An illustration of three standard audio synthesis pipelines. Top (text-to-speech): this pipeline generates speech from text. It begins with text input, which is converted into a phonetic representation by a linguistic frontend. An acoustic model generates acoustic features from the phonemes, and a vocoder synthesizes the final waveform output. Middle (voice conversion): this pipeline transforms the voice of a source speaker into that of a target speaker. Source speech is processed to extract its linguistic content (content encoder) and pitch (F0 extractor). A conversion model combines this information with a target speaker embedding to create new acoustic features. A vocoder then synthesizes these features into the converted speech target voice. Bottom (voice cloning): this pipeline synthesizes a specific person’s voice from a text prompt. A text prompt and a speaker representation from a speaker encoder are fed into a pre-trained acoustic model. A vocoder then generates the final cloned voice speech