Qwen3 TTS Prompt Guide: From Voice Descriptions to Emotion Tags

Instructions vs. Inline Tags: Know the Difference

Qwen speech models offer two distinct ways to shape an output: voice instructions and inline performance tags. They solve different problems, live in different fields, and are supported by different model families.

The simplest mental model is:

Instructions define the speaker. Tags direct the performance.

Do not assume every Qwen speech model supports both. If an instruction has no effect or a tag is spoken aloud, confirm the selected model before rewriting the prompt.

---

How to Write Effective Qwen3 TTS Voice Instructions

Voice instructions describe the sound you want in natural language. Instead of tuning pitch curves or prosody parameters, you describe perceptible qualities such as age, pitch, speaking rate, emotional baseline, and vocal texture.

Three practical constraints matter:

• Use a Qwen3-TTS-Instruct-Flash model when you need instruction control.
• Write instructions in English or Chinese and avoid mixing unrelated languages.
• Keep the instruction below the 1,600-token limit.

The limit is generous, but shorter prompts are usually easier to control. A focused description of 30–80 words gives the model enough detail without introducing contradictions.

Here is an illustrative request:

response = client.audio.speech.create(
    model="qwen3-tts-instruct-flash",
    instructions=(
        "A composed middle-aged man with a deep, resonant timbre. "
        "Speak slowly and clearly for documentary narration."
    ),
    text="This forest has been growing here for three thousand years.",
)

Treat the example as a prompt pattern; use the exact request fields required by your SDK or deployment.

Where voice instructions work well

• Audiobooks and audio drama: keep a narrator or character voice consistent across many scenes.
• Advertising: test the same script with several personas before choosing a final direction.
• Games and animation: create recognizable voices for NPCs and recurring characters.
• Voice assistants: encode a stable brand personality into the baseline delivery.
• Documentary and news narration: specify steady pacing, clear diction, and restrained emotion.

---

Five Principles for Better Voice Descriptions

1. Be specific, not vague

Words such as “nice,” “normal,” or “professional” leave too much room for interpretation. Describe audible properties instead:

• Weak: A nice professional voice.
• Better: A composed woman with a mid-low pitch, crisp diction, and a steady medium pace.

2. Use several dimensions

Female voice describes only one dimension. Add age, pitch, pace, emotion, texture, and purpose to reduce variation between generations.

3. Prefer objective sound qualities

Write what a listener could hear. Higher pitch with energetic sentence endings is more actionable than a voice everyone will love.

4. Describe qualities instead of imitating a person

Avoid asking for a celebrity, actor, presenter, or other identifiable person. Describe the acoustic qualities you want: warm, gravelly, bright, restrained, theatrical, or conversational.

5. Remove redundant adjectives

Stacking near-synonyms does not multiply an effect. Gentle, soft, tender, mellow, soothing spends many words on one dimension while leaving pace, pitch, and purpose undefined.

---

The Seven-Dimension Voice Prompt Framework

When you are unsure what to write, build the instruction from seven dimensions.

Use this reusable formula:

Age + gender + pitch or timbre + speaking rate + emotional baseline + purpose

Copy-and-paste voice instruction templates

Audiobook narrator

Middle-aged male, around 40, with a mid-low pitch and a warm,
slightly magnetic timbre. Use a slow, steady pace and restrained
emotion for long-form novel narration.

E-commerce product video

Young woman, around 25, with a higher pitch and crisp, cutting-through
timbre. Speak slightly fast with warm, persuasive energy and gently
rising sentence endings.

Premium brand film

Middle-aged woman with a rounded low timbre. Use a slow pace,
generous pauses, and a restrained but assured delivery for a
high-end brand film.

Game antagonist

Middle-aged man with a low pitch and raspy, gravelly timbre.
Speak slowly with a cold, menacing emotional baseline for a
game antagonist.

Everyday voice assistant

Young adult with a neutral-leaning female voice, mid pitch, and a
clean, clear timbre. Use a medium pace that feels friendly, natural,
and unforced for everyday questions and answers.

Science documentary

Male, age 35 to 45, with a mid-low pitch and precise articulation.
Use a medium-slow pace, rational tone, and a subtle sense of
curiosity for science documentary narration.

---

Emotion Tags and Rich-Language Tags

The qwen-audio-3.0-tts-plus and qwen-audio-3.0-tts-flash models can interpret supported tags placed directly inside the text. There are two categories, and their scopes are different.

Emotion and delivery tags: span scope

A control tag affects the words that follow it until another control tag appears or the text is segmented.

Example:

[serious]Please pay attention to the safety instructions.
[excited]All right, let's begin!

The second tag replaces the first emotional direction for the text that follows.

Rich-language tags: point scope

Rich-language tags insert a vocal event at one position without changing the surrounding emotional baseline.

You can place a point event inside a tagged emotional span:

[excited]The weather is fantastic today!
[laughing]Let's go out together!

[excited] controls the ongoing delivery. [laughing] inserts laughter at that point, after which the excited delivery continues.

Radio-drama example

[tired][sighing]Another day of working until midnight.
[curious]Hm—who left this letter on my desk?
[gasp][panicked]This... this cannot be happening!

Use punctuation as part of the direction. Dashes, ellipses, commas, and sentence boundaries help the model infer timing without adding a tag to every phrase.

---

Can You Combine Instructions and Tags?

Think of instructions and tags as complementary concepts, but check model support before trying to use them in one request.

• Qwen3-TTS-Instruct-Flash uses the instructions field for speaker-level control.
• qwen-audio-3.0-tts-plus and qwen-audio-3.0-tts-flash use inline tags for performance control.

Based on those support boundaries, do not assume a single model accepts both mechanisms. Choose the control system that matches your task:

• Need a stable, custom speaker persona? Use instruction control.
• Need emotion changes and vocal events inside a script? Use a tag-capable Qwen Audio TTS model.

For a larger production, keep a reusable persona brief alongside the script even when the selected model only accepts tags. That brief remains useful for casting a base voice, reviewing consistency, and moving the project between compatible models.

Three rules for natural tagged speech

1. Do not over-tag. One control tag per sentence or emotional beat is usually enough.
2. Place point events logically. A sigh normally belongs before or after a thought, not in the middle of an unrelated phrase.
3. Restate the mood in long scripts. Because control tags persist, repeat the intended tag at paragraph boundaries when consistency matters.

---

Troubleshooting Qwen TTS Prompts

---

A Repeatable Prompt-Tuning Workflow

1. Write a baseline voice instruction. Use the seven-dimension formula and keep it under roughly 40 words for the first test.
2. Use one fixed test passage. Include a statement, a question, a number, and a longer sentence so you can compare pacing and articulation.
3. Generate several samples. Listen for changes in timbre, pace, and emotional baseline.
4. Change one dimension at a time. Replace medium pace with slightly slow without rewriting everything else.
5. Choose the correct model for performance control. If the script needs mid-scene emotion changes, move to a tag-capable model.
6. Add tags only at key beats. Start with one tag per sentence, then remove any tag that does not improve the result.
7. Save the validated pattern. Store the final persona brief, model name, test script, and tag conventions as a reusable project template.

A compact test passage

Welcome to the demonstration. Today we will review three practical ideas.
Why do they matter? Because a clear voice makes complex information easier
to understand. Take a short pause, then finish with calm confidence.

This passage exposes pace, question intonation, number pronunciation, transitions, and the ending tone in a small amount of text.

---

Frequently Asked Questions

What makes a good Qwen3 TTS prompt?

A good prompt describes audible qualities across several dimensions: age, gender, pitch, rate, emotional baseline, vocal texture, and purpose. It stays concise and avoids contradictory adjectives.

How long should a voice instruction be?

The supported limit is 1,600 tokens, but a focused 30–80-word instruction is usually easier to tune. Start shorter and add only the dimensions the output is missing.

Why is Qwen TTS reading my emotion tags aloud?

The selected model probably does not support inline emotion and rich-language tags. Use qwen-audio-3.0-tts-plus or qwen-audio-3.0-tts-flash for the supported tags listed in this guide.

Should I add an emotion tag before every sentence?

No. Add a new control tag only when the intended delivery changes or when a long passage needs the mood restated. Excessive tags can make prosody sound fragmented.

Can I ask Qwen3 TTS to sound like a celebrity?

Describe vocal qualities rather than requesting imitation of an identifiable person. This produces a clearer prompt and avoids unnecessary personality-rights and copyright concerns.

---

Final Prompt Checklist

Before generating, confirm that:

• The selected model supports the control mechanism you are using.
• The voice instruction describes several audible dimensions.
• Every adjective contributes new information.
• No two directions contradict each other.
• Emotion tags appear before the text they should control.
• Vocal-event tags sit at natural moments in the sentence.
• Long scripts restate important control tags at paragraph boundaries.
• You changed only one variable since the previous test.

Qwen speech prompting is less about writing a dramatic paragraph and more about giving the model a clean production brief. Define the stable voice with specific, objective language; direct changing emotions only where the model supports inline tags; and tune one dimension at a time. That process is faster, more repeatable, and far easier to scale across a full project.