AI Voice over Text to Speech is WAY TOO GOOD - Overdub AI by Descript

AI voice cloning and text-to-speech are getting dramatically more usable, and Descript’s “Overdub” is positioned as a practical way to generate realistic voice variations from a short sample—then immediately use that voice to read new text. The core workflow is straightforward: upload a recording of someone’s speech (in this case, the user’s own voice), wait for processing (about a day), and then type any script to have the system produce speech in that cloned voice. Descript also ties this capability to a broader video editing and transcription toolset, making voice generation part of a production pipeline rather than a standalone experiment.

The transcript highlights two major capabilities: (1) generating speech from stock voices and (2) converting a submitted voice into an “Overdub” voice that can read arbitrary text. Early tests include a simple question—“is mayonnaise an instrument”—followed by longer, recognizable scripts. When the system uses stock voices (including an announcer-style option), the output lands as high-quality, trailer-like narration. The speaker notes that repeated words can sometimes reveal the synthetic nature, but many lines can pass as human speech, especially when delivered in a confident, announcer cadence.

The more striking demonstration comes from using the user’s own voice. After uploading a video sample recorded with a specific microphone, the system produces speech that closely matches the speaker’s tone and delivery. The transcript describes the results as “pretty darn good,” with occasional differences: the cloned voice can sound more monotone or slightly robotic depending on the text and how the voice was trained. Still, the system handles long passages well enough to deliver comedic effect when reading the “B-movie” script, including the rapid, rhythmic phrasing and character-like lines.

To test flexibility, the transcript pushes beyond plain narration into performance-style text. A rap excerpt from Eminem’s “Rap God” is used as a stress test, with playback sped up to match the cadence. The output is presented as impressive enough to sound like the speaker rapping along, even though the system is limited by how much text can be processed in one batch. Finally, the transcript uses a Wikipedia description of “Bikini Bottom” from SpongeBob SquarePants to evaluate pronunciation and handling of complex vocabulary. Most of the difficult terms are rendered correctly, and the voice remains consistent while reading unfamiliar words.

Alongside Descript’s Overdub, the transcript briefly surveys Meta/Facebook research on long-form text-to-video generation—described as time-agnostic VQGAN and time-sensitive transformers—citing examples like character reactions to typed prompts, short 8 fps clips, and longer sequences such as sunsets, clouds, and tai chi practice. The takeaway is that both audio and video generation are moving toward controllable, production-ready outputs: voice cloning becomes a practical editing tool, while text-to-video systems show increasing realism and length. The transcript frames this as a meaningful step for creators who want faster iteration—writing text, generating voice, and producing finished narration without hiring voice talent, provided they have legal rights to the voice being replicated.