Tensor TownISSUE #95 of 120

text-to-speech · vits · xtts · elevenlabs

NeuraVSThe Overfit Ogre

Neura saysText-to-speech (VITS, XTTS, ElevenLabs) generates natural, even cloned, voices from text.

Modern TTS produces speech nearly indistinguishable from human, with control over voice, emotion, and pacing — and can clone a voice from seconds of audio. Models like VITS (open), XTTS, and services like ElevenLabs power audiobooks, assistants, dubbing, and accessibility. Voice cloning raises real consent and fraud concerns (deepfake calls), so responsible use and detection matter.

Power-ups you unlock

Near-human synthetic speech
Control voice, emotion, pacing
Voice cloning from seconds of audio
Consent/fraud concerns (deepfakes)

The Overfit Ogre attacks — common mistakes

Cloning voices without consent
Assuming synthetic speech is detectable
Ignoring misuse/fraud potential

Boss battleName a beneficial and a harmful use of voice cloning.

Example code

<!doctype html><html><head><meta charset="utf-8"></head>
<body style="background:#06040d;color:#e6e0ff;font-family:monospace;padding:20px"><pre>text → natural speech (voice, emotion, pace)
clone a voice from ~seconds of audio
good: accessibility · harmful: deepfake fraud</pre></body></html>

▶ Open the interactive comic issue

‹ Controlnet · Guiding Generation Speech Recognition · Whisper ›