DSH ASR Service

An Android app that provides system-level offline speech recognition.

Install it once, and any app that uses the standard android.speech.SpeechRecognizer API gets offline recognition — no cloud, no API key, audio never leaves the device.

中文说明见下方

---

Why

Android's speech recognition depends on a recognizer being installed on the device. Many phones — especially Chinese ROMs — ship without one, so voice input simply fails. This app fills that gap: it registers itself as a system RecognitionService, so other apps can use it through the standard API with zero code changes.

Features

• System-level RecognitionService — selectable as the device's recognizer
• Multiple models, download source follows the system language: HuggingFace official for non-Chinese, hf-mirror.com for Chinese (overridable in Settings)
• Streaming and offline decoding paths
• Bilingual (Chinese + English), plus multilingual models
• Built-in self-check — verifies service registration, model loading and the full recognition chain, and tells you exactly what to fix
• One-tap "set as system default" — via Shizuku (adb identity) or root; falls back to clear instructions when neither is available
• Minimal permissions — microphone and network only
• Dark UI, no ads, no tracking, no telemetry

Models

Models are hosted on Hugging Face; the app picks the source by system language (HuggingFace official for non-Chinese, the China-friendly hf-mirror.com mirror for Chinese; overridable in Settings). Sizes are the actual file sizes.

Model	Languages	Size	Streaming	Notes
zipformer-zh-14m-int8	Chinese	~25 MB	✅	Smallest, good starting point
zipformer-en-20m-int8	English	~30 MB	✅	Smallest English option
paraformer-zh-small-int8	Chinese	~78 MB	❌	Balanced accuracy/size
paraformer-zh-int8	Chinese	~230 MB	❌	Higher accuracy
sensevoice-small-int8	zh / en / ja / ko / yue	~230 MB	❌	Punctuation included
zipformer-bilingual-zh-en-int8	Chinese + English	~190 MB	✅	Best for mixed speech

Models are provided by sherpa-onnx (Apache-2.0).

Requirements

• Android 8.0 (API 26) or newer
• arm64-v8a device
• ~31 MB for the app, plus disk space for whichever models you download

Install

1. Install the APK.
2. Open the app → Models → download a model. Start with zipformer-zh-14m-int8 (small and fast); switch to a larger one later if accuracy matters more.
3. Make the system use it, any one of these:
   • Tap Set as system default in the app (needs Shizuku or root), or
   • System Settings → Languages & input → Voice input → select this app, or
   • Let the calling app specify this service directly (see below).
4. Tap Run check — it verifies everything and reports what to fix.

For app developers

The service is a standard RecognitionService. Two ways to use it:

// ① System default (nothing to do — just use the normal API)
val r = SpeechRecognizer.createSpeechRecognizer(context)

// ② Bind this service explicitly (works even when the ROM hides the
//    "voice input" settings page, which is common on Chinese ROMs)
val component = ComponentName("app.dsh.asr", "app.dsh.asr.AsrRecognitionService")
val r = SpeechRecognizer.createSpeechRecognizer(context, component)

For option ②, declare package visibility in your manifest (Android 11+):

<queries>
    <intent><action android:name="android.speech.RecognitionService" /></intent>
    <package android:name="app.dsh.asr" />
</queries>

The <intent> entry lets you enumerate recognizers; the <package> entry is required to bind a specific component — without it the bind silently fails.

Build

# Requires JDK 17 and Android SDK (compileSdk 36)
./gradlew assembleDebug -Pabi=arm64-v8a     # or x86_64 for emulators

The sherpa-onnx AAR is expected at app/libs/sherpa-onnx-1.13.8.aar (download from the sherpa-onnx releases).

Privacy

• Audio is processed entirely on device. Nothing is uploaded.
• The only network access is downloading models from the mirror.
• No analytics, no crash reporting, no advertising, no account required.
• Permissions: RECORD_AUDIO (obviously) and INTERNET (model downloads).

License

Apache License 2.0.

Bundled/used components:

• sherpa-onnx — Apache-2.0
• Shizuku — Apache-2.0
• Recognition models — see each model's repository for its license

---

中文说明

一个为 Android 提供系统级离线语音识别的应用。

装一次，任何使用标准 android.speech.SpeechRecognizer 接口的应用就获得了离线识别能力 —— 不联网、不要 API Key、录音不出设备。

为什么需要

Android 的语音识别依赖设备上已安装的识别服务。很多手机（尤其国内 ROM）出厂不带， 语音输入会直接失败。本应用补上这一环：把自己注册成系统 RecognitionService， 其它应用零改动即可通过标准接口使用。

特性

• 系统级识别服务 —— 可被选为设备的识别引擎
• 多个模型，下载源跟随系统语言：英文用 HuggingFace 官方源，中文用国内可直连的 hf-mirror.com（设置页可手动覆盖）
• 流式与离线两条解码路径
• 中英双语，另有多语言模型
• 内置自检 —— 检查服务注册、模型加载与完整识别链路，并明确告诉你该怎么修
• 一键设为系统默认 —— 通过 Shizuku（adb 身份）或 root；两者都没有时给出清晰指引
• 权限极简 —— 只要麦克风与网络
• 暗色界面，无广告、无追踪、无统计

模型

全部来自 hf-mirror.com（Hugging Face 的国内镜像）。体积为实际文件大小。

模型	语言	体积	流式	说明
zipformer-zh-14m-int8	中文	约 25 MB	✅	最小，建议先试这个
zipformer-en-20m-int8	英文	约 30 MB	✅	英文最小选项
paraformer-zh-small-int8	中文	约 78 MB	❌	精度/体积均衡
paraformer-zh-int8	中文	约 230 MB	❌	精度更高
sensevoice-small-int8	中英日韩粤	约 230 MB	❌	自带标点
zipformer-bilingual-zh-en-int8	中英	约 190 MB	✅	中英混说最佳

模型由 sherpa-onnx 提供（Apache-2.0）。

环境要求

• Android 8.0（API 26）及以上
• arm64-v8a 设备
• 应用约 31 MB，另需空间下载所选模型

安装使用

1. 安装 APK。安装后桌面不会出现图标 —— 这是一个纯系统服务插件， 入口在系统「语音输入 → 识别服务 → 设置」页，或由 DSH Mobile 等调用方直接拉起 （也可手动打开：adb shell am start -a app.dsh.asr.OPEN -p app.dsh.asr）。
2. 进入应用 → 识别模型 → 下载一个模型。建议先下 zipformer-zh-14m-int8 （小且快），之后若更看重准确率再换大的。
3. 让系统使用它，以下任一即可：
   • 在应用里点 设为系统默认（需 Shizuku 或 root），或
   • 系统设置 → 语言和输入法 → 语音输入 → 选中本应用，或
   • 让调用方应用直接指定本服务（见下）。
4. 点 开始检测 —— 它会验证全部环节并告诉你差什么。

给开发者

本服务是标准的 RecognitionService，两种用法：

// ① 系统默认（无需任何处理，正常调用即可）
val r = SpeechRecognizer.createSpeechRecognizer(context)

// ② 显式绑定本服务（国产 ROM 常把"语音输入"设置页藏起来时依然可用）
val component = ComponentName("app.dsh.asr", "app.dsh.asr.AsrRecognitionService")
val r = SpeechRecognizer.createSpeechRecognizer(context, component)

用第 ② 种方式时，需要在你的 manifest 里声明包可见性（Android 11+）：

<queries>
    <intent><action android:name="android.speech.RecognitionService" /></intent>
    <package android:name="app.dsh.asr" />
</queries>

<intent> 让你能枚举识别服务；<package> 才是绑定指定组件所必需的 —— 缺了它会静默绑定失败。

构建

# 需要 JDK 17 与 Android SDK（compileSdk 36）
./gradlew assembleDebug -Pabi=arm64-v8a     # 模拟器用 x86_64

sherpa-onnx 的 AAR 需放在 app/libs/sherpa-onnx-1.13.8.aar （从 sherpa-onnx releases 下载）。

隐私

• 音频全程在本机处理，不上传任何数据。
• 唯一的网络访问是下载模型。
• 无统计、无崩溃上报、无广告、无需账号。
• 权限只有 RECORD_AUDIO（麦克风）与 INTERNET（下模型）。

许可

Apache License 2.0。

使用/打包的组件：

• sherpa-onnx —— Apache-2.0
• Shizuku —— Apache-2.0
• 识别模型 —— 各自仓库的许可为准