6.2 拟人全双工语音智能体：VAD 打断与端到端流式架构

// full-duplex-audio-controller.ts
import { EventEmitter } from 'events';
 
export type AudioState = 'IDLE' | 'LISTENING' | 'AGENT_SPEAKING' | 'INTERRUPTED';
 
export class FullDuplexAudioController extends EventEmitter {
  private currentState: AudioState = 'IDLE';
  private audioPlaybackQueue: Buffer[] = [];
  private isMuted = false;
 
  constructor() {
    super();
    this.initMicrophoneStream();
  }
 
  private initMicrophoneStream() {
    // 模拟持续监听麦克风 VAD 事件
    console.log('[Audio Controller] 麦克风 VAD 全双工常驻监听就绪');
  }
 
  /**
   * 当底层 VAD 检测到人类发声时触发
   */
  handleUserVoiceDetected(amplitude: number, isSubstantiveSpeech: boolean) {
    if (this.currentState === 'AGENT_SPEAKING') {
      if (isSubstantiveSpeech) {
        // 触发实质性打断 (Barge-in)
        console.warn('⚡ [Barge-in] 检测到用户明确插话，立即熔断 Agent 播放！');
        this.interruptPlayback();
      } else {
        // 仅为 "嗯/啊" 的倾听音，忽略打断保持播放
        console.log('👂 [Backchannel] 捕获到用户倾听语气词，保持自然输出');
      }
    }
  }
 
  /**
   * 立即强行打断 Agent 的当前音频播放流
   */
  interruptPlayback() {
    this.currentState = 'INTERRUPTED';
    // 1. 物理清空声卡待播放缓冲区
    this.audioPlaybackQueue = [];
    this.isMuted = true;
    
    // 2. 通知后端大模型中断当前正在进行的自回归流式生成 (AbortController)
    this.emit('abort_generation');
    
    // 3. 瞬间切回倾听模式捕获用户新意图
    this.currentState = 'LISTENING';
    this.isMuted = false;
    this.emit('ready_for_new_input');
  }
 
  /**
   * 模拟接收并播放大模型推送过来的音频帧切片
   */
  enqueueAudioChunk(chunk: Buffer) {
    if (this.currentState === 'INTERRUPTED') {
      return; // 已被用户打断，直接丢弃后续残留音频切片
    }
    this.currentState = 'AGENT_SPEAKING';
    this.audioPlaybackQueue.push(chunk);
    // 驱动声卡 DMA 缓冲区播放...
  }
}