Available Models

Overview of Qwen3-TTS model variants and their capabilities

Model Overview

Qwen3-TTS offers a family of models optimized for different use cases, from voice cloning to instruction-based voice design. All models are built on the same architecture but are fine-tuned for specific capabilities.

Model Comparison

Model	Parameters	Features	Language Support	Streaming	Instruction Control
Qwen3-TTS-12Hz-1.7B-VoiceDesign	1.7B	Voice design from text descriptions	10 languages	✅	✅
Qwen3-TTS-12Hz-1.7B-CustomVoice	1.7B	9 premium voices with style control	10 languages	✅	✅
Qwen3-TTS-12Hz-1.7B-Base	1.7B	3-second voice cloning	10 languages	✅	❌
Qwen3-TTS-12Hz-0.6B-CustomVoice	0.6B	9 premium voices	10 languages	✅	❌
Qwen3-TTS-12Hz-0.6B-Base	0.6B	3-second voice cloning	10 languages	✅	❌

Supported Languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian

Model Variants

VoiceDesign

Create custom voices from natural language descriptions

CustomVoice

Use 9 premium pre-trained voices with style control

Base

Clone any voice from just 3 seconds of audio

Tokenizer

Encode/decode audio for compression and transport

1. Qwen3-TTS-12Hz-1.7B-VoiceDesign

Best for: Creating unique voices based on detailed descriptions

Capabilities

• Generate voices from natural language descriptions
• Control timbre, age, gender, emotion, and speaking style
• Instruction-based prosody and tone adjustment
• Adaptive emotional expression based on text semantics
• Streaming generation support

Use Cases

• Character voice creation for games and animation
• Personalized virtual assistants
• Content creation requiring unique voices
• Prototyping voice concepts without recording

Example

from qwen_tts import Qwen3TTSModel

model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign"
)

wavs, sr = model.generate_voice_design(
    text="Hello, how can I help you today?",
    language="English",
    instruct="Male, 30 years old, warm and friendly tone, "
             "professional voice with clear articulation"
)

2. Qwen3-TTS-12Hz-1.7B-CustomVoice

Best for: Production applications requiring consistent, high-quality voices

Capabilities

• 9 premium pre-trained voices
• Style control via natural language instructions
• Covers various genders, ages, languages, and dialects
• Instruction-based emotion and prosody control
• Streaming generation support

Available Voices

Speaker	Description	Native Language
Vivian	Bright, slightly edgy young female	Chinese
Serena	Warm, gentle young female	Chinese
Uncle_Fu	Seasoned male, low mellow timbre	Chinese
Dylan	Youthful Beijing male, clear natural	Chinese (Beijing Dialect)
Eric	Lively Chengdu male, slightly husky	Chinese (Sichuan Dialect)
Ryan	Dynamic male, strong rhythmic drive	English
Aiden	Sunny American male, clear midrange	English
Ono_Anna	Playful Japanese female, light nimble	Japanese
Sohee	Warm Korean female, rich emotion	Korean

Example

from qwen_tts import Qwen3TTSModel

model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice"
)

# List available speakers
print(model.get_supported_speakers())

wavs, sr = model.generate_custom_voice(
    text="I'm so excited to announce our new product!",
    language="English",
    speaker="Ryan",
    instruct="Very enthusiastic and energetic"
)

3. Qwen3-TTS-12Hz-1.7B-Base

Best for: Voice cloning and fine-tuning for custom applications

Capabilities

• Clone any voice from 3 seconds of audio
• Maintains speaker characteristics and style
• Support for reference audio in multiple formats (file, URL, bytes)
• Can be fine-tuned for specific domains or speakers
• Streaming generation support

Use Cases

• Personal voice cloning
• Voice preservation and restoration
• Audiobook narration in original voice
• Custom voice assistants
• Fine-tuning base for specialized applications

Example

from qwen_tts import Qwen3TTSModel

model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-Base"
)

wavs, sr = model.generate_voice_clone(
    text="This is a test of voice cloning.",
    language="English",
    ref_audio="path/to/reference.wav",
    ref_text="The transcript of the reference audio."
)

4. Qwen3-TTS-12Hz-0.6B Models

Best for: Applications requiring faster inference or limited resources

The 0.6B variants offer:

• Faster inference: ~1.7x faster than 1.7B models
• Lower memory: Reduced GPU memory requirements
• Good quality: Maintains high quality with slight trade-off vs 1.7B
• Same features: CustomVoice and Base variants available

The 0.6B models are ideal for edge devices, high-throughput services, or applications where speed is prioritized over maximum quality.

5. Qwen3-TTS-Tokenizer-12Hz

Best for: Audio compression, transport, and training data preparation

Capabilities

• Encode audio to discrete codes (compression)
• Decode codes back to audio (reconstruction)
• 12Hz frame rate (12 frames per second)
• 16 residual quantizers
• 2048 codebook size per quantizer
• 24kHz sample rate

Use Cases

• Efficient audio storage and transmission
• Preparing training data for TTS models
• Audio codec research
• Analyzing speech representations

Example

from qwen_tts import Qwen3TTSTokenizer

tokenizer = Qwen3TTSTokenizer.from_pretrained(
    "Qwen/Qwen3-TTS-Tokenizer-12Hz"
)

# Encode audio to codes
codes = tokenizer.encode("input.wav")

# Decode codes back to audio
wavs, sr = tokenizer.decode(codes)

Model Selection Guide

I want to create unique character voices

Recommended: Qwen3-TTS-12Hz-1.7B-VoiceDesign

This model excels at generating voices from descriptions. You can specify exact characteristics like “teenage girl with cheerful tone” or “elderly professor with authoritative voice.”

I need consistent professional voices

Recommended: Qwen3-TTS-12Hz-1.7B-CustomVoice (or 0.6B for faster inference)

Use the 9 premium pre-trained voices for production applications. They’re highly consistent and support instruction-based style control.

I want to clone a specific person's voice

Recommended: Qwen3-TTS-12Hz-1.7B-Base (or 0.6B for faster inference)

Provide 3 seconds of reference audio to clone any voice. Ideal for personal voice assistants or preserving specific speaker characteristics.

I need fast inference for high-throughput services

Recommended: Qwen3-TTS-12Hz-0.6B-CustomVoice or 0.6B-Base

The 0.6B models offer significantly faster inference while maintaining good quality. Perfect for real-time applications or serving many concurrent requests.

I want to fine-tune for my specific use case

Recommended: Qwen3-TTS-12Hz-1.7B-Base or 0.6B-Base

Base models serve as excellent starting points for fine-tuning. See the Fine-Tuning Guide for details.

I need audio compression and codec functionality

Recommended: Qwen3-TTS-Tokenizer-12Hz

Use the tokenizer for encode/decode operations, audio compression, or preparing training data.

Performance Benchmarks

Content Consistency (WER ↓)

Word Error Rate on Seed-TTS test set:

Model	Chinese	English
Qwen3-TTS-12Hz-1.7B-Base	0.77	1.24
Qwen3-TTS-12Hz-0.6B-Base	0.92	1.32
CosyVoice 3	0.71	1.45
MiniMax-Speech	0.83	1.65

Speaker Similarity (Cosine Similarity ↑)

Speaker similarity on multilingual test set:

Language	1.7B-Base	0.6B-Base
Chinese	0.799	0.811
English	0.775	0.829
Japanese	0.788	0.798
Korean	0.799	0.812

Higher scores indicate better speaker similarity preservation. The 0.6B model surprisingly outperforms 1.7B in some languages.

Hardware Requirements

GPU Memory Usage

Model	FP16/BF16	FP32
1.7B models	~4-6 GB	~8-12 GB
0.6B models	~2-3 GB	~4-6 GB
Tokenizer	~1-2 GB	~2-4 GB

Memory requirements increase with batch size and sequence length. Add 2-4 GB for FlashAttention 2 and generation overhead.

Inference Speed

Typical inference times (per second of audio):

Model	GPU	Speed (RTF)
1.7B	A100	~0.1x
1.7B	V100	~0.15x
0.6B	A100	~0.06x
0.6B	V100	~0.09x

RTF (Real-Time Factor): 0.1x means generating 1 second of audio takes 0.1 seconds. Lower is faster.

Model Downloads

Models are automatically downloaded when you call from_pretrained(). For manual downloads:

Hugging Face

pip install -U "huggingface_hub[cli]"

# Download specific model
huggingface-cli download Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign \
  --local-dir ./Qwen3-TTS-12Hz-1.7B-VoiceDesign

huggingface-cli download Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
  --local-dir ./Qwen3-TTS-12Hz-1.7B-CustomVoice

huggingface-cli download Qwen/Qwen3-TTS-12Hz-1.7B-Base \
  --local-dir ./Qwen3-TTS-12Hz-1.7B-Base

huggingface-cli download Qwen/Qwen3-TTS-Tokenizer-12Hz \
  --local-dir ./Qwen3-TTS-Tokenizer-12Hz

ModelScope

pip install -U modelscope

# Download specific model (recommended for users in Mainland China)
modelscope download --model Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign \
  --local_dir ./Qwen3-TTS-12Hz-1.7B-VoiceDesign

modelscope download --model Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
  --local_dir ./Qwen3-TTS-12Hz-1.7B-CustomVoice

modelscope download --model Qwen/Qwen3-TTS-12Hz-1.7B-Base \
  --local_dir ./Qwen3-TTS-12Hz-1.7B-Base

modelscope download --model Qwen/Qwen3-TTS-Tokenizer-12Hz \
  --local_dir ./Qwen3-TTS-Tokenizer-12Hz

Next Steps

Quick Start

Install and start using Qwen3-TTS models

Architecture

Understand how the models work internally

Language Support

Learn about multilingual capabilities

Fine-Tuning

Customize models for your specific use case