Projects

### Overview: This repository contains multilingual datasets with raw audio for speech emotion recognition, focusing on English, non-English languages, and non-speech vocalizations. It aims to provide a comprehensive set of resources for developing multilingual speech emotion recognition systems. ### Key features: - **English Datasets**: Raw audio with emotion labels from various sources. - **Non-English Datasets**: Audio samples in non-English languages. - **Non-Speech Datasets**: Affective human non-speech vocalizations. - **Multilingual Emotion Recognition**: Focuses on valence classification across multiple languages. ### Tech stack: Python ### Why it matters: This repository is crucial for researchers and developers looking to create multilingual speech emotion recognition systems, particularly for improving the quality of empathic responses in healthcare and customer service.

Overview: This repository contains the code necessary for running the multilingual DistilWhisper model from the Ferraz et al. 2024 IEEE ICASSP paper, which improves ASR performance in target languages by adding lightweight CLSR modules. Key features: - LoRA baseline: DistilWhisper/train/train_lora.sh - Model using only CLSR trained on ASR loss (no distillation): DistilWhisper/train/train_clsr.sh - Distillation model with JS loss (DistilWhisper): DistilWhisper/train/train_js_distill.sh - Distillation model with KL loss (From the appendix): DistilWhisper/train/train_kl_distill.sh Tech stack: Python Why it matters: This model offers enhanced ASR performance in target languages by leveraging language-specific experts, making it particularly useful for multilingual speech recognition tasks.

### Overview: Higgs is a Rust-based MLX server, UI, and model router designed for fast, dependency-free inference on Apple hardware. It serves models from Hugging Face, proxies to OpenAI and Anthropic, and translates between wire formats for seamless integration with existing tools. ### Key Features: - **Fast Inference:** Native MLX inference with prefix caching, continuous batching, quantized KV cache, and MoE prefill tuning. - **Single Endpoint:** A unified URL for local models and remote providers, with regex routes, aliases, and an optional auto-router. - **Drop-in for Tools:** Exports environment variables for OpenAI and Anthropic clients, making it compatible with Claude Code, Aider, and other OpenAI or Anthropic clients. - **Live Metrics:** Real-time monitoring of request times, decode tokens per second, and prefix-cache hits for local streaming requests. ### Tech Stack: Rust ### Why It Matters: - **Performance:** Native MLX inference with prefix caching and continuous batching for transformer families, quantized KV cache, and MoE prefill tuning. - **Unified Endpoint:** A single URL for local models and remote providers, with regex routes, aliases, and an optional auto-router. - **Tool Compatibility:** Exports environment variables for OpenAI and Anthropic clients, making it compatible with Claude Code, Aider, and other OpenAI or Anthropic clients. - **Real-Time Monitoring:** Real-time monitoring of request times, decode tokens per second, and prefix-cache hits for local streaming requests. ### GitHub Stars: 34

### Overview: This repository is an agentic workflow designed to tackle all open Erdős problems simultaneously. It tracks each problem as a separate computational and proof object, iterating across all problems rather than solving them individually. This approach allows for parallel research and deepens understanding through iterative refinement. ### Key Features: - Tracks 632 open Erdős problems. - Triages problems into `to-check` and `harder` categories. - Scripts, notes, and data for each problem are stored in separate directories. - Generates a progress and closure state count for each problem. - Provides a web UI for easy access and management. ### Tech Stack: JavaScript ### Why It Matters: This methodology optimizes research by leveraging parallel processing and systematic tracking, ensuring comprehensive coverage of open Erdős problems and deepening understanding through iterative refinement. ### Repository: - **GitHub Stars:** 34 - **Repository:** przchojecki/agentic-erdos - **Primary Language:** JavaScript - **GitHub Description:** Agentic workflow for tackling all open Erdős problems at once. ### Data and Notes: - **Canonical Data Directory:** `data/` - **Canonical Notes Directory:** `notes/` - **Canonical Scripts Directory:** `scripts/` ### Progress Status: - **Total Problems Tracked:** 632 - **Triaged as `to-check`:** 632 - **Triaged as `harder`:** 0 - **Latest Ref:** 2026-03-10T13:46:4

### Overview: This repository contains Python-based programming tutorials from my Italian YouTube channel and website. ### Key features: - Python in the Browser with PyScript - Creating a TCP Multi-Threaded Port Scanner - Creating an App for Smartphones (and Desktops) with Python and Kivy - Facial Recognition from Images - Part One - Facial Recognition from Images - Part Two - Real-Time Facial Recognition (Webcam) - Requests: HTTP for Humans - Tutorial - Speed Coding of a Reverse Shell Python - Creating Interface Graphics with TkInter - Creating a Complete Text Editor with Python and TkInter - Creating a Chatbot with Python and Chatterbot (updated code) - Accessing Twitter API with Python and Tweepy - Facial Recognition with Python and SpeechRecognition - Synthesizing Speech with Python and gTTS - Introduction to the Socket Module - Server-Client TCP with Socket Part 1 - Server-Client TCP with Socket Part 2 - Introduction to the ArgParse Module Part 1 - Introduction to the ArgParse Module Part 2 ### Tech stack: Python ### Why it matters: This repository provides a collection of Python-based programming tutorials from my Italian YouTube channel and website. It serves as a valuable resource for learning and practicing Python programming skills.

### Overview: VeloxQuant-MLX is a fast KV-cache quantization solution for Apple Silicon models, optimizing memory usage and accelerating inference performance. ### Key features: - **43 Methods**: 43 different compression methods adapted from published papers. - **Metal Kernels**: Runs on Metal kernels for up to 14.7× faster quantization. - **Compressed KV Cache**: Replaces the KV cache with a compressed version for improved memory efficiency. - **3 Lines of Code**: Scales any `mlx_lm` model with just 3 lines of code. ### Tech stack: - Python ### Why it matters: - **Memory Efficiency**: Reduces memory usage by up to 50% compared to standard quantization. - **Performance Acceleration**: Up to 14.7× faster quantization using Metal kernels. - **Production Validation**: Validated on 12 production models, including Llama and Mistral. ### GitHub Stars: 34 ### Repository: - **Repository**: [rajveer43/VeloxQuant-MLX](https://github.com/rajveer43/VeloxQuant-MLX) - **Primary Language**: Python - **GitHub Stars**: 34

Overview: This project is a Gradio viewer for Rerun, a tool for visualizing machine learning models, specifically designed for robotics applications. Key features: - Rerun viewer integration - Gradio interface for easy data visualization - Support for multimodal data visualization - Real-time updates and incremental streaming Tech stack: Python Why it matters: This project simplifies the integration of Rerun's visualization capabilities with Gradio, making it easier for developers to visualize and analyze machine learning models in real-time, especially useful in robotics applications where quick feedback is crucial. Repository: rerun-io/gradio-rerun-viewer GitHub stars: 34 GitHub description: Rerun viewer with Gradio README excerpt: ...

### Overview: TypelessMLX 是一款 macOS 語音聽寫應用程式，使用 MLX Whisper 和 Breeze-ASR-25 模型進行完全離線推理，專門設計以提供最佳的中文語音辨識體驗。 ### Key Features: - **一鍵聽寫**：按右 Option（可自訂）開始/停止錄音，自動貼入游標位置。 - **台灣中文優化**：使用Mac 語音辨識，採用台灣中文優化的 Breeze-ASR-25 模型，效果最佳。 - **自動標點符號**：語音辨識後自動補上逗號、句號、問號等標點。 - **AI 文字修正**（選用）：macOS 26+ 可啟用 Apple Intelligence 進一步精煉標點與語句。 - **完全離線**：所有推理在本機完成，無任何資料上傳。 - **多模型支援**：提供 Breeze-ASR-25（台灣中文）、Qwen3-ASR-0.6B（中文精度最佳）、Whisper Large v3 / Medium / Small（多語言）等模型。 - **語助詞過濾**（選用）：自動移除「呃」「嗯」等猶豫音。 - **聽寫歷史**：保留最近 N 筆紀錄，可一鍵複製。 - **系統選單列 App**：

### Overview: Xenolith is a pure C inference engine designed and optimized for Intel Xe-LP iGPUs and Gemma 4 26B-A4B QAT. It is specifically tuned for systems with 32 GB of RAM and Linux, and it targets the HP Envy 17 with an i7-13700H processor and DDR4-3200 dual channel memory. The engine is designed to maximize performance on common setups, leveraging the unique features of these hardware configurations. ### Key Features: - **GPU Optimization**: Xenolith is designed to work seamlessly with Intel Xe-LP iGPUs and Gemma 4 26B-A4B QAT, providing optimal performance for tasks that require both compute and bandwidth. - **Unified Memory**: The engine is compatible with systems that have unified memory, making it ideal for environments where memory bandwidth is a limiting factor. - **Comprehensive Compatibility**: Xenolith supports a wide range of Intel Xe integrated GPUs, including those based on the Xe architecture, and it is compatible with Linux using Level Zero (`intel-compute-runtime`). ### Tech Stack: - **Language**: C ### Why It Matters: - **Performance Optimization**: Xenolith is specifically designed to maximize performance on common setups, leveraging the unique features of Intel Xe-LP iGPUs and integrated GPUs. - **Balanced Performance**: The engine is built to provide a balanced setup for decode and prefill, allowing for a more efficient use of resources. - **Benchmark Results**: Xenolith

### Overview: This repository contains a comprehensive collection of 8 English speech emotion datasets designed for the development of speech emotion recognition (SER) models. The datasets include 7442 actors across various races and ethnicities, ranging from 20 to 74 years old, with 6 classes of emotions: Anger, Disgust, Fear, Happy, Neutral, and Sad. Additionally, the dataset includes features such as `Gender`, `Age`, and `Emotion Level`. The datasets also include text data for voice, aiding in multimodal modeling. ### Key Features: - **CREMA-D**: Contains 7442 actors across 6 classes of emotions. - **MELD**: Includes 1400+ dialogues and 13,000+ utterances from Friends, with 7 classes of emotions. - **MLEnd**: Features 42 unique languages and 154 speakers, with 4 classes of emotions. - **RAVDESS**: Includes 1440 actors with 7 classes of emotions. - **SAVEE**: Includes 120 utterances per speaker and 15 TIMIT sentences per emotion, with 7 classes of emotions. - **MELD**: Contains 13,000+ utterances from Friends, with 7 classes of emotions. - **RAVDESS**: Includes 120 utterances per speaker and 15 TIMIT sentences per emotion, with 7 classes of emotions. - **SAVEE**: Includes

### Overview: Circuitry.ai is an open-source tool that combines computer vision and large language models to detect, analyze, and explain electronic circuit diagrams. ### Key features: - Component detection using YOLOv8 - Wire detection using OpenCV - Intelligent textual explanation using LLaMA 3 ### Tech stack: HTML ### Why it matters: Circuitry.ai aims to bridge the gap between visual schematic representations and conceptual understanding, making electronics more accessible to learners, hobbyists, and engineers alike.

Overview: This Rust project implements speech detection using the Silero Voice Activity Detection (VAD) model, providing a practical tool for developers working with audio processing. Key features: - Speech detection using Silero VAD - Rust-based implementation Tech stack: Rust Why it matters: This project offers a straightforward and efficient way for developers to integrate speech detection capabilities into their Rust applications, leveraging the powerful VAD model for accurate audio analysis.

### Overview: VDTuner is a tool designed to automatically optimize vector database management systems, specifically focusing on improving performance and recall rates in vector databases. It leverages the Milvus database and the vector-db-benchmark framework to evaluate and tune the performance of vector databases. ### Key Features: - **Automated Tuning:** VDTuner optimizes the performance of vector databases by adjusting various parameters. - **Milvus Integration:** Utilizes Milvus for database management and indexing. - **Benchmark Framework:** Runs benchmark tests on vector databases to ensure performance metrics are accurate. ### Tech Stack: - Python 3.11 - Milvus 2.3.1 - Vector-db-benchmark ### Why It Matters: VDTuner is crucial for developers and system administrators looking to enhance the performance of vector databases, ensuring faster search speeds and improved recall rates, which are essential for applications requiring high search efficiency and accuracy.

Overview: A Bible verse search tool utilizing AI embedding and vector database to provide meaning-based searches. Key features: - Meaning-based search for Bible verses - Integration of AI embedding and vector database for enhanced search capabilities Tech stack: TypeScript Why it matters: This tool offers a new approach to Bible verse search by focusing on the meaning of the verses rather than keywords, potentially improving the relevance and accuracy of search results. Repository: tim-hub/bible-vector-search Primary language: TypeScript GitHub stars: 34

### Overview: ReelQL is a powerful tool that allows developers to extract meaningful insights from video content by converting it into structured JSON. This makes it easier to integrate video analysis into existing codebases, enabling developers to quickly and accurately summarize, categorize, and evaluate videos. ### Key Features: - **Video to JSON Conversion**: Converts any video link into a JSON document with timestamps and structured fields. - **Jev Judgment Model Integration**: Seamlessly integrates with TypeSafe's Jev model to extract key moments, brand safety, and other important information. - **Real-Time Analysis**: Analyzes videos in real-time, providing immediate feedback and insights. ### Tech Stack: - **TypeSafe**: Provides the Jev model for judgment. - **HTML**: Used for the frontend interface. ### Why It Matters: ReelQL simplifies the process of extracting valuable insights from video content, making it easier for developers to integrate video analysis into their applications. By leveraging Jev's judgment model, ReelQL ensures that the extracted information is accurate and contextually relevant, enhancing the overall quality of the analysis. ### Repository: - **Repository**: tomascupr/reelql - **GitHub Stars**: 34 - **GitHub Description**: Give your agent eyes: any video link in, one typed JSON out. A Claude skill + API. ### Primary Language: - **Primary Language**: HTML ### GitHub Description: - **Description**: Give your agent eyes: any video link in, one typed JSON out. A Claude skill + API. ### Examples: - **Video Links**:

Overview: This project combines Python scripting, vector embeddings, OpenAI, Pinecone, and Langchain to create an AI chatbot that allows users to quickly reference any content from a favorite YouTuber in less than 10 minutes. The chatbot uses the style and tone of the YouTuber and provides top 2-3 specific videos referenced hyperlinked. Key features: - Integrates with YouTube videos - Handles massive data efficiently - Uses vector embeddings for quick responses - Provides detailed answers with hyperlinked videos Tech stack: TypeScript Why it matters: This project simplifies the process of finding and referencing content from favorite YouTubers, making it easier for users to quickly recall specific videos and engage with the content in a more meaningful way.

### Overview: Mimora is a free, offline AI pronunciation trainer designed for English and Spanish. It uses your computer's audio capabilities to provide instant, word-level feedback, making it perfect for language learners and accent reduction practice. The app runs entirely on your machine, ensuring your voice stays local and no cloud storage is involved. ### Key Features: - **Offline Practice**: Your voice never leaves your computer, making it perfect for language learners and accent reduction practice. - **Unlimited Practice**: Mimora generates fresh phrases from your text, ensuring you never drill the same sentences. - **Instant Feedback**: Get a grade from 0 to 5 and the exact words to work on, with reference-vs-you replay. - **No Subscriptions**: Free and open-source, with no cloud storage or API keys required. - **Runs on a Normal PC**: Requires 16 GB RAM and no graphics card, with the ability to use an optional GPU for faster performance. - **Scoring Engines**: Supports `phoneme` for English (default, calibrated) and `none` for Spanish. ### Tech Stack: Python ### Why It Matters: Mimora is designed to be a practical, user-friendly tool for language learners and accent reduction practitioners. Its offline nature ensures that your voice remains local, making it ideal for those who prefer to practice without relying on the internet or cloud storage. The app's ability to generate fresh phrases from your text ensures that you never repeat the same sentences, and its instant feedback with reference-vs-you replay makes it an effective tool for improving pronunciation and accent.

Overview: This repository contains the source code for the paper "Exploring Adapter-based Transfer Learning for Recommender Systems: Empirical Studies and Practical Insights" by Westlake Research. The paper addresses the question of whether adapter-based learning techniques can achieve parameter-efficient transferable recommendations (TransRec) for recommendation systems with different item modalities. Key features: - The repository includes the source code for the paper. - It provides empirical studies to address key sub-questions. - It benchmarks existing adapters in item recommendation tasks. - It studies the effects of adapter-based TransRec by scaling up or down the training data. Tech stack: Python Why it matters: This repository offers valuable insights and practical guidance on unified and transferable recommendation, a less studied recommendation scenario.

### Overview: Gumiho is an innovative hybrid architecture designed to accelerate the auto-regressive token generation process of Large Language Models (LLMs) using speculative decoding. Unlike traditional methods, Gumiho prioritizes early tokens in the draft sequence, enhancing the overall acceptance length with improved performance and efficiency. ### Key Features: - **Hybrid Architecture**: Combines serial and parallel components for early and later tokens, respectively. - **Transformer Head**: Enhanced accuracy for early tokens. - **MLP Heads**: Efficiently parallelized for later tokens. - **Full Tree Attention (FTA)**: Improves mean accepted tokens without additional computational overhead. ### Tech Stack: Python ### Why it Matters: Gumiho significantly accelerates the auto-regressive token generation process of LLMs, leading to improved performance and acceptance lengths with reduced processing time per draft-verification cycle.

Overview: This repository contains prompts designed for GPT Image 2.5, paired with the actual images generated by the model. Each prompt is crafted based on the capabilities described in OpenAI's announcement, showcasing both the fidelity and precision of the generated images. Key features: - Prompts for GPT Image 2.5 - Realistic 2.5 images produced by the model Tech stack: TypeScript Why it matters: These prompts provide a practical way to understand and replicate the capabilities of GPT Image 2.5, offering insights into the model's strengths and weaknesses. Repository: AtlasCloudAI/awesome-gpt-image-2.5-prompts GitHub stars: 33