Image · Video · Audio

OmniVoice Studio bundles voice cloning, dubbing, dictation, and TTS into a desktop app that keeps all audio processing off the internet and away from API keys.

SGLang exists to push low-latency, high-throughput inference for LLMs and multimodal models from a single GPU up to massive clusters.

An Electron app that wraps 200+ generative models behind a single UI, with an unusual pitch: no guardrails, no cloud lock-in, and a split personality between local and remote inference.

Deezer open-sourced its TensorFlow stem splitter so developers can pull vocals, drums, bass, and piano out of a mixed track without training a model from scratch.

To give developers a single, general-purpose speech model that handles transcription, translation, and language identification by treating tasks as tokens to predict.

Voicebox is a local-first alternative to ElevenLabs and WisprFlow that clones voices, dictates anywhere, and gives AI agents a mouthpiece — all without shipping audio to the cloud.

VibeVoice is a family of open-source speech models from Microsoft built to ingest, transcribe, and generate very long audio sessions—up to an hour—in a single pass.

LocalAI wraps 36+ inference engines behind one OpenAI-compatible API and pulls them on demand, so you can run LLMs, vision, voice, and video on anything from a CPU to a Jetson.

It turns face swaps and lip-syncs into queued, retryable batch jobs instead of one-off scripts.

Fish Speech S2 Pro is a 4B-parameter dual-autoregressive TTS model that treats inline emotional tags like `[whisper]` as native tokens to synthesize speech in over 80 languages.

A reimplementation of OpenAI's Whisper that trades the original inference engine for CTranslate2 and gains up to 4× speed without sacrificing accuracy.

Because 'make it pretty' is not a prompt, and this repo treats image generation like a production pipeline, not a toy.

To squeeze multimodal understanding—vision, video, and even real-time speech—into models small enough to run natively on a handset.

A Chinese speech toolkit that bundles ASR, diarization, emotion detection, and streaming into one MIT-licensed package.

Wan2.2 is an open video generation suite that treats denoising timesteps like specialist jobs, scaling from a 5B model on an RTX 4090 up to a 14B MoE flagship that demands 80GB VRAM.

pyVideoTrans wires together speech recognition, LLM translation, and voice synthesis into a single pipeline for local or API-driven video localization.

Because you shouldn't need to reverse-engineer a black box just to swap a noise scheduler or fine-tune a diffusion model.

Real-ESRGAN turns the ESRGAN research model into a practical tool for upscaling and restoring real-world images and videos using only synthetic training data.

Buzz wraps OpenAI's Whisper in a cross-platform GUI that keeps your audio data local and adds features the raw model doesn't have.

OpenAI's Whisper is accurate but slow and timestamp-imprecise; WhisperX bolts on batching, forced phoneme alignment, and speaker diarization to fix that.