Qwen3-TTS family is now open sourced: Voice design, clone, and generation

Perceived Voice Style and Quality

• Many listeners feel the English demos sound like anime dubs, YouTube personalities, or tween-drama podcasts—highly “performed” and sometimes exaggerated.
• Some note the Japanese samples are also anime-like, and at least one Japanese line is mispronounced, leading to skepticism about Japanese quality.
• Others point out the prompts explicitly encourage that style, and that more “normal” voices lower on the page sound fine.
• The Obama and other celebrity-style clones are described as distinct and impressively close; several say many voices clone better than 11labs, though at a lower bitrate.
• One user found cloning captured vocal tone well but not natural intonation, resulting in a somewhat flat, monotonous delivery—possibly due to using a base model or missing expressiveness controls.
• Another reports emotional instability with the 0.6B model (unwanted laughter/moans between chunks); suggestions include more detailed style/emotion prompts.

Use Cases and Creative Potential

• Strong enthusiasm for:
  • Audiobooks (where other TTS still struggle).
  • Restoring / remastering old radio plays and damaged recordings where words can be inferred from context.
  • Indie games and projects, including accent correction of non-native voice actors.
  • Personalized content: podcasts, using one’s own voice, or deceased relatives reading children’s books.
  • Dubbing movies into other languages while retaining something like the “original voice”.

Safety, Scams, and Societal Impact

• Multiple comments call the tech “terrifying” and see it as crossing a major threshold: realistic voice and image deepfakes are now accessible to almost anyone.
• Concrete fears: family/emergency scams using cloned faces and voices; erosion of trust in digital evidence and greater plausible deniability (“AI made that”).
• Mitigations discussed:
  • Pre-agreed “secret words” within families.
  • Cryptographic provenance systems like C2PA.
  • Web3/NFT ideas are mentioned but questioned as to how they’d distinguish human vs AI-assisted content.
• Others argue the benefits (democratized creativity, new forms of games, film, and music) will outweigh downsides over time, though several note a painful transition period and impacts on creative livelihoods.

Running Locally and Performance

• 0.6B model runs on older GPUs like a GTX 1080 but often slower than real time; 1.7B uses ~6GB VRAM and is more robust to background noise.
• Lack of FlashAttention significantly slows inference; some users run fine without it, others are blocked by install issues.
• Mac support initially unclear, later confirmed via MLX-based tooling; CPU-only is reported as possible but slow, and edge-device viability remains uncertain.
• Hugging Face demos exist but can be overloaded; local CLI/frontends and example scripts are shared in the thread.