Qwen Audio 3.0 TTS is available on EmpirioLabs. It is Alibaba's speech synthesis model, and we ship it as a single model with a tier switch: Plus for the highest audio quality and expressiveness, Flash for real-time interaction with lower first-packet latency. Both tiers share the same parameters and the same voice library, so you pick a voice once and change tiers with one field.
What Qwen Audio 3.0 TTS is good at
It covers 16 languages, including Arabic, Chinese, English, French, German, Indonesian, Italian, Japanese, Korean, Malay, Portuguese, Russian, Spanish, Tagalog, Thai, and Vietnamese, plus 20 Chinese dialect regions. Delivery is directed two ways: a plain-English instruction that sets emotion, tone, character, pace, and style for the whole render, and inline tags such as [excited], [whispers], [laughing], or [sighing] dropped straight into the text to change expression mid-sentence.
The voice library is the reason the tier switch is painless. There are over 1,000 base voices, and every one of them exists on both tiers under the same id, so moving between Plus and Flash does not change who is speaking. On top of those, each tier has a small set of flagship system voices of its own.
Making a request
It is the standard OpenAI-shaped speech endpoint, so most clients need one base URL change:
curl https://api.empiriolabs.ai/v1/audio/speech \
-H "Authorization: Bearer $EMPIRIOLABS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen-audio-3-0-tts",
"model_tier": "flash",
"input": "[excited]The results are in, and they beat our forecast![laughing]",
"voice": "loongjameszhao",
"instruction": "Speak quickly in an upbeat, enthusiastic tone",
"response_format": "mp3",
"sample_rate": 24000
}'You get back a signed audio URL. Output supports MP3, WAV, headerless PCM, and Opus, at sample rates from 8 kHz up to 48 kHz. Speed, pitch, volume, and seed are all exposed, along with pronunciation overrides for names and acronyms, literal text replacements for brand names, and a Markdown filter so formatting characters are not read aloud.
To browse the full voice catalog, call GET /v1/voices. It supports filtering by language, gender, tier, and a free-text search across voice names, traits, and use cases.
Three things worth knowing before you build
Base voices are tier-portable, system voices are not. Any base voice id works on both Plus and Flash, so you can A/B the tiers with the voice held constant. The six flagship system voices are locked to one tier each, and if you send one to the wrong tier the API tells you which tier serves it rather than failing vaguely.
SSML is not supported by this model. Use instruction for global delivery direction and inline tags for local expression changes. That combination covers most of what people reach for SSML to do, and it is easier to author.
Pitch also changes pace. Lowering pitch stretches the clip and raising it compresses the clip, so if you want a different pitch at the original duration, adjust speed to compensate.
Pricing and getting started
Billing is per character of input text, at whichever tier you selected, and it is pay-as-you-go. Flash is the cheaper of the two. Current rates for both tiers are on the Qwen Audio 3.0 TTS model page and on the pricing page.
You can try it without writing any code in the Playground, where the tier switch, the voice picker, and every parameter are live controls. Full parameter reference is in the API documentation.
