1. Choose lesson or paste text
Math and Python are deliberately not read aloud character by character. Read the equations on screen alongside narration.
2. Natural voice controls
Ready. First neural generation may download tens or hundreds of megabytes of model files and take time, especially on CPU.
Powered by Kokoro-82M and kokoro-js under open licenses. Experimental browser integration: some browsers, devices and CDN settings may not support it. It synthesizes short segments sequentially rather than an entire textbook chapter at once. System speech is a fallback, not the preferred quality.
3. Practical guidance
Why did the old voice sound bad? The pages used speechSynthesis, which selects a built-in operating-system voice. Changing the MP3 or WAV encoder cannot improve that voice's underlying prosody. A neural speech model synthesizes different audio; its quality depends on the voice and model, not just the file format.
Neural audio is generated on your own computer, so there are no account-level API costs. You can select a short chunk of text and generate it, then repeat with a longer section once you know the performance on your device. For a production course, pre-generated chapter audio could be added as compressed downloadable files after consent and licensing checks.
References: kokoro-js documentation · Model card