Skip to content

Design a Voice Clip with Qwen3-TTS

Make a reference clip for a character without recording an actor. Qwen3-TTS VoiceDesign creates a voice from a text description, such as "a gruff man in his fifties with a low, gravelly voice". You render a short line in that voice, save it as a WAV, and the agent's voice profile clones it at runtime.

VoiceDesign runs outside Tryll, while you author content. Tryll does not use it to speak at runtime: a description-based model samples a new voice on every call, and Tryll synthesizes a reply sentence by sentence, so the voice would change in the middle of a reply. A clip fixes the voice once, and the cloning model then keeps it the same in every sentence.

Prerequisites

  • A cloning TTS model for runtime: Qwen3-TTS 12Hz 1.7B Base (Q4_K_M) or Pocket TTS (int8). See Clone a Voice.
  • A web browser for the hosted demo, or Python if you run VoiceDesign on your own machine.

1. Open VoiceDesign

  • Hosted demo. Open the Qwen3-TTS demo on Hugging Face and stay on the Voice Design tab. It has three inputs: Text to Synthesize (the words of the clip), Language, and Voice Description (who is speaking and how).
  • On your own machine. Install Qwen's qwen-tts Python package and run qwen-tts-demo Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign. It opens the same interface locally. See the Qwen3-TTS repository for setup.

The other two tabs are not needed here. Voice Clone (Base) is the same cloning that Tryll runs. TTS (CustomVoice) speaks with Qwen's preset speakers.

2. Write the voice description

The description decides the voice. Vague words give a generic voice, and a different one each time. These rules come from Alibaba Cloud's Voice Design guide:

  • Be specific. Use words for qualities you can hear, such as deep, husky, crisp, nasal, fast-paced. Avoid nice, normal, good.

  • Combine several dimensions. "Female voice" alone says almost nothing. Cover most of these:

    Dimension Examples
    Gender male, female
    Age child, teenager, young adult, middle-aged, elderly, or an age such as "in her sixties"
    Pitch high, medium, low; "slightly low"
    Speed fast, medium, slow
    Timbre resonant, husky, gravelly, breathy, crisp, mellow
    Accent Scottish accent, light Eastern European accent
    Emotion and delivery calm, cheerful, irritated, fearful, whispering
    Role tavern keeper, army sergeant, news anchor, storyteller

  • Describe the sound, not your taste. "Warm, slow, slightly breathy" works. "A voice players will love" does not.

  • Do not ask for a real person. Describe the qualities instead of naming an actor or a celebrity.
  • Keep it short. One or two sentences. Drop repeated synonyms and filler.
  • Write it in English or Chinese. The text to synthesize can be in any of the model's languages.

Some examples:

Too vague Better
"Old man." "A man in his seventies with a low, gravelly voice, speaking slowly and kindly, like a village storyteller."
"Female warrior, cool voice." "A woman in her thirties with a firm, slightly low voice and a light Scottish accent, speaking briskly with calm authority."
"A merchant." "A cheerful middle-aged man with a bright, slightly nasal voice, speaking fast and eagerly, like a market trader."

For a character who will need several moods, write the description in two parts: the identity (gender, age, pitch, timbre, accent) and the delivery (the emotion). Keep the identity part word for word, and change only the delivery for each mood. See Add Emotions to a Voice.

3. Write the text to synthesize

  • The model speaks the text word for word. Put stage directions in the description, never in the text.
  • Write a line the character would say, in the mood you want. The words help the delivery: an angry clip is easier to get from an angry line.
  • Use the language the character speaks in the game. Set Language to match, or leave it on Auto. Pocket TTS speaks English only.
  • Match the clip length to the runtime model. Qwen3-TTS Base needs 3–10 seconds. Pocket TTS works best with 10–20 seconds, so write two or three sentences. Render the whole clip in one go: takes joined from separate renders can sound like different people.

4. Render several takes, keep the best

The same description gives a slightly different voice on every render. Render a few takes, and keep the one that fits the character best. Listen for:

  • a voice that stays the same from the first word to the last;
  • clean audio with no clicks, warbling or cut-off words;
  • speech that starts right away, without a long silence.

If no take fits, change the description, not just the seed. Add the dimension that is off: the age, the pitch, or the speed.

5. Save the clip and use it

  1. Download the take with the download button on the Generated Audio player.
  2. Convert it to mono 16-bit PCM WAV (24 kHz is ideal) and trim any silence at the start. The Audacity and ffmpeg steps are in Clone a Voice → Record a reference clip.
  3. Put it in the storage folder, for example <storage root>/voices/innkeeper/neutral.wav.
  4. Point the voice profile's Reference Voice at it, as in Clone a Voice → Point the agent at the clip.

Licence and rights

Check the licence of the Qwen3-TTS model and the terms of the service you render with before you ship a generated clip in a product. Do not describe a real person's voice.