agentSdk.registerAudio2Tts(new Audio2TtsCallback(agentSdk) {

    @Override
    public void onRobotOnline(String agentId, AgentMeta agentMeta) {
        // Robot comes online, initialize resources for this robot
        // agentMeta contains wakeup word, city, and other configuration
        String wakeupWord = agentMeta.getWakeupWord();
    }

    @Override
    public void onRobotOffline(String agentId) {
        // Robot goes offline, clean up resources for this robot
    }

    @Override
    public void onRequest(String agentId, String eventId, String itemId,
                          int flag, byte[] audio, AgentParam param,
                          Audio2TtsResponse response) {
        switch (flag) {
            case 0 -> {
                // VAD start: initialize ASR session
            }
            case 1 -> {
                // Audio frame: send to ASR service for streaming recognition
                sendToAsr(audio);
            }
            case 2 -> {
                // VAD end: finalize ASR, call LLM, synthesize TTS
                String asrText = finalizeAsr();
                response.onAsrFinal(eventId, asrText);

                // Call LLM to get reply
                String llmReply = callLlm(asrText);
                response.onLlmItemDelta(eventId, itemId, llmReply);
                response.onLlmItemDone(eventId, itemId);
                response.onLlmDone(eventId);

                // Synthesize TTS audio
                byte[] ttsAudio = synthesizeTts(llmReply);
                response.onTtsItemDelta(eventId, itemId, ttsAudio);
                response.onTtsItemDone(eventId, itemId);
                response.onTtsDone(eventId);
            }
        }
    }
});

onRequest(flag=0) -> Initialize
onRequest(flag=1) -> Accumulate audio -> ASR streaming recognition
                                       -> onAsrMiddle() (optional, intermediate result)
onRequest(flag=2) -> ASR complete -> onAsrFinal()
                                  -> LLM call -> onLlmItemDelta() x N
                                              -> onLlmItemDone()
                                              -> onLlmDone()
                                  -> TTS synthesis -> onTtsItemDelta() x N
                                                   -> onTtsItemDone()
                                                   -> onTtsDone()

onRequest(text="please walk forward 10 meters")
      │
      ▼
语义理解 classification
      ├─ rejected (invalid query / false wake / noise / ...)
      │    └→ just return — do NOT call any response.on* method
      │       (no skill, no interrupt, no LLM/VLM output)
      │
      └─ valid intent
           ① first MUST call response.onInterrupt(eventId, type, tips)
           │    (type ∈ risk_control / action / chat / realtime — see "Interrupts" below)
           │
           ② then fill in the result:
           ├─ skill hit → response.onSkill(eventId, itemId, ...)
           ├─ valid chat → response.onLlmItemDelta / onLlmItemDone / onLlmDone
           ├─ multimodal → response.onVlmItemDelta / onVlmItemDone / onVlmDone
           └─ TTS reply (when applicable) → response.onTtsItemDelta / onTtsItemDone / onTtsDone

@Override
public void onRequest(String agentId, String eventId, String text,
                      AgentParam param, Asr2LlmResponse response) {

    // 语义理解 on the business side: classify `text` into
    // "skill / chat / realtime / risk_control / rejected"
    NluResult nlu = classify(text);

    // ① Rejected: just return — do not call any response.on* method
    if (nlu.isRejected()) {
        return;
    }

    String itemId = IdGenerator.generateItemId();

    // ② Valid intent — MUST interrupt first, then fill in skill / LLM / VLM / TTS
    if (nlu.isRiskControl()) {
        // Risk control: interruptTips is required
        response.onInterrupt(eventId, "risk_control",
                "You mentioned a sensitive topic; let's talk about something else");
        return; // typically no further skill / LLM output after risk control
    } else if (nlu.isSkill()) {
        response.onInterrupt(eventId, "action", null);    // skill interrupt
    } else {
        response.onInterrupt(eventId, "chat", null);      // valid-chat interrupt
        // For realtime dialog: response.onInterrupt(eventId, "realtime", null);
    }

    // ③ Skill hit: dispatch skillParam (keys/values come from the Semantic Skills reference)
    if (nlu.isSkill()) {
        AgentParam skillParam = AgentParam.create()
            .setString("movement", "move_forward")
            .setInteger("meter", 10);
        response.onSkill(eventId, itemId, "movement", "move_forward", skillParam);
    }

    // ④ LLM streaming reply (runs alongside the skill — they do not block each other)
    response.onLlmItemDelta(eventId, itemId, "Sure, walking forward.");
    response.onLlmItemDone(eventId, itemId);
    response.onLlmDone(eventId);
}

// Plain chat reply
response.onInterrupt(eventId, "chat", null);
response.onLlmItemDelta(eventId, itemId, "Sure, here is the answer.");

// Action / skill hit
response.onInterrupt(eventId, "action", null);
AgentParam skillParam = AgentParam.create().setString("action", "wave_hands");
response.onSkill(eventId, itemId, "action", "wave_hands", skillParam);

// Realtime talk-while-listening
response.onInterrupt(eventId, "realtime", null);
response.onLlmItemDelta(eventId, itemId, "Mhm, I heard you...");

// Risk-control hit (interruptTips is mandatory, used as the spoken reply)
response.onInterrupt(eventId, "risk_control",
    "You mentioned a sensitive topic; let's talk about something else");

@Override
public void onGreetSignal(String agentId, String eventId, AgentParam param, GreetResponse response) {
    // Mode 1: streamed VLM text
    response.onGreetVlmDelta(eventId, "Hello");
    response.onGreetVlmDelta(eventId, ", nice weather today");
    response.onGreetVlmDone(eventId);

    // Mode 2: streamed TTS audio — argument is a base64-encoded chunk of synthesised audio bytes
    byte[] audioChunk1 = synthesizeTts("Hello");                  // PCM / Opus / ... bytes
    byte[] audioChunk2 = synthesizeTts(", welcome home");
    response.onGreetTtsDelta(eventId, Base64.getEncoder().encodeToString(audioChunk1));
    response.onGreetTtsDelta(eventId, Base64.getEncoder().encodeToString(audioChunk2));
    response.onGreetTtsDone(eventId);

    // Error path
    // response.onError(eventId, 5101, "greeting generation failed");
}