Speech synthesis overview

​

Migrating from proprietary models?

If you're currently using ElevenLabs, OpenAI, or Google for speech synthesis, use the following table to find the equivalent QwenCloud model.

​

Standard TTS or custom voice?

Speech synthesis models convert text into natural-sounding speech. Start by deciding whether built-in voices or custom voices fit your needs:

• Use standard TTS when the preset voice library meets your needs and you want quick setup with no extra configuration.
• Use a custom voice when you need a brand-specific voice, want to replicate a specific person's voice, or need to create an entirely new character voice.

​

Voice cloning or voice design?

If you choose a custom voice, two creation methods are available:

• Use voice cloning when you have a recording of the target speaker and want to reproduce that voice in synthesized speech.
• Use voice design when you don't have a recording and want to create a new voice from a text description.

​

WebSocket or HTTP?

• WebSocket: Bidirectional streaming — supports streaming input and output, delivering audio as it's synthesized to minimize latency. Ideal for customer service bots, voice assistants, and call centers.
• HTTP: Send complete text, receive audio in a streaming response (chunked output). Ideal for audiobooks, audio content production, and other high-volume batch scenarios.

Qwen-Audio-TTS/CosyVoice models use the same model name for both WebSocket and HTTP access. For Qwen models, the model name indicates the API: models with the -realtime suffix use WebSocket, while those without it use HTTP. Access Qwen-Audio-TTS/CosyVoice and Qwen WebSocket models through the DashScope SDK (Java, Python). Other models require direct WebSocket or HTTP calls. For WebSocket access, see Real-time speech synthesis. For HTTP access, see Non-real-time speech synthesis. CosyVoice series models also support the AOQ protocol. For client-side integration that prioritizes stable latency, resilience on weak networks, and built-in full-duplex noise suppression and echo cancellation, AOQ is recommended. For a protocol comparison, see Realtime API overview.

​

Instruction control

Describe the desired expression style in natural language to control speed, emotion, and style per request. Examples include "speak gently at a slower pace" or "use an excited broadcast style." Use instruction control for emotional content production, professional broadcasting, audiobooks, and other scenarios that require rich expressiveness. Models that support instruction control: Qwen-Audio-TTS series (qwen-audio-3.1-tts-plus, qwen-audio-3.1-tts-flash, qwen-audio-3.0-tts-plus, qwen-audio-3.0-tts-flash), CosyVoice series (cosyvoice-v3.5-plus, cosyvoice-v3.5-flash, cosyvoice-v3-flash). For details, see Real-time speech synthesis > Instruction control.

​

Recommended models

For voice assistants, real-time conversations, and customer service, prioritize qwen-audio-3.1-tts-flash: it targets low-latency interaction and supports system voices, voice cloning, voice design, natural language instructions, and emotion and rich-language tags. For audiobooks, film dubbing, brand voices, and content production, prioritize qwen-audio-3.1-tts-plus: it emphasizes audio quality, naturalness, and detail, with system voices, voice cloning, voice design, and rich expression control. The following table lists the best model for each scenario. For full details, visit the Model Gallery. cosyvoice-v3.5-plus is available only in China North 2 (Beijing). For model availability by region, see Regions.

​

Apply in production

For text submission, audio reception and playback, connection and concurrency management, performance evaluation, and troubleshooting, see Speech synthesis best practices.

​

All models

​

Qwen-Audio-TTS

For connection methods and system voices for qwen-audio-3.1-tts-flash, see Qwen-Audio-TTS SDK access and Qwen-Audio-TTS voice list. Supported languages:

• System voices (varies by voice): Chinese (Mandarin), English
• Cloned voices (dialects are configured through instruction control): Chinese (Mandarin, Cantonese, Chongqing, Northeastern, Gansu, Guizhou, Zhejiang, Hebei, Henan, Hubei, Hunan, Jiangxi, Ningbo, Ningxia, Qingdao, Shaanxi, Shanxi, Shandong, Shanghainese, Sichuan, Yunnan), English, Japanese, Korean, Russian, French, German, Portuguese, Thai, Indonesian, Vietnamese, Spanish, Italian, Malaysian, Filipino, and Arabic
• Other languages (supported by qwen-audio-3.1-tts-flash): Bulgarian, Dutch, Hungarian, Romanian, Danish, Croatian, Greek, Czech, Norwegian, Swedish, Hindi, Slovak, Finnish, Urdu, Serbian, Mexican Spanish, Lithuanian, Latvian, Turkish, Persian, Hebrew, Mongolian, Slovenian, Azerbaijani, Polish, Catalan, Khmer, Brazilian Portuguese, Lao, Georgian, Afrikaans, Nepali, Belarusian, Malayalam, Luxembourgish, Albanian, Tamil, Sinhala, Telugu, Estonian, Cebuano, Bosnian, Burmese, Armenian, Latin, Icelandic, Maltese, Swahili, Kazakh, Seychellois Creole, Ukrainian.

​

CosyVoice

Some CosyVoice models support SSML (Speech Synthesis Markup Language) markup and LaTeX formula reading. cosyvoice-v3.5-plus, cosyvoice-v3.5-flash, and cosyvoice-v2 are available only in China North 2 (Beijing). Supported languages (by version):

• cosyvoice-v3.5-plus and cosyvoice-v3.5-flash (no system voices):
  • Cloned voices (dialects are configured through instruction control): Chinese (Mandarin, Cantonese, Northeastern, Gansu, Guizhou, Henan, Hubei, Jiangxi, Hokkien, Ningxia, Shanxi, Shaanxi, Shandong, Shanghainese, Sichuan, Tianjin, Yunnan), English, French, German, Japanese, Korean, Russian, Portuguese, Thai, Indonesian, and Vietnamese
  • Designed voices: Chinese (Mandarin), English
• cosyvoice-v3-plus:
  • System voices (varies by voice): Chinese (Mandarin), English
  • Cloned voices (dialects are configured through instruction control): Chinese (Mandarin, Cantonese, Northeastern, Gansu, Guizhou, Henan, Hubei, Jiangxi, Hokkien, Ningxia, Shanxi, Shaanxi, Shandong, Shanghainese, Sichuan, Tianjin, Yunnan), English, French, German, Japanese, Korean, and Russian
  • Designed voices: Chinese (Mandarin), English
• cosyvoice-v3-flash:
  • System voices (varies by voice; some dialects are directly supported by system voices, others through instruction control): Chinese (Mandarin, Cantonese, Northeastern, Henan, Hunan, Shaanxi, Shandong, Sichuan, Anhui, Hokkien), English
  • Cloned voices (dialects are configured through instruction control): Chinese (Mandarin, Cantonese, Northeastern, Gansu, Guizhou, Henan, Hubei, Jiangxi, Hokkien, Ningxia, Shanxi, Shaanxi, Shandong, Shanghainese, Sichuan, Tianjin, Yunnan), English, French, German, Japanese, Korean, Russian, Portuguese, Thai, Indonesian, and Vietnamese
  • Designed voices: Chinese (Mandarin), English
• cosyvoice-v2 (voice design not supported):
  • System voices (varies by voice): Chinese (Mandarin, Cantonese, Northeastern, Hokkien, Shaanxi), English, Japanese, and Korean
  • Cloned voices: Chinese (Mandarin), English

​

Qwen3-TTS

Supported languages (by version):

• Qwen3-TTS-Flash series (system voices) (qwen3-tts-flash, qwen3-tts-flash-realtime): Chinese (Mandarin, Beijing, Shanghainese, Sichuan, Nanjing, Shaanxi, Hokkien, Tianjin, Cantonese — varies by voice), English, German, Italian, Portuguese, Spanish, Japanese, Korean, French, and Russian