ailia Speech

The C API drives the following 3-state machine directly.

stateDiagram-v2
    direction LR
    WAITING_INPUT --> BUFFERED: pushInputData / finalizeInputData
    BUFFERED --> BUFFERED: transcribe — more chunks queued
    BUFFERED --> WAITING_INPUT: transcribe — buffer drained
    BUFFERED --> COMPLETE: transcribe — all done after finalize
    COMPLETE --> WAITING_INPUT: resetTranscribeState
        

Currently BUFFERED? ailiaSpeechBuffered() / Currently COMPLETE? ailiaSpeechComplete()

#include "ailia_speech.h"
#include "ailia_speech_util.h"

struct AILIASpeech *speech = nullptr;
ailiaSpeechCreate(&speech, env_id, AILIA_MULTITHREAD_AUTO, memory_mode,
                  AILIA_SPEECH_TASK_TRANSCRIBE, AILIA_SPEECH_FLAG_NONE,
                  ailiaSpeechUtilGetCallback(), AILIA_SPEECH_API_CALLBACK_VERSION);
ailiaSpeechOpenModelFileA(speech, encoder_path, decoder_path,
                          AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_SMALL);

ailiaSpeechPushInputData(speech, audio, channels, samples, sample_rate);
ailiaSpeechFinalizeInputData(speech);

// Drive transcribe in chunks until ailiaSpeechComplete signals done
unsigned int complete = 0;
while (complete == 0) {
    unsigned int buffered = 0;
    ailiaSpeechBuffered(speech, &buffered);
    if (buffered == 0) {
        ailiaSpeechComplete(speech, &complete);
        continue;
    }
    ailiaSpeechTranscribe(speech);

    unsigned int n;
    ailiaSpeechGetTextCount(speech, &n);
    for (unsigned int i = 0; i < n; i++) {
        AILIASpeechText text;
        ailiaSpeechGetText(speech, &text, AILIA_SPEECH_TEXT_VERSION, i);
        printf("%s\n", text.text);
    }
}

ailiaSpeechResetTranscribeState(speech);
ailiaSpeechDestroy(speech);

The C# wrapper's Transcribe runs asynchronously. After the final chunk (isFinal=true), poll IsCompleted() for completion.

stateDiagram-v2
    direction LR
    READY --> PROCESSING: Transcribe (async work begins)
    PROCESSING --> READY: drain complete (isFinal=false)
    PROCESSING --> COMPLETED: drain complete (isFinal=true)
    COMPLETED --> READY: ResetTranscribeState()
        

Currently PROCESSING? speech.IsProcessing() / Currently COMPLETED? speech.IsCompleted() / Read results: speech.GetResults() · speech.GetIntermediateText()

using ailiaSpeech;

var speech = new AiliaSpeechModel();
speech.Open(encoderPath, decoderPath, envId, memoryMode,
            AiliaSpeech.AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_SMALL,
            AiliaSpeech.AILIA_SPEECH_TASK_TRANSCRIBE,
            AiliaSpeech.AILIA_SPEECH_FLAG_NONE, "auto");

// Transcribe runs asynchronously. Pass isFinal=true on the last chunk.
speech.Transcribe(waveData, frequency, channels, /*isFinal=*/true);

// Poll until processing finishes — typically from Update() or a coroutine
while (!speech.IsCompleted()) { /* yield */ }

foreach (var text in speech.GetResults()) Debug.Log(text);

speech.ResetTranscribeState();
speech.Close();

The Flutter wrapper hides the internal state machine. Pass audio to transcribe() and the wrapper handles push / finalize / drain, returning the full result.

import 'package:ailia_speech/ailia_speech.dart';

final speech = AiliaSpeechModel();
await speech.open(
  encoderFile: encoderPath,
  decoderFile: decoderPath,
  modelType: AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_SMALL,
);
final segments = await speech.transcribe(audio, sampleRate);
for (final s in segments) print(s.text);

The JNI wrapper exposes the same 3-state machine as the C API.

stateDiagram-v2
    direction LR
    WAITING_INPUT --> BUFFERED: pushInputData / finalizeInputData
    BUFFERED --> BUFFERED: transcribe — more chunks queued
    BUFFERED --> WAITING_INPUT: transcribe — buffer drained
    BUFFERED --> COMPLETE: transcribe — all done after finalize
    COMPLETE --> WAITING_INPUT: resetTranscribeState
        

Currently BUFFERED? speech.getBuffered() / Currently COMPLETE? speech.getComplete()

val speech = AiliaSpeech(
    envId = -1,
    task = AiliaSpeech.AILIA_SPEECH_TASK_TRANSCRIBE,
    flags = AiliaSpeech.AILIA_SPEECH_FLAG_NONE,
)
speech.openModel(encoderPath, decoderPath,
    AiliaSpeech.AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_SMALL)
speech.pushInputData(audio, channels, samples, sampleRate)
speech.finalizeInputData()

// Drive transcribe in chunks until getComplete() signals done
while (speech.getComplete() == 0) {
    if (speech.getBuffered() == 0) continue
    speech.transcribe()
    for (i in 0 until speech.getTextCount()) {
        Log.d("ailia", speech.getText(i).text)
    }
}

speech.resetTranscribeState()
speech.close()