Projects sharing features with Encodec

This is a PyTorch library for implementing vector and scalar quantization to create discrete latent representations in neural networks. It provides a suite of modules and utilities for converting continuous vectors into discrete codes, supporting architectures such as vector quantized variational autoencoders. The library features specialized mechanisms to maintain codebook health and efficiency, including random vector re-initialization to prevent codebook collapse and k-means centroid initialization to accelerate convergence. It supports diverse quantization strategies such as recursive res

This is a generative AI model library containing a collection of PyTorch and TensorFlow implementations for creating synthetic data and modeling complex probability distributions. It serves as a multi-framework repository of deep learning models designed for learning and replicating data patterns. The project provides specialized implementation suites for several generative architectures. This includes Generative Adversarial Networks using competing generator and discriminator models, Variational Autoencoder frameworks that map data to a latent space, and Restricted Boltzmann Machine and Deep

VALL-E-X is a neural speech synthesis framework and zero-shot text-to-speech engine. It functions as a multilingual synthesizer capable of generating natural human speech with control over emotion, pitch, and prosody. The project specializes in zero-shot voice cloning and cross-lingual voice replication, allowing the system to produce personalized speech in multiple target languages using short audio samples without additional training. It further enables cross-language accent manipulation and the ability to match the emotional tone and acoustic environment of a provided prompt. The implemen

Fairseq is a PyTorch toolkit for sequence-to-sequence modeling, specializing in neural machine translation, automatic speech recognition, and large-scale language model training. It provides a framework for processing and aligning diverse data sources, including text, audio, and video, to support tasks such as speech-to-text conversion and multimodal sequence learning. The project is distinguished by its distributed training capabilities, which utilize parameter sharding, mixed-precision training, and CPU offloading to handle models that exceed single-device memory. It also includes specializ

Moshi is a real-time voice foundation model and speech-to-speech framework designed for bidirectional, low-latency conversations. It functions as a full-duplex voice interface that processes audio and text concurrently in a single stream, enabling natural human-machine dialogue without sequential processing delays. The system utilizes a neural audio codec to compress high-fidelity audio into low-bitrate tokens for efficient transmission. To manage complex responses and reasoning, it employs internal monologue modeling, which generates a hidden stream of thought tokens alongside audible speech

Heartlib is an audio processing library for large language models that provides tools for audio tokenization, compression, and cross-modal alignment. It implements core models for audio-text embedding, automatic speech recognition, neural codecs, and text-driven audio synthesis. The project features a text-to-audio synthesis engine capable of generating high-fidelity music and speech from text descriptions or reference files. It also includes a neural audio codec designed for low-bitrate compression that preserves acoustic structure and sound quality. Additional capabilities cover audio-text

Amphion is an audio generation toolkit designed for the research and development of models that synthesize speech, music, and environmental sound effects. It provides a standardized framework for reproducible audio synthesis, incorporating a text-to-speech engine and a voice conversion framework. The project specializes in transforming audio identities, allowing for the modification of speaker accents and voice identities while preserving original rhythm and style. It also includes capabilities for singing voice synthesis and the generation of environmental soundscapes from text descriptions

Zonos is a controllable audio synthesis engine and large language model for text-to-speech. It serves as a multilingual speech generator capable of producing audio in English, Japanese, Chinese, French, and German. The system provides zero-shot voice cloning, allowing the replication of specific human voices using short audio samples. It supports the capture of nuanced behaviors, such as whispering, and provides parametric control over speaking rate, pitch, frequency, and emotional tone. The project covers a broad range of expressive speech synthesis and custom audio generation capabilities,

CSM is a conversational speech generation model and text-to-speech engine that converts text and audio inputs into synthetic speech. It utilizes a large language model architecture to predict and decode audio tokens for voice synthesis. The system functions as a zero-shot voice cloner, replicating specific speaker identities using short audio samples without requiring additional training. This enables precise control over speaker identity and the creation of synthetic speech that mimics a specific person. The model covers conversational speech synthesis and text-to-speech generation, transfo

This project is a music information retrieval library and research dataset designed for audio feature extraction and music genre classification. It provides a framework for training and evaluating machine learning models that categorize audio tracks into hierarchical genre structures, supported by a collection of open-licensed MP3 tracks and pre-computed features. The project includes a music metadata API client to fetch structured track, album, and artist information from external data sources. It utilizes these external integrations to map parent-child relationships between genres and organ

Python speech features is a signal processing toolkit and library for extracting standard speech recognition features from raw audio signals. It provides computational capabilities to calculate mel-frequency cepstral coefficients, raw and log filterbank energies, and spectral subband centroids for automated speech recognition systems and acoustic analysis. The library implements audio signal transformations including pre-emphasis filtering, overlapping windowed frame segmentation, Fast Fourier Transform spectral analysis, mel-filterbank projection with configurable filter weights and frequenc

This project is a PyTorch implementation of a text-to-image transformer. It is a generative AI model designed to map discrete text tokens to image pixels using a transformer network to create visual content from textual descriptions. The system utilizes a discrete VAE image encoder to compress visual data into tokens for transformer processing. It supports classifier-free guidance to adjust the influence of text prompts during inference and includes capabilities for ranking generated images based on their similarity to text prompts. The architecture incorporates sparse attention mechanisms a

MegaTTS3 is a bilingual speech synthesis system that generates natural-sounding speech in Chinese and English, including seamless code-switching within a single utterance. It functions as a text-to-speech engine, voice cloning system, and speech-to-text alignment tool, built around an acoustic latent compression model that encodes high-resolution audio into compact representations for efficient processing. The system distinguishes itself through accent intensity control, allowing adjustment of a speaker's accent strength in generated speech, and voice cloning from short audio samples for pers

ToonCrafter is a model that combines latent diffusion, reference-based colorization, and sketch-guided control for cartoon animation and interpolation. It functions as a cartoon video interpolation model, a reference-based colorization model, and a sketch-guided animation tool, all built on a latent diffusion animation framework. The project distinguishes itself by integrating three core capabilities into a single pipeline: generating smooth intermediate frames between two cartoon images using diffusion-based priors, transferring color and style from a reference image onto black-and-white ske

LatentSync is an audio-driven video generator and latent diffusion lip sync model designed to synchronize a speaker's lip movements in a video to a target audio track. It provides a lip synchronization training framework for developing synchronization networks on custom video and audio datasets. The system utilizes a video preprocessing pipeline to clean, segment, and align face data. It includes a visual sync evaluation tool that calculates confidence scores to measure the accuracy of audio and visual alignment in generated videos. The project covers capabilities for custom synchronization

AudioKit is an audio framework for iOS, macOS, and tvOS that provides tools for digital audio synthesis, signal processing, and audio analysis. It functions as a synthesis engine for generating audio waveforms and textures, a processing library for modifying tonal characteristics, and a toolkit for extracting frequency and amplitude data from sonic signals. The framework utilizes a modular node architecture and graph-based signal routing to connect audio generators, processors, and outputs. It wraps low-level audio primitives in high-level classes to facilitate sound generation and modificati

This project is a PyTorch implementation of a discrete variational autoencoder designed to compress high-resolution imagery into discrete latent representations. It functions as an image autoencoder that encodes visual data into discrete codes and decodes those codes back into reconstructed images. The system utilizes a latent space image compressor to convert images into a compressed, discrete format. This allows for generative image synthesis and the analysis of image compression by transitioning between raw pixels and discrete code sequences. The implementation covers latent space visuali

This project is a scientific computing framework for the .NET ecosystem, providing a comprehensive suite of libraries for numerical analysis, statistics, and mathematical optimization. It serves as a foundational toolkit for developing applications in machine learning, digital signal processing, and computer vision. The framework provides specialized toolkits for training and deploying predictive models, including neural networks, support vector machines, and decision trees. It further distinguishes itself with deep integrations for real-time visual analysis, such as object tracking and facia

TRELLIS.2 is a generative image-to-3D system that creates high-resolution 3D assets with physically based rendering materials from 2D images. It utilizes a sparse voxel representation to handle complex topologies and internal structures without relying on iso-surface fields. The project features a structured latent space representation that maps geometry and texture attributes to maintain visual fidelity. It employs an optimization-free geometry reconstruction process to decode latent representations directly into voxel grids and includes a PBR texture generator for synthesizing base color, r

This project is an end-to-end text-to-speech engine and deep learning voice synthesizer. It functions as a neural speech synthesis framework that converts written text directly into audio waveforms using a single neural network. The system implements an adversarial framework and a conditional variational autoencoder to generate high-fidelity artificial speech. It utilizes a generative adversarial network to ensure synthesized audio is indistinguishable from real human speech. The toolkit provides capabilities for neural speech synthesis, text-to-audio generation, and the training of custom v

Kolors is a generative model implementation for synthesizing photorealistic images from natural language descriptions and visual references. It utilizes a latent diffusion model framework to produce high-fidelity imagery, operating within a compressed latent space to improve generation efficiency and quality. The system functions as a multilingual image generator, interpreting text prompts in multiple languages to produce semantically accurate visual outputs. It includes a custom model training pipeline that uses low-rank adaptation to teach the model specific subjects or artistic styles from

This project is an AI voice training framework and singing voice conversion tool. It uses VITS and SoftVC technologies to transform the timbre of singing and spoken audio recordings, allowing a user to change the vocal characteristics of a recording to match a specific target speaker. The system provides a web-based voice converter interface for managing model checkpoints and performing timbre transformation and pitch shifting. It supports exporting trained models to the ONNX format for use in external interfaces and lightweight runtimes. The framework covers the full production pipeline, in

Librosa is a Python audio analysis library and digital signal processing framework. It functions as a feature extraction suite and music information retrieval tool designed to analyze the structural and sonic characteristics of audio signals. The library provides specialized capabilities for music analysis, including dynamic tempo tracking to identify rhythmic pulses and spectral feature extraction to compute harmonic spectra, chroma variants, and onset points. It also serves as a time-series audio processor for synchronizing audio streams. The system covers a broad range of audio processing

This project is a diffusion model training framework and image synthesis pipeline. It provides the tools necessary to train generative models to learn image data distributions through an iterative denoising process. The framework includes a generative model evaluation tool consisting of automated scripts used to measure the quality and accuracy of produced samples. The system covers model training pipelines and performance evaluation for generative diffusion models.

Taming Transformers is a generative system for high-resolution image synthesis that combines a vector-quantized GAN image encoder with an autoregressive transformer. It utilizes a discrete latent space to represent images as codebook tokens, enabling the production of high-fidelity visuals through a hybrid architecture. The project provides specialized capabilities for layout-based scene synthesis, allowing for the creation of complex images by placing objects according to defined bounding box coordinates. It also includes tools for image inpainting to fill missing sections of an image by ana

Tortoise-tts is a neural text-to-speech engine and voice cloning toolkit designed for high-quality audio generation. It functions as a zero-shot synthesis system, meaning it can generate speech for unseen speakers without requiring additional training or fine-tuning for each new voice. The system specializes in replicating human vocal characteristics using small sets of reference audio clips. It allows for the extraction of voice latents to mimic specific speakers, the generation of random synthetic identities, and the blending of multiple voice profiles to create hybrid vocal identities. Th

This project is a research framework and toolkit designed for training large-scale vision transformers and multimodal language models. It provides a comprehensive suite for vision-language pretraining, enabling the development of models that map images and text into shared latent spaces. The framework is distinguished by its capabilities in high-fidelity image generation and multimodal research, utilizing normalizing flows and variational autoencoders to produce images from text prompts or class labels. It supports the development of both generative and contrastive models, allowing for a wide