Getting Started — Android

speech-android provides on-device speech processing for Android using ONNX Runtime. The pipeline runs VAD + STT + TTS with barge-in support, fully offline after model download.

Requirements

• Android 8+ (API 26)
• arm64-v8a architecture

Download the pre-built demo app to try it immediately:

app-release.apk

Gradle Dependency

Add the SDK to your build.gradle.kts:

implementation("audio.soniqo:speech:0.0.9")

Quick Start

val modelDir = ModelManager.ensureModels(context)
val pipeline = SpeechPipeline(SpeechConfig(modelDir = modelDir))
pipeline.events.collect { event ->
    when (event) {
        is SpeechEvent.TranscriptionCompleted -> println(event.text)
        is SpeechEvent.ResponseDone -> pipeline.resumeListening()
        else -> {}
    }
}
pipeline.start()
pipeline.pushAudio(samples) // 16kHz mono float32

System Voice Input (RecognitionService)

The SDK ships a ready-made SpeechRecognitionService that plugs into Android’s framework SpeechRecognizer API — no code to write. Once your app is selected as the default voice recognizer, any third-party app calling SpeechRecognizer.createSpeechRecognizer(context) (with no ComponentName) gets fully on-device STT through your pipeline.

1. Declare RECORD_AUDIO and the service in AndroidManifest.xml:

<uses-permission android:name="android.permission.RECORD_AUDIO" />

<application>
    <service
        android:name="audio.soniqo.speech.service.SpeechRecognitionService"
        android:exported="true"
        android:permission="android.permission.RECORD_AUDIO">
        <intent-filter>
            <action android:name="android.speech.RecognitionService" />
        </intent-filter>
        <meta-data
            android:name="android.speech"
            android:resource="@xml/recognition_service" />
    </service>
</application>

2. Add app/src/main/res/xml/recognition_service.xml:

<?xml version="1.0" encoding="utf-8"?>
<recognition-service xmlns:android="http://schemas.android.com/apk/res/android" />

3. Set the service as the system default (Settings → System → Languages & input → Voice input picker on stock Android, or via adb):

adb shell settings put secure voice_recognition_service \
  your.package/audio.soniqo.speech.service.SpeechRecognitionService

4. Verify by running the demo app’s Recognizer test screen, which calls SpeechRecognizer.createSpeechRecognizer(ctx) (no component) and logs every framework callback — useful for confirming the binder round-trip without needing logcat.

The service implements onCheckRecognitionSupport (API 33+) returning the 27 BCP-47 languages Parakeet TDT v3 covers, marked installedOnDeviceLanguage once models are present (or pendingOnDeviceLanguage while they’re downloading). Audio focus is acquired with AUDIOFOCUS_GAIN_TRANSIENT for the duration of a session.

Models

All models run via ONNX Runtime with NNAPI acceleration. INT8 quantized by default.

Source code: github.com/soniqo/speech-android

Next Steps

• Benchmarks — Android inference performance
• Linux C API — embedded Linux setup