Projects

Overview: This is a React Native library for speech recognition, wake-up, and synthesis provided by Baidu. Key features: - Speech Recognition - Voice Wake-Up - Voice Synthesis Tech stack: Java Why it matters: It simplifies the integration of speech capabilities into React Native applications, making it easier to implement voice-based features. Repository: gdoudeng/react-native-baidu-asr GitHub stars: 33

Overview: This repository contains a collection of small Tkinter-based applications created by the author. Key features: - A simple calculator - A password manager - A to-do list application Tech stack: Python Why it matters: These projects showcase basic Tkinter functionalities and are useful for learning or creating small applications quickly.

### Overview: This repository provides a complete workflow for fine-tuning Qwen3-TTS (0.6B-Base or 1.7B-Base) to generate high-quality Hong Kong Cantonese speech with authentic accent, natural intonation, prosody, and local slang. It includes single-speaker voice cloning, multi-speaker LoRA export, automatic speaker clustering, LoRA fine-tuning, and a voice clone server for instant inference. ### Key features: - Single-speaker voice cloning / fine-tuning - Multi-speaker LoRA export - Automatic speaker clustering - LoRA fine-tuning (recommended for 8–12 GB GPUs) - Voice Clone Server - Model evaluation (WER/CER, speaker similarity, UTMOS) - Cantonese dataset preparation pipeline - Training with MLflow experiment tracking - Gradient accumulation + Flash Attention-2 support ### Tech stack: Python 3.11 or 3.12, FFmpeg, pip install -r requirements.txt ### Why it matters: This repository offers a comprehensive solution for generating high-quality Cantonese speech using the Qwen3-TTS model, ensuring authentic accent, natural intonation, prosody, and local slang. The workflow includes all necessary components to fine-tune the model, prepare the dataset, and evaluate its performance, making it a valuable resource for developers and researchers in the field of speech synthesis and natural language processing.

### Overview: SelfJev is a self-hosted decision model that turns text and questions into structured decisions using a 4B model running on your own infrastructure. It provides a Python SDK and a Jev-compatible HTTP API, making it easy to integrate into applications. ### Key Features: - **Self-hosted Decision Model**: Runs on your own GPU, providing complete control over the model and infrastructure. - **Python SDK**: Simplifies integration into applications. - **Jev-Compatible API**: Compatible with the Jev decision model, allowing seamless integration with existing systems. ### Tech Stack: Python ### Why It Matters: SelfJev enables developers to turn text and questions into structured decisions, making it easier to integrate into applications and provide contextually relevant answers directly from your infrastructure.

### Overview: Anda DB is a specialized Rust library designed as a knowledge memory system for AI agents. It provides an embedded schema-aware document database with three retrieval modes: B-Tree indexes for exact match and range filters, BM25 indexes for full-text search, and HNSW indexes for vector similarity search. The library also offers a portable object-store-backed persistence layer, a KIP (Knowledge Interaction Protocol) SDK, and two reference KIP engines. ### Key features: - **Document Database**: An embedded schema-aware document database with three retrieval modes. - **KIP Protocol**: Provides a portable object-store-backed persistence layer and a KIP SDK. - **Knowledge Graph Memory**: Records who claimed what, on what evidence, and what changed when through KIP and a Cognitive Nexus. - **Deployment Modes**: Supports long-term memory for AI agents, embedded hybrid retrieval, and knowledge-graph memory. ### Tech stack: Rust ### Why it matters: Anda DB is designed for applications that require more than a plain key-value store but less than a full external database service. It provides a modular and durable solution for long-term memory, embedded retrieval, and knowledge-graph memory, making it suitable for AI agents, knowledge management systems, and private or regulated deployments.

Overview: ChatIR is an innovative chat-based image retrieval system designed for NeurIPS 2023, leveraging large language models to enhance image search capabilities. Key features: - Engaging in conversations with users to clarify search intents - Leveraging large language models to generate follow-up questions - Achieving over 78% success rate in image retrieval after 5 dialogue rounds - Comparing with strong baselines derived from Reinforcement Learning Tech stack: Python Why it matters: ChatIR offers a more user-friendly and effective image retrieval solution by engaging in conversations with users, significantly improving the accuracy of image searches compared to traditional methods. Repository: levymsn/ChatIR GitHub stars: 32 GitHub description: Official repository of "Chatting Makes Perfect: Chat-based Image Retrieval"

Overview: ChatterPal is a self-hosted English speaking practice application that offers guided conversations, recording confirmation, feedback, and spaced review to improve your spoken English skills. Key features: - Contextual dialogues for familiar scenarios - Audio recording and verification - Original text feedback and review - Delayed review for independent practice - Historical progress and data persistence - Free practice mode with customizable options Tech stack: React + FastAPI Why it matters: ChatterPal provides a comprehensive tool for English language learners to practice their spoken skills through engaging and personalized conversations, ensuring continuous improvement and retention of the material. Repository: likebeans/ChatterPal GitHub stars: 32 GitHub description: Self-hosted English speaking practice with guided conversations, recording confirmation, feedback, and spaced review. React + FastAPI.

Overview: This Python script allows you to summarize YouTube videos by transcribing the audio with OpenAI's Whisper and generating concise video summaries using GPT. Key features: - Transcribes YouTube video audio with Whisper - Generates a summary using GPT - Downloads YouTube video's audio - Supports multiple YouTube video formats Tech stack: Python, Whisper, OpenAI, Pytube Why it matters: This script provides a practical solution for summarizing YouTube videos, making it easier for users to quickly grasp the main points of a video without having to watch it in full.

Overview: This repository implements the "Task-conditioned Domain Adaptation for Pedestrian Detection in Thermal Imagery" paper, focusing on improving pedestrian detection in thermal imagery using a task-conditioned network architecture. Key features: - Task-conditioned network architecture for simultaneous classification and detection tasks - Improved training procedures including learning rate strategy and validation recall adjustment - Comparison of Log Average Miss Rate with state-of-the-art on KAIST dataset - Monitoring training loss and validation performance during training - Functions for loading and saving weights, model, dataloader, evaluation, and detection Tech stack: Python Why it matters: This implementation showcases advancements in domain adaptation for pedestrian detection, demonstrating the effectiveness of task-conditioned networks in improving detection accuracy in thermal imagery. Repository: mrkieumy/task-conditioned GitHub stars: 33

### Overview: A framework in Italian and English that enables chatting with your documents in RAG, including multimedia (audio, video, images, and OCR). It is an Italian and English framework, which allows you to chat with your documents in RAG, including multimedia (audio, video, images and OCR). ### Key Features: - Chat with a GPT model without RAG. - Create personalized workspaces. - Indicate a folder with various types of documents for indexing. - Transcribe audio and video in documents, perform OCR on images, and describe 10 frames from the video. - Send a single file through upload and the system will describe it. - Connect to the internet only for downloading models from Hugging Face, and disconnect after starting. - Use the Tkinter package if Tkinter is not working in Windows. ### Tech Stack: Python 3.12.4 ### Why It Matters: This framework simplifies the process of interacting with documents in RAG, making it easier to chat with your files, transcribe audio and video, and perform OCR on images. It's particularly useful for those who need to quickly access and understand their documents, whether they're in Italian or English.

Overview: Whisper Speech-to-Text is a JavaScript library for real-time transcription of speech to text using OpenAI's Whisper ASR system, designed for web applications. Key features: - Real-time transcription of speech to text using OpenAI's Whisper ASR system. - Easy to use API for starting, pausing, resuming, and stopping recordings. - Automatic handling of microphone permissions and audio recording. Tech stack: TypeScript Why it matters: This library provides a convenient and efficient way to integrate speech-to-text functionality into web applications, making it easier for developers to implement natural language processing in their projects. Repository: nitaiaharoni1/whisper-speech-to-text Primary language: TypeScript GitHub stars: 33

Overview: TinyCog is a small robot platform designed for embedded systems, including the Raspberry Pi, featuring speech, vision, and movement functionalities. It leverages OpenCog for high-level reasoning, language understanding, generation, and movement planning. Key features: - Speech recognition and synthesis - Vision-based emotion and gesture analysis - Robotic movement control - OpenCog scripting system for high-level reasoning Tech stack: C++ Why it matters: TinyCog provides a comprehensive solution for small robots, enabling them to interact with their environment through speech and vision, and perform coordinated movement, all powered by OpenCog's advanced AI capabilities. Repository: opencog/TinyCog GitHub stars: 33

Overview: This project introduces a pipeline for giving any Python CLI tool a `-w` flag that takes plain English commands, using locally-trained SLMs (CPU inference). The solution is designed for local, offline use, with a small model per library, and it comes with the package. Key features: - Generating flag grammar for training - Auditing the dataset against held-out prompts - Fine-tuning on a free Colab T4 or single GPU - Quantizing and shipping the model Tech stack: Python Why it matters: This project addresses the common issue of having to look up command-line flags every time, providing a solution that is both local and efficient, using a small model per library, and it comes with the package. Repository: pranavkumaarofficial/nlcli-wizard Primary language: Python GitHub stars: 33 GitHub description: Natural language control for Python CLI tools using locally-trained SLMs (CPU inference) README excerpt: # nlcli-wizard [](https://github.com/pranavkumaarofficial/nlcli-wizard/actions/workflows/tests.yml) [](LICENSE) [](https://www.python.org/downloads/)

Overview: This repository identifies a voice as male or female using acoustic properties of the voice and speech. Key features: - Dataset sourced from Kaggle - Includes 3,168 voice samples from male and female speakers - Measures acoustic properties such as mean frequency, standard deviation, median, quantiles, interquartile range, skewness, kurtosis, spectral entropy, spectral flatness, mode frequency, frequency centroid, peak frequency, average fundamental frequency, minimum fundamental frequency, maximum fundamental frequency, average dominant frequency, minimum dominant frequency, maximum dominant frequency, range of dominant frequency, modulation index, and label (male or female) Tech stack: Jupyter Notebook Why it matters: This dataset and analysis can be used for gender recognition in speech and voice applications, such as in security systems, entertainment, or educational tools, providing a robust method based on acoustic properties for identifying speakers' genders.

Overview: Proximum is a high-performance, embeddable vector database designed for Clojure and Java, offering Git-like versioning and copy-on-write branching. It provides persistent data structure semantics, efficient immutability, and true persistence, making it ideal for RAG applications, semantic search, and ML experimentation. Key features: - Git-like versioning - Copy-on-write branching - Immutability - True persistence - SIMD-accelerated search with competitive recall - Pure JVM implementation Tech stack: Clojure Why it matters: Proximum offers a robust, efficient, and scalable solution for vector search, making it suitable for applications requiring version tracking, A/B testing, and reproducible search results. Repository: [replikativ/proximum](https://github.com/replikativ/proximum) GitHub stars: 33 GitHub description: Versioned, fast and scalable nearest neighbor search.

Overview: This repository is an official implementation of the Listen, Attend and Spell (LAS) model for articulatory features estimation at the Interspeech 2019 conference. Key features: - Recipes for common datasets: Librispeech, Common Voice, TIMIT, L2 articulatory, SLR32, and VCTK. - Support for phones, words, characters, and binary features data collection. - Decoder for phonological features. - Multitask training for phones recognition and indicators estimation. - Joint training with CTC loss on the encoder. Tech stack: Python Why it matters: This implementation provides a practical solution for estimating articulatory features using the Listen, Attend and Spell architecture, making it useful for researchers and developers working on speech recognition and articulatory feature detection.

### Overview: This repository provides a comprehensive explanation of OpenAI's paper on DALL-E, a zero-shot text-to-image generation model. It details the methodology used to create the model, including the discrete Variational Auto-Encoder (dVAE) and the autoregressive transformer. The notebook demonstrates how to generate images from text captions using this model, showcasing the capabilities and limitations of the approach. ### Key Features: - **DALL-E Model**: An autoregressive language model trained on 250 million image-text pairs. - **Discrete Variational Auto-Encoder (dVAE)**: A VAE that compresses images into 32x32 grids of discrete image tokens. - **Transformer Autoregressive Model**: A transformer that models the text and image tokens in a single stream of data. - **Output Generation**: The notebook shows how to generate images from text captions using the dVAE and autoregressive transformer, demonstrating the consistency of the generated images with the input text. ### Tech Stack: - **Jupyter Notebook**: The primary language used for the notebook. ### Why It Matters: - **Zero-Shot Text-to-Image Generation**: The ability to generate images from text inputs without explicit training on a fixed set of prompts. - **Consistency**: The model's output is consistent with the input text, making it useful for creative and artistic applications. ### Repository: - **simonsanvil/DALL-E-Explained**: The primary language is Jupyter Notebook. - **GitHub Stars**: 33 - **GitHub Description

### Overview Starfish is a Python library designed to simplify the process of generating synthetic data for AI models. It leverages structured outputs and efficient parallel processing to adapt to your workflow, allowing you to define the exact format of your data and scale seamlessly from experiments to production. ### Key Features - **Structured Outputs**: First-class support for structured data through JSON schemas or Pydantic models. - **Model Flexibility**: Supports any LLM provider—local models, OpenAI, Anthropic, or your own implementation via LiteLLM. - **Dynamic Prompts**: Built-in Jinja2 templates for dynamic prompts. - **Easy Scaling**: Transforms any function to run in parallel across thousands of inputs with a single decorator. - **Resilient Pipeline**: Automatic retries, error handling, and job resumption—pause and continue your data generation anytime. - **Complete Control**: Shares state across your pipeline, extends functionality with custom hooks. ### Tech Stack Starfish is built with Python, leveraging its robust ecosystem for data processing and machine learning. ### Why It Matters Starfish simplifies the complex process of generating synthetic data, making it easier to fuel AI models. By providing a flexible and efficient solution, it helps developers focus on their core tasks without worrying about the intricacies of data generation. ### GitHub Stars Starfish has been starred by 33 developers, indicating its growing popularity and usefulness in the community. ### Installation Install Starfish using the following command: ```bash pip install starfish-core ``` For optional dependencies, you can install specific parsers or

Overview: This repository collects and processes multimodal datasets, including visual, audio, and text modalities from various sources such as TV series "Friends," dyadic conversations, and robot-human interactions. It provides a structured directory for easy access and processing. Key features: - Extracts raw datasets into a structured format - Supports multiple datasets including MELD, IEMOCAP, and CarLani - Provides detailed README files for each dataset - Includes face, visual, audio, and text features extracted from the datasets Tech stack: Python Why it matters: This repository offers a comprehensive collection and processing framework for multimodal datasets, making it useful for researchers and developers working on multimodal AI applications.

### Overview: A reverse-engineered Apple Neural Engine (ANE) LLM inference engine designed for Apple Silicon, providing a native native inference solution for large language models (LLMs) on the M1, M2, M3, M4, and M5 chips. This engine matches the decode speed of llama.cpp and achieves 3.6 TFLOPS performance with fused ANE mega-kernels. ### Key features: - **ANE Inference Engine**: Runs Qwen3.5-2B on three Apple Silicon accelerators simultaneously. - **Metal GPU Decode**: Single-token decode with 13 custom compute shaders, optimized for performance. - **CPU Parallel GEMV**: Utilizes NEON and Accelerate BLAS for parallel processing. - **Reverse Engineering**: Cracked open Apple's private Neural Engine framework, including key discoveries and successful implementations. ### Tech stack: Rust + Obj-C + Metal. ### Why it matters: This engine provides a low-power, high-performance inference solution for large language models, leveraging the strengths of Apple's hardware and reverse-engineered private APIs to achieve superior performance on Apple Silicon devices.