Optimum Intel is a Hugging Face integration layer that connects Transformers, Diffusers, Sentence Transformers, and timm with OpenVINO to optimize and accelerate model pipelines on Intel hardware. It supports exporting models to OpenVINO IR, running inference with OpenVINO Runtime, and applying quantization/optimization techniques such as INT8/INT4 compression and static quantization.

audio.cpp is a high-performance native C++ audio inference framework built on ggml for running local audio models efficiently across tasks like TTS, ASR, voice cloning, voice conversion, diarization, VAD, source separation, music generation, and alignment. This repository is a downstream distribution that adds a Windows-friendly WebUI, local launcher scripts, an API server, and workflow/pipeline support on top of the upstream core runtime.

Vox Director is an agent skill that turns a single topic, talking-head video, or photo into a finished Vox-style paper-collage explainer or ad video. It automates the full pipeline: beat mapping, style selection, collage keyframe generation, motion animation, narration, music, captions, and final ffmpeg assembly.

Nanospeech is a research-oriented, hackable text-to-speech system built in PyTorch and MLX. It provides a small pretrained English TTS model, voice matching from reference audio, built-in sample voices, command-line generation, a Gradio demo, and training scripts for both the main model and a duration predictor.

AutoShorts is a GPU-accelerated pipeline for turning long gameplay recordings into vertical short-form clips for platforms like TikTok, YouTube Shorts, and Reels. It uses AI scene analysis to find highlights, then automatically crops, renders, adds subtitles, and can generate multilingual AI voiceovers with style-adaptive narration.

This repository provides a browser-based AI audiobook/story production tool that uses an LLM to split scripts, detect emotions, and automatically arrange multi-character TTS narration, sound effects, background music, and audio filters. It also supports exporting finished WAV audio, SRT subtitles, and MP4 videos with generated background-image prompts, while allowing users to manage prompts, voice resources, and project files locally in the browser.

This repository provides a one-line embeddable AI virtual human widget for websites, featuring Live2D/VRM characters, browser-based speech input/output, lip-sync, emotion expressions, and optional local LLM or knowledge-base answering. It is designed as a mostly frontend solution with iframe-based embedding, configurable character/knowledge/personality settings, and optional serverless neural TTS support.

This repository is a PyTorch implementation of Google's Parallel Tacotron 2, a non-autoregressive text-to-speech model with differentiable duration modeling. It includes code for data preparation, preprocessing, training, evaluation, and speech synthesis, along with a bundled HiFi-GAN vocoder for waveform generation.

Wyoming OpenAI is a Python proxy middleware that bridges the Wyoming protocol with OpenAI-compatible APIs for speech-to-text and text-to-speech. It lets clients like Home Assistant use a variety of local or cloud backends (OpenAI, Speaches, LocalAI, Kokoro, Mistral Voxtral, Edge TTS, Groq, Chatterbox) and supports both standard and streaming/realtime transcription plus incremental TTS streaming with sentence chunking.

OrkasVideoStudio is a coding-agent-driven video studio for composing, editing, generating, and automatically assembling videos from a readable, diffable `plan.json` timeline. It provides a CLI (`ovs`) and MCP server so agents like Claude Code or Codex can create, modify, and verify videos using deterministic tools for rendering, ffmpeg-based editing, transcription, and BYO-key generation providers.