Multilingual, directable speech synthesis
Learn how Qwen Audio 3.0 TTS handles 16 languages, natural-language voice direction, inline expression tags, long-form generation, and robust voice cloning.
Qwen Audio 3.0 TTS is a production-oriented speech synthesis system designed for consistent content, recognizable speaker identity, natural prosody, controllable delivery, multilingual output, and difficult real-world reference audio. It uses a low-frame-rate speech tokenizer to reduce decoding work while retaining speech content and speaker information.
The system supports both broad instructions and local edits. You can describe an overall performance in natural language, then place tags in the script where a laugh, breath, emotion shift, or other non-verbal detail should occur.
Facts summarized from the official Qwen Audio 3.0 TTS release.
The practical workflow combines a script, a selected or authorized voice, an optional performance instruction, and a model variant matched to the job.
Start with the script you want to synthesize. Select a preset voice or provide an authorized reference recording when you need to preserve a specific speaker identity.
Write a natural-language instruction for emotion, role, pace, accent, or speaking scenario. Add inline tags such as [giggles], [gasp], or [angry] when a specific phrase needs local control.
Use Flash for interactive experiences where response time matters. Use Plus for finished narration and production work where naturalness and timbre fidelity take priority.
The system combines an efficient speech representation, staged training, directable performance control, broad deployment support, and evaluation across real speech-production conditions.
A supervised 12.5 Hz speech tokenizer lowers the amount of autoregressive decoding needed for synthesis while retaining the content of the speech and the identifying qualities of the speaker.
Click the record to listen
The training process combines separate language-model and flow-matching pretraining, joint training with high-quality data annealing, language-model reinforcement learning, flow-matching robustness training, and final reinforcement learning. These stages target content consistency, natural prosody, voice fidelity, audio quality, and robustness together.
Click the record to listen
Qwen Audio 3.0 TTS follows free-form instructions for role, emotion, speaking style, rate, timbre, and accent. It also adds 86 fine-grained inline tags for phrase-level and word-level control, including transitions, laughter, breathing, coughing, and sighing.
Click the record to listen
Qwen Audio 3.0 TTS supports 16 languages, including seven newly added languages. The model also handles difficult text normalization, one-pass synthesis up to about three minutes, and degraded reference prompts without a separate denoising mode. Speaker fine-tuning and vocoder super-resolution support target-voice adaptation and 48 kHz output.
Click the record to listen
The reported evaluation covers zero-shot voice cloning, multilingual and cross-lingual synthesis, free-form instruction following, fine-grained control, text normalization, long-form generation, and robustness to difficult reference audio. Results combine objective benchmarks with arena-style human evaluation.
Click the record to listen
Both variants share the Qwen Audio TTS family, but they prioritize different production needs. Choose based on whether response time or finished-audio quality matters more.
| Comparison | FlashReal-time interaction | PlusHigher-quality generation |
|---|---|---|
| Positioning | Tuned for real-time interaction and conversational experiences. | Tuned for finished audio where generation quality takes priority. |
| Best for | Interactive speech | Finished narration |
| Primary advantage | Lower response delay | Higher priority on naturalness |
| Response profile | First-packet latency at the 300 ms level, as described in the release article. | Prioritizes naturalness, speaker similarity, and timbre fidelity over the fastest first packet. |
| Typical uses | Voice assistants, interactive characters, and live experiences | Dubbing, narration, and production audio |
Variant positioning and latency wording are based on the Tongyi Lab release article.
Preview eight speech styles across narration, multilingual delivery, character performance, podcasts, and production voiceover.
Audiobook and long-form speech
Global news and announcements
Expressive character performance
Product and campaign audio
Stories for families and children
Meditation and wellbeing audio
Natural multi-speaker dialogue
Games and cinematic scenes
The same control system can support localized content, long-form narration, and expressive characters without reducing every performance to a fixed preset.
Keep a consistent speaker identity while generating localized dialogue in multiple languages. Natural-language direction can adapt the performance to a documentary, product video, drama, or short-form clip without building a separate acoustic preset for every scene.
Generate audiobook excerpts, training material, explainers, and editorial narration with control over pace and tone. One-pass long-form synthesis reduces the need to split every passage into short fragments before generation.
Direct character voices with descriptions of role, mood, accent, and speaking style. Inline tags help place laughter, breaths, gasps, and emotional changes at the exact point where the script calls for them.
Buy credits once and use them across Qwen text to speech, AI Voice Design, AI Voice Cloning, and multilingual voice generation. Your credits do not expire.
One-time credit packs with no recurring subscription
Direct answers about languages, model variants, style control, voice cloning, and access.
Qwen Audio 3.0 TTS is a production-oriented text-to-speech system from Qwen Audio, Token Foundry, and Alibaba Group. It combines multilingual speech generation, voice cloning, natural-language direction, inline control tags, and long-form synthesis.
Qwen Audio 3.0 TTS supports 16 languages: Chinese, Arabic, English, French, German, Indonesian, Italian, Japanese, Korean, Malay, Portuguese, Russian, Spanish, Tagalog, Thai, and Vietnamese. The announcement notes that full language support is rolling out.
Flash is tuned for interactive use and lower first-packet latency. Plus is tuned for higher-quality generation where naturalness and timbre fidelity are more important than the fastest response.
Yes. Qwen Audio 3.0 TTS supports voice cloning and cross-lingual generation. Only use recordings you own or have permission to use, and disclose synthetic speech where applicable.
You describe how the line should sound in normal language, including emotion, role, pace, timbre, accent, or scenario. The model uses that instruction to guide the overall performance without requiring manual acoustic settings.
Inline tags are markers placed inside the target text to control a local moment. The release adds 86 tags for details such as laughter, breathing, coughing, sighing, emotion changes, and other non-verbal events.
This page provides a Qwen Audio 3.0 TTS generator for English system voices. Choose Flash or Plus, preview a voice, set the delivery instruction and audio controls, then generate and download the result.
Read the first-party release for technical results and audio demos, or try the Qwen3 TTS tools available on this website for text-to-speech, voice design, and voice cloning workflows.