# 核心训练目标：Conditional Flow Matching
# x_0: 干净梅尔频谱, x_1: 带噪/掩码频谱, t ∈ [0,1]
# v_θ(x_t, t, c) 预测速度场, c = 上下文条件 (已知音频片段 + 文本/说话人 embedding)

def cfm_loss(model, x_0, x_1, cond, t):
    x_t = (1 - t) * x_1 + t * x_0          # 线性插值路径 (Optimal Transport)
    v_target = x_0 - x_1                    # 真实速度场
    v_pred = model(x_t, t, cond)
    return F.mse_loss(v_pred, v_target)

┌─────────────┐   ┌──────────────┐   ┌─────────────┐   ┌──────────────┐
│ 网络抖动缓冲 │→  │ 丢包检测 &   │→  │ 生成式 PLC  │→  │ 交叉淡变 &   │
│ (JitterBuf) │   │ 掩码构建     │   │ 推理引擎    │   │ 后处理       │
└─────────────┘   └──────────────┘   └─────────────┘   └──────────────┘
      │                   │                   │                   │
  缓冲 600ms          标记丢包区间         1-step CFM         10ms 交叉淡变
  提供上下文          生成条件掩码         推理 < 15ms        平滑过渡

// 伪代码：自适应掩码扩展，避免「修复边界伪影」
struct PLCMaskBuilder {
    void build(const FrameBuffer& buf, float loss_rate) {
        // 1. 基础丢包标记
        mask = buf.get_loss_flags();  // [T], bool
        
        // 2. 连续丢包合并 + 保护带扩展
        // 前向扩展 20ms (防止起音模糊)，后向扩展 40ms (防止尾音截断)
        mask = dilate(mask, left=2, right=4);  // 10ms/frame
        
        // 3. 高丢包率下的「岛屿合并」：间隔 < 50ms 的非丢包帧纳入修复
        if (loss_rate > 0.3) merge_islands(mask, max_gap=5);
        
        // 4. 生成条件掩码：1=已知(条件), 0=待生成
        cond_mask = 1.0f - mask; 
    }
};

def plc_inference(audio_ctx, mask, net_stats):
    try:
        # 1. 尝试生成式修复
        if net_stats.rtt < 300 and net_stats.cpu_usage < 70%:
            return cfm_plc(audio_ctx, mask)  # < 20ms
    except (OOMError, TimeoutError) as e:
        log.warn(f"Generative PLC fallback: {e}")
    
    # 2. 降级到轻量 LPC 外推 (NETEQ 兼容)
    return neteq_plc_fallback(audio_ctx, mask)

# 重参数化训练时的 RepVGG Block → 推理时融合为单层 3x3 Conv
class RepConv(nn.Module):
    def __init__(self, c1, c2):
        super().__init__()
        self.rbr_dense = Conv(c1, c2, 3, padding=1)
        self.rbr_1x1 = Conv(c1, c2, 1)
        self.rbr_identity = nn.BatchNorm2d(c1) if c1 == c2 else None
    
    def forward(self, x):
        return self.rbr_dense(x) + self.rbr_1x1(x) + (self.rbr_identity(x) if self.rbr_identity else 0)
    
    def fuse(self):
        # 合并 BN 到 Conv 权重，部署时仅保留 fused_conv
        ...

网络层感知 → 丢包预测 (LSTM 预测未来 200ms 丢包概率)
     ↓
PLC 预热：提前对高风险区间执行「投机生成」，缓存备用
     ↓
实际丢包发生 → 直接拼接预生成片段 → **感知延迟 ≈ 0ms**

graph LR
    A[阶段一: 纯净语料<br/>LibriSpeech/VCTK/内部2000h] --> B[阶段二: 合成弱网<br/>丢包模式+编码器伪影]
    B --> C[阶段三: 真实硬例挖掘<br/>线上Badcase回流]
    C --> D[持续微调 Loop]

def simulate_codec_artifacts(clean_wav, codec='opus', bitrate=24kbps, plc_mode='default'):
    # 1. 编码 -> 强制丢包 -> 解码 (触发编码器内部 PLC)
    decoder = CodecDecoder(codec, bitrate)
    decoder.reset_state()  # 模拟会议中途加入/切换流
    
    packets = encoder.encode(clean_wav)
    corrupted = apply_loss_pattern(packets, loss_pattern)
    
    # 2. 关键：保留编码器 PLC 产生的「脏音频」作为模型输入
    #    而非简单的零填充/静音填充
    dirty_audio = decoder.decode(corrupted) 
    
    # 3. 目标：干净音频
    return dirty_audio, clean_wav

class UnifiedAudioRestoration(nn.Module):
    def __init__(self, dim=512, n_layers=12):
        super().__init__()
        # 共享骨干: Conformer/Transformer
        self.encoder = ConformerEncoder(dim, n_layers)
        
        # 轻量任务头 (2-layer Transformer Decoder)
        self.plc_head = PLCHead(dim)           # 预测速度场 v_t
        self.denoise_head = DenoiseHead(dim)   # 预测噪声掩码/谱
        self.dereverb_head = DereverbHead(dim) # 预测早期反射/晚期混响
        
        # 任务路由器: 根据输入条件动态激活
        self.task_router = nn.Linear(dim, 3)  # [plc, denoise, dereverb] logits
    
    def forward(self, x_t, t, cond, task_mask=None):
        # cond: [B, T, D] 包含上下文音频、噪声估计、RIR embedding
        feat = self.encoder(cond)
        
        # 动态任务融合 (推理时通常 PLC 为主，辅以去噪)
        if task_mask is None:
            task_weight = F.softmax(self.task_router(feat.mean(1)), dim=-1)
        else:
            task_weight = task_mask
            
        v_plc = self.plc_head(feat, t)
        n_denoise = self.denoise_head(feat)
        r_dereverb = self.dereverb_head(feat)
        
        # 加权融合速度场 (去噪/去混响视为修正项)
        v_final = v_plc + task_weight[:,1:2,None,None] * n_denoise + task_weight[:,2:3,None,None] * r_dereverb
        return v_final

[CPU: 预处理/掩码构建] 
    → (共享内存 fd) → 
[NPU: CFM DiT Block 推理 (INT4)] 
    → (共享内存 fd) → 
[DSP: Vocos 解码 + 交叉淡变 (向量化汇编优化)] 
    → (环形缓冲区) → 
[AudioTrack 播放]

# 伪代码：每日离线跑批 Pipeline
def daily_hardcase_mining(date_str):
    # 1. 从 ClickHouse 拉取低质量会话
    bad_sessions = ch.query(f"""
        SELECT session_id, uid, plc_segments 
        FROM plc_telemetry 
        WHERE dt='{date_str}' AND plc_pesq_est < 2.5
        SAMPLE 10000
    """)
    
    # 2. 切片对齐 & 去重 (Perceptual Hash)
    segments = extract_segments(bad_sessions, duration=4.0) # 含上下文
    unique_segments = dedup_perceptual_hash(segments, threshold=0.95)
    
    # 3. 多维自动标注
    labeled = []
    for seg in unique_segments:
        label = {
            'audio_dirty': seg.dirty_wav,
            'audio_clean': seg.clean_wav, # 端侧缓存的最近干净帧或服务端转码回源
            'loss_pattern': seg.loss_mask,
            'snr': estimate_snr(seg.dirty_wav),
            'reverb_rt60': estimate_rt60(seg.dirty_wav),
            'speaker_emb': extract_spk_emb(seg.context_wav),
            'lang_id': asr_lang_id(seg.context_wav),
            'codec': seg.codec_type,
            # 关键：教师模型生成伪标签目标
            'teacher_target': teacher_model.generate(seg.dirty_wav, seg.loss_mask) 
        }
        labeled.append(label)
    
    # 4. 写入训练数据湖 (Parquet + WebDataset)
    write_to_datalake(labeled, f"plc_hardcase/{date_str}")
    
    # 5. 触发训练流水线
    trigger_training_job(config="plc_ft_lora", data_path=f"plc_hardcase/{date_str}")

m=audio 49170 RTP/SAVPF 111 112
a=rtpmap:111 opus/48000/2
a=fmtp:111 minptime=10;useinbandfec=1;usedtx=1
; --- Generative PLC Negotiation ---
a=extmap:15 urn:ietf:params:rtp-hdr-ext:plc-mask  ; PLC 掩码扩展头
a=rtcp-fb:111 goog-remb
a=rtcp-fb:111 transport-cc
a=plc-capability:generative;model-id="cfm-v3.2-int4";max-gap=500ms;latency=15ms
a=plc-capability:neteq;version=3.14               ; 降级能力声明

# plc_prod_config.yaml
model:
  arch: "CFM-DiT-Tiny"
  dim: 384
  layers: 8
  heads: 6
  cond_dim: 512 # Context Encoder + Speaker Emb + Phoneme Posteriorgram
  solver: "Euler" # 1-step Consistency Distilled
  steps: 1
  guidance_scale: 1.0 # 无 CFG，蒸馏已烘焙

inference:
  chunk_size_ms: 600   # 上下文窗口
  lookahead_ms: 200    # 未来条件窗口 (若可用)
  overlap_ms: 10       # 交叉淡变
  max_gap_ms: 500      # 最大修复长度
  fallback_gap_ms: 300 # 超过此长度强制降级 NETEQ

deployment:
  precision: "INT4_ASYM" # 权重 INT4, 激活 INT8
  runtime: "ONNXRuntime 1.18+ / MLC-LLM / CoreML"
  thread_pool: 4
  memory_pool_mb: 48
  npu_delegate: true # 优先 NPU

monitoring:
  pesq_estimator: "DNS-MOS-Lite" # RTF < 0.01
  artifact_detector: "CNN-Small" # 2-layer Conv, 检测金属音/静音洞
  alert_webhook: "https://alert.internal/api/plc"