Tortoise TTS

Tortoise TTS earned its name honestly: it is a text-to-speech system by James Betker that trades speed for quality, producing speech with highly realistic prosody and intonation. Architecturally it chains an autoregressive decoder with a diffusion decoder, an unusual combination at the time of release. It supports multiple built-in voices and can clone a new voice from a few short reference clips, with random voice generation also available. Originally very slow, it has been substantially optimized: with DeepSpeed, KV caching, and half precision enabled it reaches around 0.25 to 0.3 real-time factor on a 4 GB GPU, and a streaming mode brings first-audio latency under half a second. It runs on NVIDIA GPUs, with Apple Silicon supported through nightly PyTorch builds, and installs via pip or Docker with both command-line and Python APIs. Released under the Apache 2.0 license, it remains popular among developers, researchers, and hobbyists producing narration and character voices locally.