Name age, gender and accent
Start with who the speaker is. Age and accent change the voice more than any other words.

Qwen Audio is Alibaba’s family of speech models. This Qwen Audio studio puts it in your browser: type text, pick one of 26 Qwen TTS voices, clone a voice from a short recording or describe a new one, then download the audio.
Independent studio · Not affiliated with Alibaba Cloud or the Qwen team · Runs Qwen3-TTS-Flash and Qwen3-TTS
Premium, cinematic American English voice
Qwen Audio runs Alibaba's Qwen3-TTS-Flash and the open Qwen3 TTS models through the WaveSpeed API. Independent site, not affiliated with Alibaba.
Each example was made in this Qwen Audio studio. Tap the wand to load its voice and text, then press Generate.
Product launch
Jennifer · Qwen3-TTS-Flash · English
Movie trailer
Ryan · Qwen3-TTS-Flash · English
Bedtime story
Aiden · Qwen3-TTS · English
Science explainer
Elias · Qwen3-TTS-Flash · English
App welcome
Cherry · Qwen3-TTS-Flash · English
Spanish cooking show
Katerina · Qwen3-TTS-Flash · Spanish
Every Qwen TTS take costs 1 credit per 200 characters, in every mode.
Every take below was made in this Qwen Audio studio with the settings it shows. Press play to hear Qwen TTS, or tap the wand to load the same voice and text into the generator.








Qwen Audio is the name Alibaba’s Qwen team now uses for its speech models. The line started with models that listen and grew into Qwen TTS models that speak.
Text
up to 1,000 chars
Your voice
3–30 s sample
Description
a new voice
Qwen3-TTS · Flash
26
Built-in voices
11
Languages
3 s
To clone a voice
10
Signup credits
The first Qwen Audio model, Qwen-Audio (November 2023), took sound in and wrote text out, and Qwen2-Audio added voice chat in 2024. Qwen TTS output started with Qwen3-TTS-Flash in September 2025, followed by voice cloning and voice design in December 2025.
On January 22, 2026 Alibaba open-sourced Qwen3 TTS under Apache 2.0: 1.7B and 0.6B models with 9 built-in voices, 3-second voice cloning and voice design in 10 languages. That release is why most people search for Qwen TTS today.
The newest generation carries the Qwen Audio name itself. Qwen-Audio-3.0-TTS (July 2026) and the Qwen-Audio-3.1 models (September 2026) are hosted on Alibaba Cloud Model Studio only, with no open weights.
This studio runs Qwen3-TTS-Flash and the open Qwen3 TTS models through the WaveSpeed API, so you can use Qwen Audio without a GPU, an install or a cloud account. We are an independent site with no link to Alibaba.
Three names, three different models. Here is how they relate and which ones this Qwen Audio studio runs.
Qwen TTS is the everyday name for the whole Qwen speech line. Qwen3-TTS-Flash is Alibaba’s fast hosted model with 17 voices; in the Qwen Audio studio it renders a take in about 5 to 20 seconds.
Qwen3 TTS is the open Qwen TTS release from January 2026. Its built-in voices follow a style instruction, and the same family clones a voice from a few seconds of audio or builds one from a description. Takes run 30 to 90 seconds.
Qwen-Audio-3.0-TTS is the newer hosted model with 16 languages. It is not open source and not on WaveSpeed, so this Qwen Audio studio does not offer it yet; we will add it once we can run the real model.
The same price for every voice and mode. Every plan below shows its price per credit.
| Topic | Qwen3-TTS-Flash | Qwen3 TTS (open) | Qwen-Audio-3.0-TTS |
|---|---|---|---|
| Released | Sep 2025 | Jan 22, 2026 | Jul 20, 2026 |
| Open weights | No | Yes, Apache 2.0 | No |
| Voices here | 17 | 9 + clone + design | — |
| Languages | 11 | 10 | 16 |
| Style instruction | No | Yes | Yes |
| In this studio | Text to speech | Speech, clone, design | Not offered |
Release dates and specs from the Qwen GitHub, Qwen blog, Alibaba Cloud Model Studio and WaveSpeed model pages, checked Nov 18, 2026.
What this Qwen Audio studio supports today. Each value below is a real setting in the generator above.
Need speech to text, live voice chat or a longer audiobook in one file? Split long scripts into takes of 1,000 characters; Qwen ASR and realtime voice are not part of this Qwen Audio studio yet.
Record or upload a short sample and Qwen3 TTS reads any new text in that voice. Below, one of our own Flash voices was used as the reference, so no real person’s voice was copied.

Reference: Katerina (Flash)
Cloned take · new text
Use 3 to 30 seconds of one person speaking clearly, with no music or crowd noise. Ten seconds is plenty.
Add a transcript of the sample if you can. The Qwen voice clone model copies tone and rhythm more closely when it knows the words.
Write the new text in any of the 10 supported languages; the cloned voice carries over.
Clone only your own voice or a voice you have permission to use. Never use it to impersonate or mislead anyone.
Your sample goes straight to the model provider for this take. Qwen Audio keeps the finished audio in your account, not the recording.
Clone a voiceNo sample needed: write who is speaking and Qwen3 TTS builds that voice. These four habits get better results; every example was made here.
Start with who the speaker is. Age and accent change the voice more than any other words.
Deep, bright, breathy, gravelly or husky: say what the voice sounds like, not just how it feels.
Fast and loud or slow and calm. Qwen TTS follows the description, so match it to the text.
A late-night radio host or a museum guide: a role tells the model how close and how formal to sound.
The Qwen3 TTS weights are free to download. Here is where to run Qwen TTS yourself, and when the online Qwen Audio studio is the quicker route.
QwenLM/Qwen3-TTS has the code, a quick start and the model list: Base, CustomVoice and VoiceDesign, in 1.7B and 0.6B sizes.
OpenThe Qwen3-TTS collection hosts the weights and the official demo Space. Expect a queue at busy hours.
OpenCommunity nodes run Qwen TTS inside ComfyUI workflows, handy if you already make video or images there.
The README sets no minimum, but recommends FlashAttention 2 with float16 or bfloat16, so plan on a recent NVIDIA GPU. The 0.6B models are the lighter choice.
No GPU, or no time for setup? This Qwen Audio studio runs the same Qwen3 TTS models in the cloud and gives you a file to download.
Three steps from text to an audio file with Qwen Audio. New accounts get 10 free credits, enough for two 1,000-character takes.
1Text to Speech uses a built-in Qwen TTS voice, Voice Clone copies a recording and Voice Design builds a voice from your words.
2Choose Flash for speed or Qwen3 TTS for style control. Leave the language on Auto or set it for mixed text.
3The credit cost shows before you click. Play the take, download it, or reuse its settings for the next line.
Qwen TTS shines on short, expressive reads. Mix voices, languages and styles for each part of a project.

Narrate YouTube videos, Shorts and product demos with Qwen Audio. Keep each take to a scene, then line them up in your editor.

Give each character a voice with Voice Design and read chapters in parts. A slow, soft style suits bedtime stories.

Record intros, sponsor reads and segment stingers in a voice that stays the same every episode.

Turn lesson notes into Qwen TTS audio, or hear the same sentence in 11 languages with one voice.

Prototype NPC lines, narrators and villains before you book voice actors. Describe the role and let Qwen Audio voice it.
Every Qwen Audio take costs 1 credit per 200 characters, the same for every voice and mode. Your first two takes are on us; after that, top up or subscribe.
Answers on Qwen Audio, Qwen TTS, Qwen3 TTS, voice cloning, pricing and running the models yourself.
Yes. Qwen Audio is Alibaba’s speech family: Qwen-Audio and Qwen2-Audio understand sound, while Qwen3 TTS and Qwen-Audio-3.0-TTS speak. This studio lets you use Qwen TTS online.
Yes. Qwen TTS covers Qwen3-TTS-Flash, the open Qwen3 TTS models and the hosted Qwen-Audio-3.0-TTS. Here you can run the first two in your browser.
Qwen3 TTS is one of the strongest open text to speech models: natural prosody, 10 languages, voice cloning and voice design. Listen to the examples above and judge for yourself.
No. Qwen-Audio-3.0-TTS is only offered through Alibaba Cloud Model Studio. The open model is Qwen3 TTS (Apache 2.0), which this studio runs.
New accounts get 10 free credits, enough for two 1,000-character Qwen TTS takes in any mode. After that, credit packs and plans are on the pricing page.
Qwen does not publish a minimum. The README recommends FlashAttention 2 with float16 or bfloat16, so a recent NVIDIA GPU is the safe choice; the 0.6B models need less memory than 1.7B. Online, you need no GPU at all.
Install the qwen-tts package from the QwenLM/Qwen3-TTS GitHub repo and load a model from Hugging Face, or skip the setup and use the Qwen Audio studio on this page.
Qwen3.5-Omni can listen and talk, but it is a chat model. For reading text aloud with a chosen voice, Qwen TTS models like Qwen3 TTS are the right tool.
Up to 1,000 characters, about a minute of speech. For longer scripts, split the text into parts and join the files in any audio editor.
Audio you make here follows our terms of service. Qwen3 TTS is released under Apache 2.0. For cloned voices you need the speaker’s permission.
No. A Qwen TTS clone sample is sent to the model provider for that take only; we store the finished audio in your account, and you can delete it any time.
No. This Qwen Audio studio (QwenAudio) is independent and not affiliated with or endorsed by Alibaba Cloud. The official Qwen site is qwen.ai.
Sign in with Google, pick a Qwen TTS voice and hear your text in seconds. Clone your own voice or design a new one, with no GPU and no install.
Try for free