DashScope provides cloud-hosted Qwen3-TTS APIs for fast, scalable, and production-ready text-to-speech generation. Skip infrastructure management and start generating speech in minutes.
Overview
The DashScope API offers:
- Managed infrastructure: No GPU setup or maintenance required
- Low latency: Optimized cloud deployment for fast generation
- High availability: Production-grade reliability and uptime
- Real-time streaming: Stream audio as it’s generated
- All model variants: Access to CustomVoice, VoiceClone, and VoiceDesign models
API Types
DashScope provides three API endpoints corresponding to the three Qwen3-TTS model types:
1. Custom Voice API
Generate speech using predefined speaker voices with optional instruction control.
Features:
- 9 premium speaker voices (Vivian, Serena, Ryan, Aiden, etc.)
- Multi-language support (Chinese, English, Japanese, Korean, and more)
- Natural language instruction control (emotion, tone, speaking style)
- Streaming and non-streaming modes
Documentation:
2. Voice Clone API
Clone any voice from a reference audio sample and generate new speech in that voice.
Features:
- 3-second rapid voice cloning
- High-fidelity speaker similarity
- Supports reference audio + transcript
- Cross-lingual voice cloning
Documentation:
3. Voice Design API
Create custom voices from natural language descriptions.
Features:
- Design voices with text descriptions (age, gender, tone, emotion)
- Fine-grained control over voice characteristics
- Generate unique voice profiles on-demand
- Support for creative and expressive voice styles
Documentation:
Quick Start
1. Get API Credentials
-
Sign up for DashScope account:
-
Create API key in the console
-
Note your API endpoint and key
2. Make Your First Request
Refer to the official documentation for complete API reference, authentication, and code examples:
API Comparison
| Feature | Custom Voice | Voice Clone | Voice Design |
|---|
| Predefined speakers | ✅ 9 speakers | ❌ | ❌ |
| Custom voice from audio | ❌ | ✅ | ❌ |
| Voice from description | ❌ | ❌ | ✅ |
| Instruction control | ✅ | ❌ | ✅ |
| Streaming | ✅ | ✅ | ✅ |
| Languages | 10 languages | 10 languages | 10 languages |
| Best for | Production apps with consistent voices | Voice cloning, personalization | Creative voice design |
Self-Hosted vs Cloud API
| Aspect | Self-Hosted (qwen-tts package) | DashScope API |
|---|
| Setup | Install package, download models | Sign up, get API key |
| Infrastructure | Requires GPU (16-24GB VRAM) | Managed cloud infrastructure |
| Scaling | Manual scaling, load balancing | Auto-scaling, built-in load balancing |
| Cost | GPU hardware/cloud compute costs | Pay-per-use API pricing |
| Latency | Local: lowest; Cloud: depends on network | Optimized cloud infrastructure |
| Maintenance | Self-maintained, model updates | Fully managed, automatic updates |
| Customization | Full control, fine-tuning support | API parameters only |
| Data privacy | Complete control, data stays local | Data sent to cloud (check privacy policy) |
When to Use Self-Hosted
- Need complete data privacy/control
- High-volume processing with predictable load
- Custom fine-tuning requirements
- Network-isolated environments
- Long-term cost optimization for stable workloads
When to Use DashScope API
- Fast prototyping and development
- Variable or unpredictable traffic
- No GPU infrastructure available
- Need high availability without ops overhead
- Small to medium scale deployments
- Want automatic model updates and optimizations
Pricing
For pricing information, contact Alibaba Cloud or refer to the official documentation:
Rate Limits
API rate limits vary by account tier. Check the official documentation for current limits:
Regional Availability
DashScope APIs are available in multiple regions:
- China: Beijing, Shanghai, Hangzhou, Shenzhen
- International: Singapore, US, Europe
Select the region closest to your users for optimal latency.
Best Practices
Error Handling
- Implement retry logic with exponential backoff
- Handle rate limit errors gracefully
- Log API errors for debugging
- Validate inputs before API calls
- Use streaming mode for real-time applications
- Batch requests when possible
- Cache frequently generated audio
- Choose nearest regional endpoint
Cost Optimization
- Cache common phrases and responses
- Use appropriate audio quality settings
- Monitor usage and set alerts
- Batch process when real-time isn’t required
Security
- Never expose API keys in client-side code
- Rotate API keys regularly
- Use environment variables for credentials
- Implement rate limiting on your application layer
- Validate and sanitize user inputs
Migration Guide
From Self-Hosted to API
Migrating from the qwen-tts package to DashScope API:
Before (Self-Hosted):
from qwen_tts import Qwen3TTSModel
import torch
model = Qwen3TTSModel.from_pretrained(
"Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice",
device_map="cuda:0",
dtype=torch.bfloat16,
)
wavs, sr = model.generate_custom_voice(
text="Hello world",
language="English",
speaker="Ryan",
)
From API to Self-Hosted
Reverse migration for data privacy or cost reasons:
- Set up GPU infrastructure
- Install qwen-tts package
- Download model weights
- Adapt API calls to library calls
- Implement caching and optimization
Support
For DashScope API support:
- Documentation: DashScope API Docs
- Technical Support: Alibaba Cloud support portal
- Community: Qwen GitHub Issues and Discord
For self-hosted model support:
Official Documentation Links
For complete API reference, authentication details, request/response formats, and code examples, please refer to the official DashScope documentation linked above.
Next Steps