Skip to main content

Android Offline STT SDK

SDK Version: 1.0.0

The STT SDK transcribes local WAV audio without a network connection. It prepares the bundled Vosk model, validates the audio stream, and returns text. Recording, permissions, transcoding, and UI remain host responsibilities.

Requirements​

  • WAV/RIFF container
  • 16-bit PCM, mono, 16 kHz
  • A URI that ContentResolver can open
  • A release package containing the Vosk/JNA native libraries and bundled models

Add the supplied STT SDK artifact and preserve its transitive/native dependencies.

APIs​

object XTalkStt {
  suspend fun prepare(context: Context, languageTag: String? = null)
  suspend fun transcribeWavUri(
    context: Context,
    wavUri: String,
    languageTag: String? = null,
  ): XTalkSttResult
}

data class XTalkSttResult(val text: String)
class XTalkSttException(val code: XTalkSttErrorCode, ...)

Tags beginning with zh select the bundled Chinese model. Other or null tags select the English model.

Example​

suspend fun transcribe(context: Context, uri: Uri): String =
  withContext(Dispatchers.IO) {
    XTalkStt.prepare(context, "en-US") // optional warm-up
    XTalkStt.transcribeWavUri(
      context = context,
      wavUri = uri.toString(),
      languageTag = "en-US",
    ).text
  }

Both calls are suspending. Warm-up is optional because transcription prepares the selected model automatically.

Errors​

XTalkSttErrorCodeMeaning and action
ENGINE_NATIVE_MISSINGVosk/JNA native library is absent or ABI-incompatible; check packaging.
MODEL_MISSINGThe selected model asset is missing; replace the release package.
MODEL_CORRUPTModel extraction/loading failed; clear the app model cache and retry once.
AUDIO_UNSUPPORTEDInput is not mono 16-bit 16 kHz PCM WAV; transcode before retrying.
TRANSCRIBE_FAILEDEngine or result parsing failed; retain the classified error and retry only when appropriate.
try {
  val text = XTalkStt.transcribeWavUri(context, uri.toString(), "en-US").text
} catch (e: XTalkSttException) {
  showSttError(e.code)
}

Do not send audio or model contents to logs. The SDK performs no cloud recognition.