k2-fsa/sherpa-onnx
OFFLINE SPEECH PROCESSING ON ONNXRUNTIME
Library for speech recognition and synthesis, diarization and audio enhancement. Runs locally on 12 programming languages and embedded systems without internet.
// KEY FEATURES
- ASR, TTS, diarization, VAD, source separation and speech enhancement in one package
- Support for 12 programming languages, WebAssembly, Flutter and Tauri
- Runs on x86, ARM, RISC-V, RK NPU, Ascend NPU, Qualcomm NPU without internet
- Streaming and non-streaming recognition, forced alignment, language identification
// DEEP DIVE: HOW IT WORKS
Sherpa-onnx is built on the next-gen Kaldi and uses ONNX Runtime for inference. All components — ASR, TTS, VAD, diarization, source separation, speech enhancement, speaker and language identification, Keyword Spotting, punctuation and forced alignment — work locally without network. Unified API for 12 languages: C++, Python, Java, Swift, Kotlin, Dart, Go, Rust, C#, JavaScript, Pascal and WebAssembly.
Unlike cloud services and heavy frameworks, sherpa-onnx targets edge devices: Raspberry Pi, RISC-V, RK NPU, Ascend NPU, Qualcomm QNN, Axera NPU, Intel OpenVINO. Prebuilt binaries for Android, iOS, HarmonyOS, Windows, macOS, Linux (x86, x86_64, arm32, arm64, riscv64). Flutter and Tauri plugins with demo apps accelerate integration.
ONNX models required — downloaded separately or trained independently. NPU acceleration needs vendor SDKs (RKNN, QNN, CANN, AXERA SDK, OpenVINO). Streaming ASR with low latency, quality depends on model. No training — inference only.
// USE CASES
Best for: Developers of voice interfaces on embedded systems and mobile devices
- Add offline speech recognition to a mobile app on Flutter or Tauri
- Run diarization and audio enhancement on Raspberry Pi or RK NPU without cloud
- Embed streaming ASR and TTS into an embedded system on RISC-V or ARM
New repositories every 30 minutes
REDDYX AI scans GitHub 24/7 and ships the best AI/ML/Web3 projects to Telegram.
Join on Telegram