Generate natural, expressive speech online with Alibaba Qwen-Audio-3.0-TTS. Create multilingual text-to-speech audio, clone voices, use Chinese dialects, emotion tags, and instruction control for videos, podcasts, games, audiobooks, and product narration.
Trusted by 12,000+ creators
Voice workbench
Shape the script, voice, and delivery in one calm, focused workspace.
Model guide
All three are speech synthesis model families available through Alibaba Cloud Model Studio, but they are separate families rather than consecutive versions of one model. Their voice sources, voice cloning, voice design, instruction control, and integration methods differ.
DS Speech provides online voice generation powered by Qwen-Audio-3.0-TTS. The family supports text to speech, voice cloning, voice design, and natural-language instruction control for emotion, tone, character, speed, and delivery. Cloned voices support multiple languages and many Chinese dialects, making them useful for video narration, podcasts, audiobooks, game characters, and branded voices.
Qwen3-TTS belongs to the Qwen-TTS family and includes separate models for standard synthesis, instruction control, voice design, and voice cloning. It uses different model names and some different integration methods from Qwen-Audio-3.0-TTS, so voices and API parameters are not interchangeable.
CosyVoice is another independent speech synthesis family with real-time and non-real-time generation. Different versions provide voice cloning, voice design, and instruction control, with integration options designed for real-time interaction, voice clients, and more demanding network conditions.
DS Speech currently focuses on the Qwen-Audio-3.0-TTS online experience, so you can select voices, clone a voice, control delivery, and generate audio without configuring models, APIs, or storage yourself.
Simple, transparent pricing
Membership allowance and credit packs reset each calendar month. Every plan uses USD pricing.
The first-purchase offer can be used once per user. Yearly plans show the full annual charge; the monthly equivalent is for comparison only.
Creator reviews
From short videos and podcasts to games, courses, and branded content, creators use DS Speech to produce natural, expressive AI narration faster.
Qwen-Audio-3.0-TTS is fast enough for daily production. I can revise a script and hear a stable voice result in seconds instead of booking another recording session.
林澈Short-form Video DirectorThe emotional delivery does not feel robotic. A short instruction gives podcast intros, ads, and transitions a much more intentional tone.
Mika ChenPodcast ProducerThe public voice library covers character narration, NPC dialogue, and tutorial prompts, so reviews no longer rely on temporary machine voices.
周远Game Narrative DesignerDialect and multilingual text to speech make regional training content sound closer to the audience and less generic than ordinary TTS.
Nina HayesTraining ProducerWe use voice cloning for course series. When an instructor cannot record again, updates still keep a consistent voice across lessons.
何嘉宁Corporate Learning ProducerQwen-Audio-3.0-TTS is fast enough for daily production. I can revise a script and hear a stable voice result in seconds instead of booking another recording session.
林澈Short-form Video DirectorThe emotional delivery does not feel robotic. A short instruction gives podcast intros, ads, and transitions a much more intentional tone.
Mika ChenPodcast ProducerThe public voice library covers character narration, NPC dialogue, and tutorial prompts, so reviews no longer rely on temporary machine voices.
周远Game Narrative DesignerDialect and multilingual text to speech make regional training content sound closer to the audience and less generic than ordinary TTS.
Nina HayesTraining ProducerWe use voice cloning for course series. When an instructor cannot record again, updates still keep a consistent voice across lessons.
何嘉宁Corporate Learning ProducerI use AI voice cloning to test ad reads quickly and compare several scripts and delivery directions in one afternoon.
Alex MorganGrowth LeadLong-form text generation is easy to manage, and emotion tags help shape each chapter before the final pacing pass.
宋怡Audiobook EditorA cloned character voice stays consistent across episodes, which makes ongoing series much easier to produce.
Kenta MoriAnimation Channel ProducerWe use online text to speech for product explainers and launch teasers, even when the final script changes close to release.
陈玥Brand Content StrategistThe voice plaza has enough range for explainers, shorts, and product clips without starting a voice search from scratch every time.
Oliver GrantVideo ProducerI use AI voice cloning to test ad reads quickly and compare several scripts and delivery directions in one afternoon.
Alex MorganGrowth LeadLong-form text generation is easy to manage, and emotion tags help shape each chapter before the final pacing pass.
宋怡Audiobook EditorA cloned character voice stays consistent across episodes, which makes ongoing series much easier to produce.
Kenta MoriAnimation Channel ProducerWe use online text to speech for product explainers and launch teasers, even when the final script changes close to release.
陈玥Brand Content StrategistThe voice plaza has enough range for explainers, shorts, and product clips without starting a voice search from scratch every time.
Oliver GrantVideo ProducerLonger scripts fit into each generation, so the text needs fewer cuts and the finished delivery sounds more connected.
梁书瑶Educational CreatorMultilingual AI narration is much faster now. We can compare English, Chinese, Japanese, and Korean versions in the same review.
Sophie CarterBrand StrategistRegional accents matter for local content. Comparing standard Mandarin and dialect versions helps the team choose what feels closest to the audience.
高铭Local Content ManagerEmotion tags make it quick to separate tense dialogue from calm guidance, giving each character more expressive range.
Haruto SatoGame Audio DirectorCourse updates depend on speed. I can revise a lesson and regenerate clean narration without scheduling a full recording session.
Rachel MooreCourse CreatorLonger scripts fit into each generation, so the text needs fewer cuts and the finished delivery sounds more connected.
梁书瑶Educational CreatorMultilingual AI narration is much faster now. We can compare English, Chinese, Japanese, and Korean versions in the same review.
Sophie CarterBrand StrategistRegional accents matter for local content. Comparing standard Mandarin and dialect versions helps the team choose what feels closest to the audience.
高铭Local Content ManagerEmotion tags make it quick to separate tense dialogue from calm guidance, giving each character more expressive range.
Haruto SatoGame Audio DirectorCourse updates depend on speed. I can revise a lesson and regenerate clean narration without scheduling a full recording session.
Rachel MooreCourse CreatorQwen-Audio-3.0-TTS is a high-performance speech synthesis family available through Alibaba Cloud Model Studio. It supports online text to speech, voice cloning, voice design, and instruction-based delivery control. It is a separate model family from Qwen3-TTS and CosyVoice. DS Speech focuses on making Qwen-Audio-3.0-TTS voice cloning and AI speech generation easy to use online.
Enter text in the workbench, choose a preset voice, or upload a clear reference audio clip to save your own voice model, then generate a preview.
Yes. Before using someone else's voice, make sure you have legal authorization from the voice rights holder. If you use your own voice, or a voice generated by you that does not involve third-party rights, it can be used for commercial projects without additional commercial-use restrictions from us.
Chinese (Mandarin, Guangdong dialect, Chongqing dialect, Northeastern dialect, Gansu dialect, Guizhou dialect, Zhejiang dialect, Hebei dialect, Henan dialect, Hubei dialect, Hunan dialect, Jiangxi dialect, Ningbo dialect, Ningxia dialect, Qingdao dialect, Shaanxi dialect, Shanxi dialect, Shandong dialect, Shanghai dialect, Sichuan dialect, and Yunnan dialect), English, Japanese, Korean, Russian, French, German, Portuguese, Thai, Indonesian, Vietnamese, Spanish, Italian, Malaysian, Filipino, and Arabic.
Supported formats: WAV, MP3, and M4A. Recommended duration is 10~20 seconds, with a maximum of 60 seconds. File size must be ≤ 10 MB, and sample rate must be ≥ 16 kHz.
For stereo audio, only the first channel is processed, so make sure the first channel contains valid speech.
The audio must contain at least 5 seconds of continuous, clear reading without background sound. Other parts may only contain short pauses of ≤ 2 seconds. Avoid background music, environmental noise, or other voices. Use normal spoken audio; do not upload songs or singing recordings.
Do not use the product for fraud, impersonation, harassment, hate, obscene or pornographic content, or any illegal scenario. If discovered, your account rights will be terminated immediately.
Membership allowance and separately purchased credit packs are cleared each calendar month. Please plan usage within the current month.
Yes. Tags and instruction prompts can control some emotion, dialect, and character styles.
Instruction examples:
Standard broadcast style: clear and precise articulation.
Young energetic female voice: faster pace with an upward intonation, suitable for fashion products.
Calm middle-aged male: slow pace and deep magnetic tone, suitable for news or documentary narration.
Chinese dialect instruction example:
请用河南话表达
Tag example:
[excited]The weather is great today![laughing]Let's go out together!
No. After your membership expires or you downgrade from Pro to Plus, your created voice models will be retained and not deleted. However, you can only use the most recently created voice models within the quantity supported by your current membership level.
For example, if you were a Pro member and created 30 voice models, after downgrading to Plus you can continue using the 10 most recently created voice models. The remaining models are retained but temporarily unavailable.
When you reactivate Pro, all 30 original voice models will become available again.
Audio files generated with text-to-speech are retained for only 3 days. After that, they can no longer be played or downloaded, so please download and store them promptly.
