From a speech synthesizer to SayFable: the in-house NLP stack has become a library

SayFable started out as a local text-to-speech synthesizer for a single language, but has grown into an offline library featuring NLP annotation, text-to-speech support for 68 languages, and a book-based chat feature. This is more important than just another list of AI features: the developer had to turn a set of separate linguistic tools into a cohesive app that can be used on a phone.

It includes a full stack of features: lemmatization, POS tagging, NER, sentiment analysis, embeddings, and speech synthesis. Initially, everything ran on a Python/Flask stack on the desktop, but it was later ported to a Swift app on iOS.

This transition is the most interesting part of the project. A local prototype can be a convenient way to test an idea, but a personal library requires a different level of integrity: supporting different types of content, working with files and links, and linking NLP analysis with reading and voice synthesis. SayFable integrates books, articles, podcasts, videos, and music into a single, cohesive workflow.

The project was developed over about half a year, working on it in the evenings and on weekends. The app is currently in beta, and it claims to support CarPlay. This is a good example of how an AI application can grow not from a large platform, but from the gradual expansion of a single local tool.

However, it’s too early to view SayFable as a ready-made replacement for major library services: the available description lacks details on the quality of annotation, processing speed, and limitations for specific languages. Therefore, the main achievement at this stage isn’t a promise of perfect NLP, but a working component of an ambitious local stack that has already made its way into a real-world mobile scenario.

Source: All Articles / Machine Learning / Habr