Prompt Qwen for text-to-speech

Last updated: October 12, 2026

Qwen speaks your script word for word and takes its direction from a separate field, in a few plain words. Pick it for short character lines, and for giving a character a voice of its own: a voice you clone from one clip or design from a description speaks through Qwen.

Let the agent write it

You don't have to write any of this yourself. The Layer agent knows how to prompt each model it runs, from the same guidance this page is built on. Name the model, the line and the moment, for example: Voice this line with Qwen 3 TTS: a scout whispers a warning to the party, tense and slow. "The bridge is out. Take the river path, stay low, and don't light a torch until you reach the mill." The agent keeps the line clean, puts the delivery in its own field, splits a long script into runs, shows you its plan, and runs it when you approve. Use the rest of this guide when you want to write or adjust the line yourself, and see Write a brief for the agent for how to brief it.

When to use it

Use Qwen for barks, callouts, tutorial prompts and short dialogue lines, one or two sentences at a time. Its strength is the voice: train a voice reference set on Qwen, cloned from a short clip or designed from a written description, and every line for that character comes out in the same voice, in any of ten languages (Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian).

Without a voice reference set, every line is spoken in one default voice, a bright young female voice. For a different character, attach a voice of your own.

Versions on Layer

Qwen 3 TTS

The 1.7B Qwen speech model. It speaks the script in the default voice, or in a voice reference set trained on it, and takes Style instructions for the delivery when no trained voice is attached.

How to write the prompt

The prompt is only the words the character says. The delivery goes in Style instructions:

Prompt: The bridge is out. Take the river path, stay low, and don't light a torch until you reach the mill.
Style instructions: whispering, tense, speaking slowly

Qwen reads every word of the prompt aloud, so anything that is not dialogue belongs in the style field. Punctuation sets the rhythm: each full stop and comma becomes a pause. That line and style are the ones in the clip above.

What works

The delivery in its own field

Write the direction in Style instructions, in a few words of tone, emotion, pace and loudness. We voiced the same line three ways. With whispering, tense, speaking slowly it ran 8.1 seconds, with a clear pause at each comma, and came out noticeably quieter than the line with no style (7.2 seconds). With urgent, shouting, speaking fast it ran 6.3 seconds. Every take said every word. We measured length and loudness rather than listening, one take each.

One or two sentences per run

Each run speaks about 16 seconds of audio. When we gave Qwen a 75-word welcome speech, it stopped mid-word after 16 seconds and the last two sentences were never spoken. Keep each run to a sentence or two, about 35 words, and voice longer narration as several runs, one per line or beat. It also makes a retake cheaper: a bad word costs one line, not the whole speech.

Punctuation for the pauses

Write the script with the breaks you want heard. Full stops give the longer pauses, commas shorter ones.

Acoustic words, not story

Describe how the line should sound, not why: calm, warm, unhurried rather than she has forgiven him. Keep the words consistent with each other; shouting quietly gives the model two directions to choose between.

A voice of your own

To cast a character, train a voice reference set on Qwen. Clone it from one clip of the voice, or design it from a Voice description. For a designed voice, describe how it sounds: age and gender, pitch, texture, emotion, pace and accent, in two to four sentences, and name the language for an accent ("a thick French accent speaking English"). A cloned voice keeps its sound in every language Qwen speaks.

What to avoid

Directions in the prompt

Anything in the prompt is spoken. When we wrote "Say this in a tense whisper:" before the line, the take began with those exact words. Move the direction to Style instructions.

Inline tags

Qwen has no tag syntax for sounds or pauses. In our test, [sighs] at the start of the line was not read as a word, but it slowed the whole line and left an unexplained sound before it. Leave tags out and use the style field and punctuation instead.

Style instructions with a trained voice

When a voice reference set is attached, Qwen ignores Style instructions: the voice clip sets the delivery. For a whispered or shouted line in a cloned voice, clone from a clip recorded in that delivery.

Long scripts in one run

A script longer than about 16 seconds of speech is cut off mid-word. Split the script before you run it.

Settings that matter

Style instructions holds the delivery, in a few words; it applies to the default voice only. Reference set attaches a voice you trained on Qwen. Temperature sets how much the delivery varies from take to take: lower it for a predictable read, keep it high to audition different performances. Speech language translates your script into the language you choose before it is spoken; leave it empty to speak the script as written. Style instructions, Temperature and Speech language sit under Advanced.

Prompts to start from

Whispered warning

Prompt: The bridge is out. Take the river path, stay low, and don't light a torch until you reach the mill.
Style instructions: whispering, tense, speaking slowly

Battlefield call

Prompt: The bridge is out. Take the river path, stay low, and don't light a torch until you reach the mill.
Style instructions: urgent, shouting, speaking fast

Use it in Layer

  1. In a session, open Create speech from the side panel.

  2. Click the model name at the top of the panel and choose Qwen 3 TTS.

  3. In the prompt box, write only the words the character says, a sentence or two.

  4. Open Advanced and write the delivery in Style instructions, in a few words.

  5. Click Create speech. The cost shows on the button before you click.

  6. When the line lands on the board, click View to play it. Change one word of the style at a time and run it again.

Learn more

  • Qwen3-TTS (Qwen): the preset voices and their native languages, natural-language instructions for tone and emotion, voice design, and three-second voice cloning.
  • Qwen3-TTS-12Hz-1.7B-CustomVoice (Qwen): how instructions steer tone, speaking rate and emotion on a preset voice, and why each voice sounds best in its own language.
  • Qwen3-TTS technical report (Qwen): how the model follows voice descriptions, how a cloned voice carries across all ten languages, and long-form speech.
  • Qwen 3 TTS [1.7B] API (fal): the hosted endpoint Layer runs, with a preset voice or a cloned speaker embedding, and its style prompt.
  • How to design AI voices in minutes using Qwen3-TTS (Stream): a worked framework for describing a voice when you design one, and contradictions to avoid.

Related