Changelog

Move punctuations removal to SpeechTokenizer directly

Improve whitespace and break handling in the SpeechTokenizer

Transform URLs into spoken form in the SpeechTokenizer

Support floating point number parsing in SpeechDurationTokenizer

New SpeechDurationTokenizer consistently outperforming the old tokenizer by 400%+ on Markdown and 100%+ on plain text.

Allow TTS pace to be updated mid-job and adjust accordingly

Add trailing and leading whitespace supports to SpeechTokenizer to support incremental tokenization

Reduce sidebar width to 33 in life dev

New SpeechTokenizer to pre-process LLM responses for TTS, with full support for CommonMark, GFM, LaTeX, MDX, and Mermaid Markdown flavors/syntaxes

Solve SpeechDurationTokenizer.take() returning the entire text for 0 or negative indexes

Refactor markdown utilities and add mdast extensions to parse Life.js syntaxes

Persist ONNX state of Silero VAD jobs

Support Sonic 3 model from Cartesia

Models API refactoring, renamed types, simplified job structure, use job in VAD

Upgrade to Silero VAD 6.2 and simplify the class implementation

New <Markdown/> component from life/react to render Markdown streams

Allow createLifeClient to be provided with no options object

Resolve incorrect child process path in server

Ensure env vars are injected BEFORE the CLI code gets loaded

Remove the need for the patched version of ink + bump deps

Include the patched version of ink in the life package build

Upgrade to Vitest 4.0