sherpa_onnx

Supported functions

Speech recognition	Speech synthesis	Source separation
✔️	✔️	✔️

Speaker identification	Speaker diarization	Speaker verification
✔️	✔️	✔️

Spoken Language identification	Audio tagging	Voice activity detection
✔️	✔️	✔️

Keyword spotting	Add punctuation	Speech enhancement
✔️	✔️	✔️

Supported platforms

Architecture	Android	iOS	Windows	macOS	linux	HarmonyOS
x64	✔️		✔️	✔️	✔️	✔️
x86	✔️		✔️
arm64	✔️	✔️	✔️	✔️	✔️	✔️
arm32	✔️				✔️	✔️
riscv64					✔️

Supported programming languages

1. C++	2. C	3. Python	4. JavaScript
✔️	✔️	✔️	✔️

5. Java	6. C#	7. Kotlin	8. Swift
✔️	✔️	✔️	✔️

9. Go	10. Dart	11. Rust	12. Pascal
✔️	✔️	✔️	✔️

It also supports WebAssembly.

Supported frameworks

Flutter

Architecture	Android	iOS	Windows	macOS	Linux	Web
x64	✔️		✔️	✔️	✔️	✔️
x86	✔️					✔️
arm64	✔️	✔️	✔️	✔️	✔️	✔️
arm32	✔️					✔️

• Pre-built Flutter demo apps: flutter releases

Tauri

Architecture	Android	iOS	Windows	macOS	Linux
x64	✔️		✔️	✔️	✔️
x86	✔️
arm64	✔️	✔️	✔️	✔️	✔️
arm32	✔️

• Pre-built Tauri demo apps: tauri releases

Supported NPUs

1. Rockchip NPU (RKNN)	2. Qualcomm NPU (QNN)	3. Ascend NPU
✔️	✔️	✔️

4. Axera NPU	5. Intel NPU (OpenVINO)
✔️	✔️

Join our discord

Introduction

This repository supports running the following functions locally

• Speech-to-text (i.e., ASR); both streaming and non-streaming are supported
• Text-to-speech (i.e., TTS)
• Speaker diarization
• Speaker identification
• Speaker verification
• Spoken language identification
• Audio tagging
• VAD (e.g., silero-vad)
• Speech enhancement (e.g., gtcrn, DPDFNet)
• Keyword spotting
• Source separation (e.g., spleeter, UVR)

on the following platforms and operating systems:

• x86, x86_64, 32-bit ARM, 64-bit ARM (arm64, aarch64), RISC-V (riscv64), RK NPU, Ascend NPU
• Linux, macOS, Windows, openKylin
• Android, WearOS
• iOS
• HarmonyOS
• NodeJS
• WebAssembly
• NVIDIA Jetson Orin NX (Support running on both CPU and GPU)
• NVIDIA Jetson Nano B01 (Support running on both CPU and GPU)
• Raspberry Pi
• RV1126
• LicheePi4A
• VisionFive 2
• 旭日X3派
• 爱芯派
• RK3588
• SpacemiT-K1
• SpacemiT-K3
• etc

with the following APIs

• C++, C, Python, Go, C#
• Java, Kotlin, JavaScript
• Swift, Rust
• Dart, Object Pascal

Links for Huggingface Spaces

You can visit the following Huggingface spaces to try sherpa-onnx without installing anything. All you need is a browser.

Description	URL	中国镜像
Speaker diarization	Click me	镜像
Speech recognition	Click me	镜像
Speech recognition with Whisper	Click me	镜像
Speech synthesis	Click me	镜像
Generate subtitles	Click me	镜像
Audio tagging	Click me	镜像
Source separation	Click me	镜像
Spoken language identification with Whisper	Click me	镜像

We also have spaces built using WebAssembly. They are listed below:

Description	Huggingface space	ModelScope space
Voice activity detection with silero-vad	Click me	地址
Real-time speech recognition (Chinese + English) with Zipformer	Click me	地址
Real-time speech recognition (Chinese + English) with Paraformer	Click me	地址
Real-time speech recognition (Chinese + English + Cantonese) with Paraformer-large	Click me	地址
Real-time speech recognition (English)	Click me	地址
VAD + speech recognition (Chinese) with Zipformer CTC	Click me	地址
VAD + speech recognition (Chinese + English + Korean + Japanese + Cantonese) with SenseVoice	Click me	地址
VAD + speech recognition (English) with Whisper tiny.en	Click me	地址
VAD + speech recognition (English) with Moonshine tiny	Click me	地址
VAD + speech recognition (English) with Zipformer trained with GigaSpeech	Click me	地址
VAD + speech recognition (Chinese) with Zipformer trained with WenetSpeech	Click me	地址
VAD + speech recognition (Japanese) with Zipformer trained with ReazonSpeech	Click me	地址
VAD + speech recognition (Thai) with Zipformer trained with GigaSpeech2	Click me	地址
VAD + speech recognition (Chinese 多种方言) with a TeleSpeech-ASR CTC model	Click me	地址
VAD + speech recognition (English + Chinese, 及多种中文方言) with Paraformer-large	Click me	地址
VAD + speech recognition (English + Chinese, 及多种中文方言) with Paraformer-small	Click me	地址
VAD + speech recognition (多语种及多种中文方言) with Dolphin-base	Click me	地址
Speech synthesis (Piper, English)	Click me	地址
Speech synthesis (Piper, German)	Click me	地址
Speech synthesis (Matcha, Chinese)	Click me	地址
Speech synthesis (Matcha, English)	Click me	地址
Speech synthesis (Matcha, Chinese+English)	Click me	地址
Speaker diarization	Click me	地址
Voice cloning with ZipVoice (Chinese+English)	Click me	地址
Voice cloning with Pocket TTS (English)	Click me	地址

Links for pre-built Android APKs

You can find pre-built Android APKs for this repository in the following table

Description	URL	中国用户
Speaker diarization	Address	点此
Streaming speech recognition	Address	点此
Simulated-streaming speech recognition	Address	点此
Text-to-speech	Address	点此
Voice activity detection (VAD)	Address	点此
VAD + non-streaming speech recognition	Address	点此
Two-pass speech recognition	Address	点此
Audio tagging	Address	点此
Audio tagging (WearOS)	Address	点此
Speaker identification	Address	点此
Spoken language identification	Address	点此
Keyword spotting	Address	点此

Links for pre-built Flutter APPs

Real-time speech recognition

Description	URL	中国用户
Streaming speech recognition	Address	点此

Text-to-speech

Description	URL	中国用户
Android (arm64-v8a, armeabi-v7a, x86_64)	Address	点此
Linux (x64)	Address	点此
macOS (x64)	Address	点此
macOS (arm64)	Address	点此
Windows (x64)	Address	点此

Note: You need to build from source for iOS.

Links for pre-built Lazarus APPs

Generating subtitles

Description	URL	中国用户
Generate subtitles (生成字幕)	Address	点此

Links for pre-trained models

Description	URL
Speech recognition (speech to text, ASR)	Address
Text-to-speech (TTS)	Address
VAD	Address
Keyword spotting	Address
Audio tagging	Address
Speaker identification (Speaker ID)	Address
Spoken language identification (Language ID)	See multi-lingual Whisper ASR models from Speech recognition
Punctuation	Address
Speaker segmentation	Address
Speech enhancement	Address
Source separation	Address

Some pre-trained ASR models (Streaming)

Please see

• https://k2-fsa.github.io/sherpa/onnx/pretrained_models/online-transducer/index.html
• https://k2-fsa.github.io/sherpa/onnx/pretrained_models/online-paraformer/index.html
• https://k2-fsa.github.io/sherpa/onnx/pretrained_models/online-ctc/index.html

for more models. The following table lists only SOME of them.

Name	Supported Languages	Description
sherpa-onnx-streaming-zipformer-bilingual-zh-en-2023-02-20	Chinese, English	See also
sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16	Chinese, English	See also
sherpa-onnx-streaming-zipformer-zh-14M-2023-02-23	Chinese	Suitable for Cortex A7 CPU. See also
sherpa-onnx-streaming-zipformer-en-20M-2023-02-17	English	Suitable for Cortex A7 CPU. See also
sherpa-onnx-streaming-zipformer-korean-2024-06-16	Korean	See also
sherpa-onnx-streaming-zipformer-fr-2023-04-14	French	See also

Some pre-trained ASR models (Non-Streaming)

Please see

• https://k2-fsa.github.io/sherpa/onnx/pretrained_models/offline-transducer/index.html
• https://k2-fsa.github.io/sherpa/onnx/pretrained_models/offline-paraformer/index.html
• https://k2-fsa.github.io/sherpa/onnx/pretrained_models/offline-ctc/index.html
• https://k2-fsa.github.io/sherpa/onnx/pretrained_models/telespeech/index.html
• https://k2-fsa.github.io/sherpa/onnx/pretrained_models/whisper/index.html

for more models. The following table lists only SOME of them.

Name	Supported Languages	Description
sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8	English	It is converted from https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2
Whisper tiny.en	English	See also
Moonshine tiny	English	See also
sherpa-onnx-zipformer-ctc-zh-int8-2025-07-03	Chinese	A Zipformer CTC model
sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17	Chinese, Cantonese, English, Korean, Japanese	支持多种中文方言. See also
sherpa-onnx-paraformer-zh-2024-03-09	Chinese, English	也支持多种中文方言. See also
sherpa-onnx-zipformer-ja-reazonspeech-2024-08-01	Japanese	See also
sherpa-onnx-nemo-transducer-giga-am-russian-2024-10-24	Russian	See also
sherpa-onnx-nemo-ctc-giga-am-russian-2024-10-24	Russian	See also
sherpa-onnx-zipformer-ru-2024-09-18	Russian	See also
sherpa-onnx-zipformer-korean-2024-06-24	Korean	See also
sherpa-onnx-zipformer-thai-2024-06-20	Thai	See also
sherpa-onnx-telespeech-ctc-int8-zh-2024-06-04	Chinese	支持多种方言. See also

Useful links

• Documentation: https://k2-fsa.github.io/sherpa/onnx/
• Bilibili 演示视频: https://search.bilibili.com/all?keyword=%E6%96%B0%E4%B8%80%E4%BB%A3Kaldi

How to reach us

Please see https://k2-fsa.github.io/sherpa/social-groups.html for 新一代 Kaldi 微信交流群 and QQ 交流群.

Projects using sherpa-onnx

Sherpa Voice / @siteed/sherpa-onnx.rn

React Native wrapper and demo app for validating sherpa-onnx on iOS, Android, and Web, including ASR, TTS, VAD, KWS, speaker ID, diarization, language ID, punctuation, audio tagging, and speech enhancement.

• NPM package
• Live demo

Speed of Sound

A voice-typing application for the Linux desktop (GTK4/Adwaita). It captures microphone audio, transcribes it offline using Sherpa ONNX ASR models, optionally polishes the text with an LLM, and types the result into the active window via XDG Remote Desktop Portal keyboard simulation.

VoxSherpa TTS

VoxSherpa TTS is a 100% offline Android Text-to-Speech app powered by Sherpa-ONNX. It supports Kokoro-82M, Piper, and VITS engines with multilingual support including Hindi, English, British English, Japanese, Chinese and 50+ more languages.

• Download APK (All Versions)
• Android 11+ · 100% offline · No telemetry

---

BreezeApp from MediaTek Research

BreezeAPP is a mobile AI application developed for both Android and iOS platforms. Users can download it directly from the App Store and enjoy a variety of features offline, including speech-to-text, text-to-speech, text-based chatbot interactions, and image question-answering

• Download APK for BreezeAPP
• APK 中国镜像

1	2	3

Open-LLM-VTuber

Talk to any LLM with hands-free voice interaction, voice interruption, and Live2D taking face running locally across platforms

See also https://github.com/t41372/Open-LLM-VTuber/pull/50

voiceapi

Streaming ASR and TTS based on FastAPI

It shows how to use the ASR and TTS Python APIs with FastAPI.

腾讯会议摸鱼工具 TMSpeech

Uses streaming ASR in C# with graphical user interface.

Video demo in Chinese: 【开源】Windows实时字幕软件（网课/开会必备）

lol互动助手

It uses the JavaScript API of sherpa-onnx along with Electron

Video demo in Chinese: 爆了！炫神教你开打字挂！真正影响胜率的英雄联盟工具！英雄联盟的最后一块拼图！和游戏中的每个人无障碍沟通！

Sherpa-ONNX 语音识别服务器

A server based on nodejs providing Restful API for speech recognition.

QSmartAssistant

一个模块化，全过程可离线，低占用率的对话机器人/智能音箱

It uses QT. Both ASR and TTS are used.

Flutter-EasySpeechRecognition

It extends ./flutter-examples/streaming_asr by downloading models inside the app to reduce the size of the app.

Note: [Team B] Sherpa AI backend also uses sherpa-onnx in a Flutter APP.

sherpa-onnx-unity

sherpa-onnx in Unity. See also #1695, #1892, and #1859

xiaozhi-esp32-server

本项目为xiaozhi-esp32提供后端服务，帮助您快速搭建ESP32设备控制服务器 Backend service for xiaozhi-esp32, helps you quickly build an ESP32 device control server.

See also

• ASR新增轻量级sherpa-onnx-asr
• feat: ASR增加sherpa-onnx模型

KaithemAutomation

Pure Python, GUI-focused home automation/consumer grade SCADA.

It uses TTS from sherpa-onnx. See also ✨ Speak command that uses the new globally configured TTS model.

Open-XiaoAI KWS

Enable custom wake word for XiaoAi Speakers. 让小爱音箱支持自定义唤醒词。

Video demo in Chinese: 小爱同学启动～˶╹ꇴ╹˶！

C++ WebSocket ASR Server

It provides a WebSocket server based on C++ for ASR using sherpa-onnx.

Go WebSocket Server

It provides a WebSocket server based on the Go programming language for sherpa-onnx.

Making robot Paimon, Ep10 "The AI Part 1"

It is a YouTube video, showing how the author tried to use AI so he can have a conversation with Paimon.

It uses sherpa-onnx for speech-to-text and text-to-speech.

1

TtsReader - Desktop application

A desktop text-to-speech application built using Kotlin Multiplatform.

MentraOS

Smart glasses OS, with dozens of built-in apps. Users get AI assistant, notifications, translation, screen mirror, captions, and more. Devs get to write 1 app that runs on any pair of smart glasses.

It uses sherpa-onnx for real-time speech recognition on iOS and Android devices. See also https://github.com/Mentra-Community/MentraOS/pull/861

It uses Swift for iOS and Java for Android.

flet_sherpa_onnx

Flet ASR/STT component based on sherpa-onnx. Example a chat box agent

achatbot-go

a multimodal chatbot based on go with sherpa-onnx's speech lib api.

fcitx5-vinput

Local offline voice input plugin for Fcitx5 (Linux input method framework). It uses C++ with offline ASR for speech recognition, supporting push-to-talk, command mode, and optional LLM post-processing.

Video demo in Chinese: fcitx5-vinput

Wake Word

A VS Code extension for hands-free voice-activated coding. It uses sherpa-onnx for real-time keyword spotting (KWS) to detect custom wake phrases and trigger VS Code commands by voice. Audio capture is handled by decibri, a cross-platform Node.js microphone streaming library with prebuilt native binaries.

• VS Code Marketplace
• Open VSX
• decibri integration guides for sherpa-onnx

SmartSub

SmartSub is a local-first cross-platform desktop application for the complete subtitle production pipeline: audio/video transcription, subtitle translation, proofreading, and subtitle burning/muxing.

It natively integrates sherpa-onnx to power three offline ASR engines — FunASR, Qwen3-ASR, and FireRedASR — delivering high-accuracy Chinese and multilingual speech recognition entirely on-device, with no file uploads required.

Fono

Talk to your computer. Fono is an open-source (GPL-3.0), complete voice-AI stack (speech-to-text, text-to-speech, a local LLM, wake word, speaker ID) in one small binary for the Linux, macOS, and Windows desktop. Press a key and speak: it types into any app, answers as a voice assistant, or drives a coding agent.

It uses sherpa-onnx's Supertonic TTS models for text-to-speech, with the Supertonic text frontend ported to Rust. See also #3750, which fixes dropped diacritics in the Supertonic frontend for Romanian, Czech, Polish, Hungarian, Turkish, Vietnamese, and other non-English Latin-script languages.

Domia

Domia is an open-source local AI companion you talk to, living entirely on hardware you own. Each Domia is its own character, with a distinct personality, voice, emotional state, and memory of you and its place — and it acts, not just talks: skills connect it to the real world over MCP, Home Assistant included. Domia nodes form a local peer-to-peer mesh that shares speech and reasoning compute, while room satellites bring each companion into the spaces where you live without giving up its identity. No cloud, no accounts, and no data leaving your network.

Domia runs sherpa-onnx in-process for keyword spotting, voice activity detection, streaming and offline speech recognition, and text-to-speech. Speech engines and models are selected through per-companion configuration, making it easy to match supported sherpa-onnx models to different hardware.

qwen-audio-agent

A realtime full-duplex voice runtime for AI coding agents (Claude Code, Codex, OpenCode, and other ACP agents). It uses sherpa-onnx keyword spotting with the 3M-parameter zh-en transducer model to implement a local voice wake word: the app sleeps on idle but keeps the microphone open, and the wake phrase brings the session back with no cloud calls.