Whisper Android — Prebuilt on-device speech-to-text AAR

Run whisper.cpp on Android with one Gradle line. No NDK, no source build.

whisper-android is a prebuilt AAR that bundles whisper.cpp — a fast, on-device speech-to-text engine — behind a clean Kotlin API. Everything runs locally on the device: no network, no cloud, no API keys. 99 languages, optional translation to English.

You bring an audio file and a model file → you get text with timestamps.

---

Install

Pick one of the two methods.

A) Maven Central (recommended)

// build.gradle.kts (module)
dependencies {
    implementation("dev.ffmpegkit-maintained:whisper-android:0.1.2")
}

B) JitPack

// settings.gradle.kts
dependencyResolutionManagement {
    repositories {
        google()
        mavenCentral()
        maven { url = uri("https://jitpack.io") } // add this
    }
}

// build.gradle.kts (module)
dependencies {
    implementation("com.github.ffmpegkit-maintained.whisper:whisper-android:v0.1.2")
}

C) Direct AAR download

Grab whisper-android-<version>.aar from the Releases page, drop it in app/libs/, and add implementation(files("libs/whisper-android-0.1.2.aar")).

---

Quick Start

A complete, copy-paste example — even if you have never touched whisper.cpp or the NDK.

1. Add the dependency (see Install above).

2. Download a model (see Model Download) and ship it, or push it during dev:

adb push ggml-base.en.bin /sdcard/Android/data/<your.app.id>/files/models/

3. Transcribe:

import androidx.lifecycle.lifecycleScope
import dev.ffmpegkit.whisper.Whisper
import dev.ffmpegkit.whisper.WhisperConfig
import kotlinx.coroutines.launch
import java.io.File

class MainActivity : AppCompatActivity() {
    override fun onCreate(savedInstanceState: Bundle?) {
        super.onCreate(savedInstanceState)

        lifecycleScope.launch {
            // Audio can be WAV/MP3/FLAC at any sample rate (decoded + resampled internally).
            val modelPath = File(getExternalFilesDir("models"), "ggml-base.en.bin").absolutePath
            val audioPath = File(getExternalFilesDir(null), "speech.wav").absolutePath

            val model  = Whisper.loadModel(this@MainActivity, modelPath)
            val result = Whisper.transcribe(model, audioPath, WhisperConfig(language = "en"))

            Log.i("Whisper", "Text: ${result.text}")
            result.segments.forEach { s ->
                Log.i("Whisper", "[${s.startMs}–${s.endMs} ms] ${s.text}")
            }

            Whisper.releaseModel(model)
        }
    }
}

That's it. Whisper.transcribe is a suspend function — call it from a coroutine.

Audio input: WAV, MP3 or FLAC at any sample rate — the library decodes and resamples to 16 kHz mono automatically (via whisper.cpp's built-in miniaudio decoder, no FFmpeg). No manual conversion needed.

---

Model Download

The model is not bundled in the AAR (models are 75 MB – 1.5 GB — far too big). Download the one that fits your speed/quality/size budget from Hugging Face:

Model	Size	Speed	Quality	Languages	Download
tiny.en	~75 MB	⚡⚡⚡ fastest	★★	English only	ggml-tiny.en.bin
base	~142 MB	⚡⚡ fast	★★★	99 languages	ggml-base.bin
base.en	~142 MB	⚡⚡ fast	★★★	English only	ggml-base.en.bin
small	~466 MB	⚡ slower	★★★★	99 languages	ggml-small.bin

Which one? Start with base (or base.en for English-only) — the best speed/quality trade-off on a phone. Use tiny.en if you need real-time-ish speed on low-end devices, or small when accuracy matters more than latency.

Ship the model with your app (assets or a first-run download), then load it with Whisper.loadModel(context, path) or Whisper.loadModelFromAsset(context, "models/ggml-base.bin").

---

Compatibility

ABI	arm64-v8a (covers >90% of modern Android devices)
Android	API 24+ (Android 7.0 and up)
Android 15	✅ 16 KB page size aligned
NEON	✅ enabled
compileSdk / targetSdk	35

Need x86_64 (emulators, Chromebooks), real-time streaming, VAD, or quantized models? Those are in the Pro build — see jokobee.com.

---

Documentation

Full guides on the Wiki: Installation · Quick Start · Model Download · FAQ · Troubleshooting.

API at a glance

object Whisper {
    suspend fun loadModel(context: Context, modelPath: String): WhisperModel
    suspend fun loadModelFromAsset(context: Context, assetName: String): WhisperModel
    suspend fun transcribe(model: WhisperModel, audioPath: String, config: WhisperConfig = WhisperConfig()): WhisperResult
    fun releaseModel(model: WhisperModel)
    fun getSystemInfo(): String
}

---

Publisher

Jokobee · https://www.jokobee.com · contact@jokobee.com Maintained under the ffmpegkit-maintained organisation.

License

MIT — see LICENSE. whisper.cpp is also MIT (© Georgi Gerganov and contributors).