Our take
- Activity
- 84
- Community
- 62
- Issues
- 80
- Pull requests
- 80
+654 stars in the last 30 days
- Merges more pull requests than 88% of the projects we track
- Releases more often than 85% of the projects we track
- More contributors than 78% of the projects we track
Project history
Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.
Stars
16,011Forks
1,882Change since Nov 17, 2026: +19 stars·0 forks
Overview
sherpa-onnx is a C++ speech and audio inference toolkit built around ONNX Runtime. It lets applications run speech models on-device or on their own systems, avoiding a required cloud connection for inference.
The project suits developers who need speech capabilities across varied hardware and programming languages. Its scope includes both real-time and non-streaming speech recognition, alongside synthesis and other audio tasks. Model choice and deployment setup depend on the target platform and use case.
Key Features
- Streaming and non-streaming speech-to-text.
- Text-to-speech, including support for voice cloning models.
- Speaker diarization, identification, and verification.
- Voice activity detection, keyword spotting, and spoken-language identification.
- Audio tagging, speech enhancement, and source separation.
- APIs for 12 languages, including C++, C, Python, JavaScript, Java, C#, Go, Rust, Dart, and Pascal.
- Platform support includes Linux, Windows, macOS, Android, iOS, HarmonyOS, and WebAssembly, plus several embedded architectures.
- NPU options include Rockchip, Qualcomm, Ascend, Axera, and Intel, as listed in the project documentation.
Use Cases
- Build offline transcription features into mobile, desktop, or embedded applications when audio should stay on-device.
- Add real-time speech input to voice interfaces, captioning tools, or meeting utilities using a streaming ASR model.
- Generate spoken output in applications that need local text-to-speech rather than a remote service.
- Analyze recorded or live audio for speaker turns, voice activity, keywords, or audio categories.
- Deploy speech workloads on constrained ARM or RISC-V devices, selecting an appropriate model and supported runtime for the hardware.
Getting Started
Start with the official documentation and repository README to choose a supported model, platform, and API. The README links to model catalogs, platform-specific examples, and prebuilt demos. This repository does not provide a single universal install command in the supplied README excerpt.
Alternatives
- whisper.cpp: whisper.cpp focuses on offline speech recognition with Whisper models, while sherpa-onnx supports multiple speech and audio tasks through ONNX Runtime.
- tortoise.cpp: tortoise.cpp focuses on local text-to-speech with Tortoise, while sherpa-onnx supports a broader range of speech and audio tasks.
| Project | Language | License | Stars | Status |
|---|---|---|---|---|
| sherpa-onnx | C++ | Apache-2.0 | 15.1k | Not checked yet |
| whisper.cpp | C++ | MIT | 54.1k | Active |
| tortoise.cpp | C++ | MIT | 192 | Inactive |
Considerations
- sherpa-onnx is an inference runtime, not a single ready-made speech model. Choose and obtain a compatible pretrained model for the task and language you need.
- Cross-platform support does not mean every feature or architecture is available everywhere. Check the platform tables and feature-specific documentation before committing to a target.
- Embedded and NPU deployments require hardware-specific setup and compatible model/runtime combinations.
- The project supports many APIs and targets, so integration and packaging steps vary. Consult the relevant examples rather than assuming one build process applies to all platforms.
Found this useful?
Share it with someone who would like sherpa-onnx.