How to Use the Qwen Audio 3.1 TTS Next API

Disclosure: This article was written with AI assistance and reviewed by EmpirioLabs AI.

Qwen Audio 3.1 TTS Next is available on EmpirioLabs. It is Alibaba's audio generation model. Describe a scene in plain language, including who speaks, what they say, how they sound, and what can be heard around them, and it returns one finished audio file with the narration, dialogue, sound effects, ambience, and music together. It runs on POST /v1/audio/speech, in the API and in the Playground.

Watch the launch film

Every sound in the film is a real Qwen Audio 3.1 TTS Next output, each scene from one written description: a night market in the rain, a haunted hallway at midnight, and a song by a campfire. The pictures are Qwen Image 2.1 outputs, and the score was generated by ACE-Step on EmpirioLabs. Watch it on YouTube.

How it differs from Qwen Audio 3.1 TTS

Qwen Audio 3.1 TTS reads your text aloud in one of 19 preset voices and can stream it as it is generated. Qwen Audio 3.1 TTS Next has no voice list: you describe each voice in input or clone it from a reference clip, and the same description adds sound effects, ambience, and music. Use Qwen Audio 3.1 TTS to read text as written in a fixed voice, and Qwen Audio 3.1 TTS Next when the result is a produced scene.

Your first request

curl https://api.empiriolabs.ai/v1/audio/speech \
  -H "Authorization: Bearer $EMPIRIOLABS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen-audio-3-1-tts-next",
    "input": "A calm female narrator says: \"Every morning, the harbor wakes before the town does.\" Gentle waves and distant gulls in the background."
  }'

The response carries a URL to the generated audio, MP3 at 48,000 Hz in stereo by default. response_format also takes wav or pcm, sample_rate takes 8,000 to 48,000 Hz, and channels takes 1 for mono or 2 for stereo. speed, volume, and seed are adjustable, bit_rate sets a constant MP3 bitrate, and watermark adds an AI-generated content watermark.

Writing the description

• Voices. Describe each speaker, such as a calm female narrator or an excited child, and put each spoken line in quotation marks.
• Delivery. Say how each line is spoken, for example whispers, answers warmly, or cheerfully says.
• Sound. Add sound effects, ambience, and music in the same text, such as rain tapping on a window or soft piano in the background.
• Languages. Chinese and English.
• Length. Up to 3,000 characters per request and up to 120 seconds of audio, or 240 seconds for a podcast.

Voices from reference clips

references takes up to three clips as URLs or base64 audio data URIs, each up to 30 seconds and 10 MB, in WAV, MP3, or OGG Opus. Name them in input as @voice1, @voice2, and @voice3, in the order you send them. Two clips make a two-voice dialogue:

curl https://api.empiriolabs.ai/v1/audio/speech \
  -H "Authorization: Bearer $EMPIRIOLABS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen-audio-3-1-tts-next",
    "references": ["https://example.com/visitor.mp3", "https://example.com/bookseller.mp3"],
    "input": "In a quiet bookshop, @voice1 whispers: \"Have you seen the new poetry shelf?\" @voice2 answers warmly: \"I set it up this morning. Take your time.\" A page turns, and rain taps on the window.",
    "response_format": "wav",
    "sample_rate": 24000,
    "channels": 1
  }'

Operational notes

• All direction goes in input. A request with instruction or with SSML is rejected with a 400 that says so.
• Every @voice needs its clip. @voice2 in input needs at least two clips in references; otherwise the request is rejected.
• No preset voices. voice is accepted so OpenAI-compatible clients work unchanged, and it has no effect. Describe the voice in input.
• One complete file per request. This model does not stream, and POST /v1/audio/speech:stream returns 400. A short clip takes about 15 to 40 seconds, so allow a client timeout of several minutes for long scenes.
• Formats. MP3, WAV, and PCM. Opus is not offered.

Pricing

Token billing. Input tokens cover the description and each reference clip, and output tokens cover the generated audio, about 200 per second. A request with reference clips bills more input tokens than the same description without them. Current rates are on the model page and the pricing page.

Start building

Try Qwen Audio 3.1 TTS Next in the Playground, where every parameter is a live control, and read the API documentation for the full parameter reference.