Spear TTS

Spear TTS (Speak, Read and Prompt) is a multi-speaker text-to-speech system from Google Research, published on arXiv in February 2023. It casts TTS as two sequence-to-sequence tasks: text to discrete semantic tokens (reading) and semantic tokens to acoustic tokens (speaking), which are decoded by a SoundStream decoder.

Decoupling the two stages lets the speaking module train on abundant audio-only data, so the full system needs only about 15 minutes of parallel data. Speaker identity is controlled by example prompting: a 3-second sample is enough to generalize to an unseen voice with no speaker labels.

The paper's experiments matched ground-truth speech on naturalness and acoustic quality in subjective tests. Google did not release official weights, so this entry is catalogued as a research system with closed weights; community reimplementations exist in PyTorch.