Skip to content
← Models
Open weightsk2-fsa (Next-gen Kaldi team)/10 trustVerified

Sherpa-ONNX

One codebase for offline speech AI on every device: streaming ASR, TTS, diarization, and enhancement on ONNX Runtime.

Visit official page →

About

Sherpa-ONNX is the deployment arm of the next-gen Kaldi (k2-fsa) project: an open-source speech stack that runs entirely offline on ONNX Runtime. It covers streaming and batch speech recognition, text-to-speech, speaker diarization, voice activity detection, speech enhancement, source separation, and keyword spotting, with pretrained Zipformer transducer models and bindings for around a dozen programming languages.

Its range of targets is the selling point. The same codebase ships to Android, iOS, HarmonyOS, Raspberry Pi, RISC-V boards, and NPU-equipped hardware, plus servers over a WebSocket interface. That makes it the default pick for teams that need ASR on devices with no reliable connectivity, from hearing-assistive apps to on-device live captioning, without paying per-minute API fees.

Specs

Context window

Streaming real-time or batch; per-model (tiny Zipformer to large)

Pricing

Free, self-hosted

Free quota

No quota; everything runs locally with no internet connection

API access

No hosted API; embeddable libraries in ~12 languages, CLI tools, and a WebSocket server/client

Weights

Open weights

Released

Not disclosed

Benchmarks

No published standard benchmark headlines; streaming Zipformer transducers with ~100 ms partial latency in community builds (third-party report)

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

on-device-asrstreaming-asrvoice-assistantslive-captioningembedded-speech

Trust score — 8/10

/10 trustVerified

Scored 8/10 (verified tier, metrics-v1). Evidence: Apache-2.0 license and 9,000+ GitHub stars confirmed via package mirrors; platform coverage (Android, iOS, HarmonyOS, Raspberry Pi, RISC-V, NPUs) and the 12-language binding count from the project's own docs. Strengths: the broadest on-device speech stack in open source, active maintenance (v1.13.x, 2026), one deployment path for ASR/TTS/diarization. Docked for: (1) no published standard benchmark figures to anchor accuracy claims; (2) it is a toolkit, not a single model, so quality depends heavily on which pretrained model you pick. License: Apache-2.0 (code and models).

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
9.0/10
Maintenance · recency, releases, issue load
9.0/10
Openness · license + docs
10.0/10
Manual review · sanity read + red-flag check
7.0/10
  • · 9,000+ GitHub stars; v1.13.x (2026); 12 programming language bindings
  • · Runs offline on Android/iOS/HarmonyOS/Raspberry Pi/RISC-V/NPUs; WebSocket server/client for servers

Sources