Qwen Audio 3.1 TTS Next vs StepAudio 3 Gen: one radio scene

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

Short answer: from the same written description, Qwen Audio 3.1 TTS Next performed all three spoken lines and none of the stage directions. StepAudio 3 Gen speaks the text of its prompt as written, so it read the stage directions aloud and stopped partway through the second line. Given the scene in its own multi-speaker format, StepAudio 3 Gen performed the first two lines, returned its file in about 5 seconds against about a minute, and cost well under half as much. Both run on EmpirioLabs. Listen to each clip below; we are not declaring a winner.

The scene

A late-night radio studio, with soft jazz playing in the background. A warm male host says: "Good evening, and welcome back to Night Owls. Our first caller is on the line." A phone rings once. A cheerful female caller says: "Hi! I have been listening for years, and I finally got through!" The host laughs and says: "Well, we are glad you did."

Listen to the results

Results at a glance

How we ran it

Each model received the same written scene, one shot each, through the EmpirioLabs API on October 18, 2026, with every other setting at its default. Qwen Audio 3.1 TTS Next took the scene as input, and StepAudio 3 Gen took it as prompt. StepAudio 3 Gen documents its own format for multi-speaker scenes, so we also sent the same scene that way: the setting in instruction, the two speakers in roles, the three lines in scripts, the phone ring in square brackets and the laugh in parentheses. We checked which lines each clip speaks by transcribing it with two speech recognition models, which agreed. Prices are what these exact requests cost at current rates. Times run from sending the request to the finished file; StepAudio 3 Gen answers with a job, timed from creation to completion.

What to listen for

Stage directions. Qwen Audio 3.1 TTS Next turns the description into the scene itself. StepAudio 3 Gen speaks the text it is given, so the setting belongs in instruction and the sound cues in square brackets.

The room. Listen for the jazz under the voices, the phone ring between the lines, and the host's laugh, and for how each clip balances them against the speech.

Speed and price. StepAudio 3 Gen returned its clips in seconds at well under half the price. Qwen Audio 3.1 TTS Next took about a minute and bills the audio it generates, so a longer scene costs more.

Run the same test on EmpirioLabs

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 late-night radio studio, with soft jazz playing in the background. A warm male host says: \"Good evening, and welcome back to Night Owls. Our first caller is on the line.\" A phone rings once. A cheerful female caller says: \"Hi! I have been listening for years, and I finally got through!\" The host laughs and says: \"Well, we are glad you did.\""
  }'

StepAudio 3 Gen, in roles and scripts:

curl https://api.empiriolabs.ai/v1/audio/generations \
  -H "Authorization: Bearer $EMPIRIOLABS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "stepaudio-3-gen",
    "instruction": "A late-night radio studio, with soft jazz playing in the background.",
    "roles": [
      {"name": "Host", "description": "A warm male radio host"},
      {"name": "Caller", "description": "A cheerful female caller"}
    ],
    "scripts": [
      {"speaker": "Host", "text": "Good evening, and welcome back to Night Owls. Our first caller is on the line."},
      {"speaker": "Caller", "text": "[A phone rings once] Hi! I have been listening for years, and I finally got through!"},
      {"speaker": "Host", "text": "(laughs) Well, we are glad you did."}
    ]
  }'

StepAudio 3 Gen answers with a job; read the finished clip from GET /v1/jobs/{job_id}. Try both in the playground.

Frequently asked questions

Were the clips edited or retried?

No. Each clip is the first audio its request returned, unedited. The second StepAudio 3 Gen clip is a different request, the same scene in the model's documented format, not a retry of the first.

Why did StepAudio 3 Gen read the description aloud?

StepAudio 3 Gen speaks the text of prompt as written. Text in square brackets or parentheses is taken as a sound cue or a delivery direction instead of being read out, so the phone ring and the laugh in the second request were not spoken. Put the setting and music in instruction.

Which one should I use?

Qwen Audio 3.1 TTS Next takes a scene written in plain language and can clone up to three voices from reference clips. StepAudio 3 Gen is faster and costs less when the lines and cues are written out in its format. Listen to the clips with your own scene.

Try it

Playground | How to use the Qwen Audio 3.1 TTS Next API | All models | Pricing