Some notes on voice cloning since everyone seems to be wrong about both how it works and its potential impact on the industry: - Voice cloning has many good op

Some notes on voice cloning since everyone seems to be wrong about both how it works and its potential impact on the industry: - Voice cloning has many good open-source implementations so it's possible for any new company to spin up easily. (the "moat" is having a higher-quality model) - Most implementations of voice cloning only require 10 seconds of clean audio, so they can be obtained quickly. - Implementations like the open-source projects and ElevenLabs do not train a model; instead it creates an encoding which represents a person's voice, and is what is stored/loaded. The encoding is used with a "base model" (trained on an often-secret dataset due to legal liability) for the generations. It also makes it difficult to narrow down which voice is used if a malicious voice is used in the wild. - As with all AI models, it's infeasible to remove a voice from the base model without retraining it. - The audio output of SaaS voice clone models are 44.1kHz, which is CD quality. Voice acting tends to require much higher quality audio for mastering. Open source models are much lower quality. On the ethics of voice cloning: - Modern voice acting tends to be collaborative process between the actor and the audio director/producers to get the best performance. - Current AI-generated voices cannot display contextual and important nuance, such as pauses and tone inflections. These often aren't things that can be added post-production. - Voice actors tend to be contractors, with the clients being the company and "owning" the performance. Voice actors alone don't have legal leverage to get voices taken down, although companies like Rockstar do. - Of course, getting the audio to say embarrassing things could land the voice actor in hot water with the company, especially those less tech-savvy. And, as of last week, some observations from the SAG-AFTRA deal: - Replica Studios notably doesn't have a individual-oriented product, instead requiring a significant cost and is more intended for teams. - Replica Studios apparently doesn't encode the voice and does a full finetune of the model on a new voice.