Speech-to-text, text-to-speech, speaker diarization, speech enhancement, source separation, and VAD using next-gen Kaldi with onnxruntime without Internet connection. Support embedded systems, Android, iOS, HarmonyOS, Raspberry Pi, RISC-V, RK NPU, Axera NPU, Ascend NPU, x86_64 servers, websocket server/client, support 12 programming languages
13.5k
Stars
1.5k
Forks
596
Open issues
30
Contributors
AI Analysis
Sherpa-ONNX is a comprehensive speech AI toolkit enabling speech-to-text, text-to-speech, speaker diarization, speech enhancement, and related audio tasks entirely offline using ONNX Runtime. It excels at deployment on resource-constrained and edge devices (Raspberry Pi, embedded Linux, mobile platforms, NPUs) across 12 programming languages, serving developers and organizations requiring privacy-preserving, disconnected audio processing. It is not a general-purpose speech library but rather ...
Inferred from signals mentioned in the README (tests, CI, type safety) — not a review of the actual code.
AI's overall editorial judgment — not an average of the bars above, can weigh other factors too.
Offline speech AI toolkit for embedded and mobile: ASR, TTS, diarization, VAD across 12 languages and dozens of platforms
sherpa-onnx is a C++-core, ONNX-runtime-based library that runs a broad suite of speech AI functions—ASR (streaming and batch), TTS, speaker diarization, VAD, keyword spotting, source separation, and more—entirely offline. It targets embedded systems, mobile (Android, iOS, HarmonyOS), edge hardware (Raspberry Pi, RK NPU, Ascend NPU, RISC-V), and servers. With bindings for 12 programming languages and pre-built packages for each, it is aimed at developers who need production-ready, privacy-preserving speech AI without cloud dependency.
Spawned from the k2-fsa (next-gen Kaldi) lineage in September 2022, sherpa-onnx emerged as the deployment-focused counterpart to the research-oriented k2/lhotse stack, choosing ONNX Runtime for maximum portability across heterogeneous hardware.
Growth appears driven by the global demand for offline/on-device speech AI, growing interest in edge NPU hardware, and the practical frustration with cloud-only SDKs. Chinese developer communities (HarmonyOS, Axera, RK chips) contributed significantly to adoption. The 77 stars/week pace suggests steady, ongoing organic interest rather than viral spikes.
Hugging Face Spaces demos are publicly accessible, indicating real-world demo usage. The presence of HarmonyOS, WearOS, and specific Chinese edge-board support (RK3588, 旭日X3派, 爱芯派) suggests real hardware deployments beyond hobbyist use. A Discord community exists. The 1,501 forks-to-star ratio (~11%) is higher than typical for pure-research repos, suggesting active integration by downstream projects. Specific production deployments are not individually documented in the README excerpt, so full production-scale adoption cannot be independently verified.
Likely a C++ core library exposing a stable C API, with thin language-specific wrappers (Python, Java/Kotlin, Swift, Dart, etc.) built on top. ONNX Runtime handles inference across CPU and NPU backends. The README suggests a model-agnostic design: users supply ONNX model files, and the library handles pipeline logic. Streaming ASR likely uses a ring-buffer/chunk architecture. WebAssembly support suggests the core compiles to wasm via Emscripten.
Not documented in README
Last push on 2026-06-23 (the current date), indicating active daily development. With 13K+ stars and 1,500+ forks accrued over roughly 3.5 years and continuous recent commits, the project shows sustained, active maintenance rather than intermittent bursts. The breadth of platform additions (SpacemiT-K3, Axera NPU, QNN) in the README suggests an expanding scope driven by real hardware partners or user requests.
ADOPT IF: you need offline/on-device speech AI (ASR, TTS, VAD, diarization) on embedded Linux, Android, iOS, HarmonyOS, or edge NPU hardware, and cloud dependency is not acceptable. AVOID IF: you need maximum ASR accuracy with no latency or resource constraints and can use server-side Python pipelines freely—cloud-native stacks will offer simpler model management. MONITOR IF: you are evaluating speech AI for desktop or mobile apps where offline capability is desirable but not yet a hard requirement, and want to track ecosystem maturity before committing.
Independent dimensions
Mainstream potential
5/10
Technical importance
9/10
Adoption evidence
6/10
- Model management complexity: users must source and convert compatible ONNX models themselves, which may be a significant friction point for new adopters compared to cloud SDKs with managed model updates.
- Documentation breadth vs. depth tradeoff: supporting 12 languages, 30+ platforms, and 10+ functions may mean per-integration documentation is thinner than specialized single-platform libraries.
- Dependency on ONNX Runtime's own roadmap and NPU backend support: breakages in upstream ONNX Runtime or NPU-specific execution providers could cascade into sherpa-onnx.
- Core team concentration risk: the k2-fsa organization appears to be a relatively small academic/research-affiliated group; bus-factor for the C++ core is unknown but potentially limited.
- Rapidly evolving hardware landscape (new NPUs, SoCs) may create a maintenance burden that outpaces the team's capacity, leading to uneven platform support quality over time.
Likely to consolidate its position as the reference offline speech AI deployment library for edge/embedded use cases, particularly in Asian hardware ecosystems. Mainstream adoption in Western cloud-first environments may remain limited, but its niche is durable and growing.
Newsletter
Get analyses like this every Monday
Free weekly digest of the most interesting open-source discoveries.
Languages
No language breakdown available.
Information
- Language
- C++
- License
- Apache-2.0
- Last updated
- 12h ago
- Created
- 47mo ago
- Analyzed with
- anthropic/claude-haiku-4-5
Stars over time
No commit data available.
Contributors over time
Top 100 contributors only — repos with more will plateau at 100.
Open issues
No open issues — clean slate.
Open pull requests
No open pull requests.
Top contributors
Contributor data not available yet.
Recent releases
No releases published yet.
Similar repos
thewh1teagle/kokoro-onnx
kokoro-onnx is a text-to-speech (TTS) library that wraps the Kokoro TTS model...
FunAudioLLM/SenseVoice
SenseVoice is a multilingual speech foundation model supporting ASR, emotion...
huggingface/speech-to-speech
Speech-to-Speech is a modular, low-latency voice-agent pipeline that chains VAD...
| Repository | Stars | Week Δ | Language | Score | Updated |
|---|---|---|---|---|---|
|
k2-fsa/sherpa-onnx
this
|
13.5k | +132 | C++ | 8/10 | 12h ago |
|
thewh1teagle/kokoro-onnx
|
2.6k | — | Python | 8/10 | 5d ago |
|
FunAudioLLM/SenseVoice
|
8.8k | — | C | 8/10 | 1d ago |
 fikrikarim/parlor
|
1.9k | — | HTML | 7/10 | 1mo ago |
|
huggingface/speech-to-speech
|
5.8k | — | Python | 8/10 | 23h ago |
 OHF-Voice/piper1-gpl
|
4.7k | — | C++ | 7/10 | 1w ago |
FunASR (18K stars, Python) focuses on high-accuracy ASR models, primarily for server-side or cloud-assisted inference. sherpa-onnx targets embedded/offline/mobile deployment with a much broader function set and multi-language bindings. They serve partially overlapping but distinct deployment scenarios.
ONNX Runtime is the inference engine sherpa-onnx builds on top of. sherpa-onnx adds speech-specific pipeline logic, model management, streaming support, and platform packaging that raw ONNX Runtime does not provide. They are complementary, not competing.
Whisper is a model family; sherpa-onnx can run Whisper as one of its supported backends. sherpa-onnx provides the deployment infrastructure (streaming, VAD integration, mobile bindings) that Whisper alone lacks.
Vosk is the closest analog: offline ASR with mobile/embedded support. sherpa-onnx covers a broader function set (TTS, diarization, source separation), supports more recent model architectures, and explicitly targets NPU hardware, giving it an edge for newer edge deployments.
Silero provides VAD and ASR models in TorchScript/ONNX. sherpa-onnx explicitly supports Silero VAD as one of its backends, making it more of an integration consumer than a competitor. sherpa-onnx adds the full cross-platform deployment layer Silero models lack.
