in enumerate(sentences):
    wavs, sr = model.generate_voice_clone(
        text=text,
        language="English",
        voice_clone_prompt=prompt_items,  # Reuse the same prompt
    )
    sf.write(f"output_{i}.wav", wavs[0], sr)
Reusing prompts avoids recomputing audio features, significantly improving performance when generating multiple outputs.

​
ICL Mode vs X-Vector Only Mode

The Base model supports two cloning modes: In-Context Learning (ICL) uses both the reference audio codes and speaker embedding: 
wavs, sr = model.generate_voice_clone(
    text="Your new text",
    language="English",
    ref_audio=ref_audio,
    ref_text=ref_text,           # Text is REQUIRED
    x_vector_only_mode=False,    # ICL mode (default)
)
Advantages:
  • Higher quality cloning
  • Better preservation of voice characteristics
  • More natural prosody
Requirements:
  • Must provide ref_text (transcript of reference audio)

​
X-Vector Only Mode

Uses only the speaker embedding (x-vector) without reference codes: 
wavs, sr = model.generate_voice_clone(
    text="Your new text",
    language="English",
    ref_audio=ref_audio,
    ref_text=None,                # Text NOT required
    x_vector_only_mode=True,      # Only use speaker embedding
)
Advantages:
  • No need for reference text
  • Faster processing
Disadvantages:
  • Lower cloning quality
  • Less accurate voice characteristics
ICL mode (x_vector_only_mode=False) is strongly recommended for best quality. Only use x-vector mode when you cannot provide reference text.

​
Batch Voice Cloning

​
Same Voice, Multiple Texts

Generate multiple outputs using the same cloned voice: 
# Create prompt once
prompt_items = model.create_voice_clone_prompt(
    ref_audio=ref_audio,
    ref_text=ref_text,
)

# Generate batch
wavs, sr = model.generate_voice_clone(
    text=["First sentence.", "Second sentence."],
    language=["English", "English"],
    voice_clone_prompt=prompt_items,
)

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

​
Different Voices, Multiple Texts

Clone multiple voices and generate in batch: 
ref_audios = [
    "https://example.com/voice1.wav",
    "https://example.com/voice2.wav",
]
ref_texts = [
    "Reference text for voice one.",
    "Reference text for voice two.",
]

# Create prompts for both voices
prompt_items = model.create_voice_clone_prompt(
    ref_audio=ref_audios,
    ref_text=ref_texts,
)

# Generate with different voices
wavs, sr = model.generate_voice_clone(
    text=["Text in voice one.", "Text in voice two."],
    language=["English", "English"],
    voice_clone_prompt=prompt_items,
)

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

​
Complete Example

Here’s the official example demonstrating all cloning modes:
examples/test_model_12hz_base.py
import os
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-Base/"
    OUT_DIR = "qwen3_tts_test_voice_clone_output_wav"
    os.makedirs(OUT_DIR, exist_ok=True)

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

    # Reference audio and text
    ref_audio = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-TTS-Repo/clone_2.wav"
    ref_text = "Okay. Yeah. I resent you. I love you. I respect you. But you know what? You blew it! And thanks to you."

    # Target text to synthesize
    syn_text = "Good one. Okay, fine, I'm just gonna leave this sock monkey here. Goodbye."
    syn_lang = "Auto"

    gen_kwargs = dict(
        max_new_tokens=2048,
        do_sample=True,
        top_k=50,
        top_p=1.0,
        temperature=0.9,
        repetition_penalty=1.05,
    )

    # Test both modes
    for xvec_only in [False, True]:
        mode = "xvec_only" if xvec_only else "icl"
        print(f"\n=== Testing {mode} mode ===")

        # Method 1: Direct generation
        t0 = time.time()
        wavs, sr = tts.generate_voice_clone(
            text=syn_text,
            language=syn_lang,
            ref_audio=ref_audio,
            ref_text=ref_text,
            x_vector_only_mode=xvec_only,
            **gen_kwargs,
        )
        t1 = time.time()
        print(f"Direct: {t1 - t0:.3f}s")
        sf.write(f"{OUT_DIR}/direct_{mode}.wav", wavs[0], sr)

        # Method 2: With reusable prompt
        prompt_items = tts.create_voice_clone_prompt(
            ref_audio=ref_audio,
            ref_text=ref_text,
            x_vector_only_mode=xvec_only,
        )

        t0 = time.time()
        wavs, sr = tts.generate_voice_clone(
            text=syn_text,
            language=syn_lang,
            voice_clone_prompt=prompt_items,
            **gen_kwargs,
        )
        t1 = time.time()
        print(f"Reusable prompt: {t1 - t0:.3f}s")
        sf.write(f"{OUT_DIR}/prompt_{mode}.wav", wavs[0], sr)

if __name__ == "__main__":
    main()

​
Generation Parameters

Customize the generation: 
wavs, sr = model.generate_voice_clone(
    text="Your text",
    language="English",
    ref_audio=ref_audio,
    ref_text=ref_text,
    # Generation parameters
    max_new_tokens=2048,
    temperature=0.9,
    top_k=50,
    top_p=1.0,
    repetition_penalty=1.05,
    non_streaming_mode=False,  # Simulate streaming
)

​
Combining Voice Design and Cloning

Create a custom voice with VoiceDesign, then clone it for consistent character voices: 
# Step 1: Design voice
design_model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign",
    device_map="cuda:0",
    dtype=torch.bfloat16,
)

ref_text = "Hello, I'm your new virtual assistant."
ref_wavs, sr = design_model.generate_voice_design(
    text=ref_text,
    language="English",
    instruct="Female, 30s, professional and friendly tone, clear articulation"
)

# Step 2: Clone the designed voice
clone_model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-Base",
    device_map="cuda:0",
    dtype=torch.bfloat16,
)

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

# Step 3: Use for all future generations
wavs, sr = clone_model.generate_voice_clone(
    text="How can I help you today?",
    language="English",
    voice_clone_prompt=voice_prompt,
)
See the Voice Design guide for more details on this workflow.

​
Troubleshooting

  • Use ICL mode instead of x-vector only mode
  • Ensure reference audio is clean and clear
  • Use longer reference audio (5-10 seconds)
  • Verify reference text exactly matches the audio
Make sure ref_text is provided when x_vector_only_mode=False. The text should accurately transcribe the reference audio.
For batch processing, reduce batch size or use the 0.6B model instead of 1.7B.

​
Next Steps

Build docs developers (and LLMs) love

Get started for freeTalk to us