Dialoguing with Hugging Face's LLMs

We are currently developing an experimental project: speech-to-speech. This library, still under development, offers a modular approach to speech-to-speech language models through a pipeline that includes speech-to-text, large language model prompting, and text-to-speech. The pipeline is fully open and modular, with a focus on leveraging models available via the Transformers library on the Hugging Face hub. While current approaches tend to focus on end-to-end models, often citing latency benefits, we have achieved a latency of around 500ms with a few technical tweaks. Moreover, the modular approach allows users to easily switch components: for example, swapping the Gemma 2B model for LLaMA 3.1 8B can be done directly from the CLI.