wlejon/kokoro-lab-py

kokoro-lab-py is an interactive tool for experimenting with a text to speech model called Kokoro, which turns written text into spoken audio. Instead of just typing text and getting one fixed voice back, this lab lets a user design a custom voice by moving sliders that control things like pitch, brightness, and speaking pace, or by starting from an existing stock voice or generating a random one. Beyond picking a voice, a user can nudge how masculine or feminine the voice sounds, and steer its emotional quality along three dials often used to describe emotion: how positive or negative it feels, how energetic it is, and how assertive it sounds. For finer control, a person can draw directly on graphs showing the pitch, loudness, and timing of the speech, and the audio re-synthesizes immediately to match, without needing to press any kind of render button. Any change made this way is remembered as an adjustment layered on top of the voice, so switching the base voice, emotion, or speed afterward keeps that manual reshaping instead of losing it. Under the surface, the tool is built in two layers. The part that lets sliders control the voice is plain mathematics with no AI model involved, working over a small set of downloaded reference files that describe the space of possible voices. The part that actually produces audio runs the Kokoro model step by step rather than all at once, which lets the tool capture every stage of the process and, importantly, redo only the back half of that process when someone edits the pitch or timing contours, making those edits fast instead of re-running the whole model each time. Getting started requires Python between versions 3.10 and 3.12, since a dependency the speech model relies on has no support yet for newer Python releases, plus downloading a folder of voice-space reference data from Hugging Face. The actual Kokoro model weights download automatically the first time the app runs. Once set up, the tool opens as a small local web app in a browser. It carries an MIT license, and it is aimed at developers and hobbyists interested in experimenting with speech synthesis and voice design rather than end users looking for a simple text to speech app.