{"id":44526392,"url":"https://github.com/miuda-ai/lele","last_synced_at":"2026-02-13T18:04:25.920Z","repository":{"id":335080623,"uuid":"1143972422","full_name":"miuda-ai/lele","owner":"miuda-ai","description":"Bare-Metal Rust Audio AI Framework","archived":false,"fork":false,"pushed_at":"2026-02-07T14:30:54.000Z","size":411,"stargazers_count":121,"open_issues_count":2,"forks_count":10,"subscribers_count":1,"default_branch":"main","last_synced_at":"2026-02-07T23:11:48.983Z","etag":null,"topics":["audio","onnxruntime","pytorch"],"latest_commit_sha":null,"homepage":"https://miuda.ai/?ref=lele","language":"Rust","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":null,"status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/miuda-ai.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":null,"code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-01-28T06:28:26.000Z","updated_at":"2026-02-07T14:30:57.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/miuda-ai/lele","commit_stats":null,"previous_names":["miuda-ai/lele"],"tags_count":0,"template":false,"template_full_name":null,"purl":"pkg:github/miuda-ai/lele","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/miuda-ai%2Flele","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/miuda-ai%2Flele/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/miuda-ai%2Flele/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/miuda-ai%2Flele/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/miuda-ai","download_url":"https://codeload.github.com/miuda-ai/lele/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/miuda-ai%2Flele/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":29414277,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-02-13T06:24:03.484Z","status":"ssl_error","status_checked_at":"2026-02-13T06:23:12.830Z","response_time":78,"last_error":"SSL_connect returned=1 errno=0 peeraddr=140.82.121.5:443 state=error: unexpected eof while reading","robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":false,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["audio","onnxruntime","pytorch"],"created_at":"2026-02-13T18:03:00.753Z","updated_at":"2026-02-13T18:04:25.907Z","avatar_url":"https://github.com/miuda-ai.png","language":"Rust","funding_links":[],"categories":[],"sub_categories":[],"readme":"# lele: Bare-Metal Rust AI Framework\n\n**lele** is a standalone, dependency-free inference engine for intelligence, built from scratch in pure Rust.\n\nIt rejects the \"general-purpose runtime\" approach (wrapping C++ libs like ORT or using heavy Torch ports) in favor of **hand-crafted, domain-specific kernels**.\n\n## Overview\n\n`lele` is designed to run deep learning models (specifically speech-related ones like SenseVoice, Silero VAD, and TTS, even `yolo26` ) with minimal overhead. \n\n## Performance Benchmarks (2026-02-12)\n\nIn-depth comparison between **lele** and **ONNX Runtime (CPU)** on macOS (Apple Silicon). All benchmarks run with single-thread affinity for fair comparison.\n\n| Model | ORT RTF (CPU) | lele RTF | Speedup |\n| :--- | :--- | :--- | :--- |\n| **Silero VAD** | 0.0031 | 0.0016 | 1.93x |\n| **SenseVoice** | **0.032** | 0.051 | 0.63x |\n| **Supertonic** | **0.122** | 0.134 | 0.91x |\n| **Yolo26** | **759.19** | 1050.56ms | 0.72x |\n\n*Note: RTF (Real-Time Factor) is defined as (Inference Time / Audio Duration). Lower is better.*\n\n## Key Features\n\n- **Zero Runtime Dependencies**: Generated models are pure Rust.\n- **AOT Compilation**: Converts ONNX models to specialized Rust code for maximum performance.\n- **SIMD Optimized**: Hand-written kernels using Apple Silicon (NEON) and x86_64 (AVX/SSE) intrinsics.\n- **Memory Efficient**: Static buffer allocation and zero-copy weight loading.\n- **Speech Optimized**: Built-in feature extraction for audio (FFT, Mel-spectrogram, LFR, CMVN).\n- **WebAssembly Ready**: Full browser compatibility with WASM SIMD128 optimizations.\n\n## WebAssembly Support\n\n**lele** compiles to WebAssembly and runs ML inference directly in the browser with **no server required**.\n\n\u003cimg src=\"docs/yolo26.png\" width=\"600\" alt=\"YOLO26 Object Detection in Browser\"\u003e\n\n### WASM Performance Optimizations\n\n| Optimization | Impact |\n|--------------|--------|\n| **WASM SIMD128** | Tiled matmul micro-kernel with `f32x4_mul`/`f32x4_add` (4x unroll) |\n| **Optimized Activations** | SIMD paths for tanh/sigmoid/relu/silu using polynomial exp approximation |\n| **Vectorized Normalization** | SIMD softmax and layer_norm with horizontal reduction |\n| **Release Settings** | `opt-level=3`, `lto=true`, `codegen-units=1`, `panic=\"abort\"` |\n| **Post-Processing** | `wasm-opt -O3` for additional 5-15% size/speed gains |\n\n**Binary Size Reduction**: Dev builds (2.9M → 1.7M for SenseVoice, 42% smaller with optimizations)\n\n**Expected Runtime Speedup**: **20-100x** over unoptimized scalar WASM (10-50x from release mode + 2-4x from SIMD128)\n\n### Try the Web Demo\n\n```bash\ncd examples/web-demo\n./build_wasm.sh\npython3 -m http.server 8080 -d web\n# Open http://localhost:8080\n```\n\nSee [examples/web-demo/README.md](examples/web-demo/README.md) for details.\n\n## Supported Models\n\n- **SenseVoiceSmall**: High-accuracy multi-lingual ASR.\n- **Silero VAD**: Reliable Voice Activity Detection.\n- **Supertonic**: Fast and high-quality Text-to-Speech.\n- **Yolo26**: Real-time object detection.\n\n## Getting Started\n\n### Prerequisites\n\n- Rust (Latest stable)\n- `cargo`\n\n### Compilation \u0026 Generation\n\nTo compile an ONNX model into Rust code:\n\n```bash\ncargo run --release --bin lele_gen -- \u003cmodel.onnx\u003e \u003coutput_path.rs\u003e\n```\n\n### Running Examples\n\n```bash\n# SenseVoice ASR\n./run_sensevoice.sh\n\n# Supertonic TTS\n./run_supertonic.sh\n\n# Silero VAD\n./run_silero.sh\n\n# Yolo26 Object Detection\n./run_yolo26.sh\n```\n\n\n## Roadmap\n\n1. Performance optimizations (SIMD, multi-threading, etc.), better than ONNX Runtime.\n2. Support for more audio models (e.g., Whisper, CosyVoice, etc.)\n3. GPU acceleration backend (wgpu); Quantization (INT8/FP16)\n4. Advanced attention mechanisms (FlashAttention, PagedAttention)\n5. Voice API server (RESTful service), including ASR/TTS/Denoise endpoints.\n\n## License\n\nMIT\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fmiuda-ai%2Flele","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fmiuda-ai%2Flele","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fmiuda-ai%2Flele/lists"}