mu-zi-lee/qwen3-tts-skill
Overview
This skill provides a complete local TTS workflow around Qwen3-TTS for high-quality single-sentence synthesis and long-form batch dubbing. It supports three synthesis modes: CustomVoice, VoiceDesign, and VoiceClone, plus a tokenizer tool for audio roundtrips. The skill includes scripts for single-shot generation, batch processing of article→dubbing JSON→TTS→merge, and a local web UI demo.
How this skill works
For single sentences, run the provided script with the chosen mode, language, text, and optional speaker or instruct flags to produce WAV output. For long-form content, the workflow converts an article into a dubbing JSON (segmenting text, labelling roles, and suggesting emotional instructions), you review and adjust the JSON, then run the batch generator which produces segment WAVs and merges them into a final audio file using FFmpeg. Models can be used from local downloads or hubs and the scripts accept device, dtype, and attention settings for performance tuning.
When to use it
- Generate high-quality narration or voice lines for short text snippets.
- Produce audiobook or long-article narration via batch dubbing.
- Create multi-character dialogues with role labels and emotion control.
- Clone a specific speaker voice from reference audio for personalization.
- Quickly iterate on custom voice designs described in natural language.
Best practices
- Use CustomVoice for most tasks; use VoiceDesign when you need a precise timbre described in natural language.
- For VoiceClone provide a clear reference audio and matching transcript for best results.
- Review and adjust the generated dubbing JSON before batch TTS to ensure correct segmentation, roles, and emotional instructions.
- Install FFmpeg for reliable merging of segment files and verify GPU/device settings for optimal performance.
- Download required models locally in advance to avoid slow network downloads during runs.
Example use cases
- Record a 30–60 second promo voiceover using CustomVoice with an instruct for tone.
- Convert a 10,000-word article into a single merged WAV via the batch dubbing pipeline.
- Produce a multi-role audio drama by labelling characters in the dubbing JSON and setting different speakers per role.
- Create a personalized greeting by cloning a short reference sample and synthesizing new lines.
FAQ
Use CustomVoice for reliable built-in speakers with emotion control; use VoiceDesign when you need a specially described timbre.
How do I get a single merged audio from a long article?
Generate the dubbing JSON, review and edit it, then run the batch generator which creates segment files and merges them into a final WAV using FFmpeg.
What do I need to run large batches efficiently?
Install FFmpeg, ensure sufficient GPU memory or set dtype/device options, and consider FlashAttention 2 if supported to reduce memory use.
