Alibaba's Qwen3-TTS can clone voices in 3 seconds or design them with text

Customizing an AI voice usually meant settling for generic presets or recording minutes of clean studio audio for training.

Alibaba's Qwen team just rolled out its new Qwen3-TTS lineup, introducing two Flash-tier models: VoiceDesign and VoiceClone.

Custom vocal personas and 3-second cloning

VoiceDesign-VD-Flash scraps preset voices altogether. Instead, it lets you sculpt a unique vocal identity from scratch using free-form text prompts that specify tone, rhythm, emotion, and persona. Alibaba claims the model beats GPT-4o-mini-tts and Gemini-2.5-pro on role-play benchmarks.

Meanwhile, VoiceClone-VC-Flash lowers the clone barrier down to just three seconds of audio. It supports speech generation across 10 languages, including Chinese, English, Japanese, and Spanish, while using context-aware cadence for more natural delivery.

Alibaba claims VoiceClone scored a 15% lower word error rate (WER) in multilingual benchmarks compared to ElevenLabs and GPT-4o-Audio. Both features are now available to test on Qwen Chat.

Why it matters

For creators and indie developers, generating bespoke character voices via prompts and cloning voices in three seconds makes narrative content and localization remarkably cheap. It is well worth trying right now on Qwen Chat.

Strategically, Alibaba is taking direct aim at voice leader ElevenLabs and OpenAI. When frontier AI labs bundle highly expressive, low-latency audio models directly into their platforms, single-feature audio startups will struggle to defend their subscription fees.

We think prompt-based voice design is a smart way to ditch lifeless presets. However, dropping the cloning threshold to just three seconds will inevitably raise red flags around deepfakes and vocal impersonation, demanding tighter safety guardrails from Alibaba.