|
| 1 | +//! 说话人分离命令（REQ-153 / v0.7.2：弱化版讲者切换离线分析）。 |
| 2 | +//! |
| 3 | +//! @ai-context: analyze_session_speakers 懒加载——会话详情打开时调用： |
| 4 | +//! ① 幂等（已有 SpeakerChange 事件直接返回，不重复分析）； |
| 5 | +//! ② 模型缺失 → 空列表（前端提示"未启用"，诚实降级不报错）； |
| 6 | +//! ③ 段边界 = session_segments（端点断句产出 = 现成语音段， |
| 7 | +//! 零新增 VAD）；音频 = session-audio/<id>.wav（16k 单声道）。 |
| 8 | +//! @ai-context: 分析成本 = 段数 × embedding 推理（~10ms/段）——百段会话 |
| 9 | +//! <2s，可接受；单段失败跳过（不阻断整体，判定器对无效段 |
| 10 | +//! 重置前驱不误判）。结果落 session_events（SpeakerChange）， |
| 11 | +//! 容量守卫防写放大（低频事件，天然安全）。 |
| 12 | +
|
| 13 | +use tauri::State; |
| 14 | + |
| 15 | +use crate::commands::AppState; |
| 16 | +use crate::session_events::{EventKind, NewSessionEvent}; |
| 17 | +use crate::speaker_change::{detect_speaker_changes, SpeechSegment}; |
| 18 | +use crate::speaker_engine::{self, SpeakerEmbeddingEngine}; |
| 19 | + |
| 20 | +/// 讲者切换输出（camelCase 契约，前端展示用）。 |
| 21 | +#[derive(Debug, Clone, serde::Serialize)] |
| 22 | +#[serde(rename_all = "camelCase")] |
| 23 | +pub struct SpeakerChangeOut { |
| 24 | + /// 切换时刻（会话纪元 ms） |
| 25 | + pub time_ms: u64, |
| 26 | + /// 切换置信度 0.0-1.0（距余弦阈值越远越高） |
| 27 | + pub confidence: f32, |
| 28 | +} |
| 29 | + |
| 30 | +/// 讲者分析结果（前端区分"未启用"与"无切换"——诚实展示）。 |
| 31 | +#[derive(Debug, Clone, serde::Serialize)] |
| 32 | +#[serde(rename_all = "camelCase")] |
| 33 | +pub struct SpeakerAnalysisResult { |
| 34 | + /// 说话人分离是否启用（模型文件存在且分析已执行） |
| 35 | + pub enabled: bool, |
| 36 | + pub changes: Vec<SpeakerChangeOut>, |
| 37 | +} |
| 38 | + |
| 39 | +/// 会话讲者切换分析（弱化版说话人分离；幂等懒加载）。 |
| 40 | +#[tauri::command] |
| 41 | +pub fn analyze_session_speakers( |
| 42 | + state: State<'_, AppState>, |
| 43 | + session_id: i64, |
| 44 | +) -> Result<SpeakerAnalysisResult, String> { |
| 45 | + if session_id <= 0 { |
| 46 | + return Err("无效的会话 id".to_string()); |
| 47 | + } |
| 48 | + // 1) 幂等：已有分析结果（SpeakerChange 事件）→ 直接返回 |
| 49 | + let existing = state |
| 50 | + .db |
| 51 | + .list_events_by_kind(session_id, EventKind::SpeakerChange) |
| 52 | + .map_err(|e| e.to_string())?; |
| 53 | + if !existing.is_empty() { |
| 54 | + return Ok(SpeakerAnalysisResult { |
| 55 | + enabled: true, |
| 56 | + changes: existing |
| 57 | + .iter() |
| 58 | + .map(|e| SpeakerChangeOut { |
| 59 | + time_ms: e.timestamp_ms, |
| 60 | + confidence: e |
| 61 | + .payload |
| 62 | + .get("confidence") |
| 63 | + .and_then(|v| v.as_f64()) |
| 64 | + .map(|v| v as f32) |
| 65 | + .unwrap_or(0.0), |
| 66 | + }) |
| 67 | + .collect(), |
| 68 | + }); |
| 69 | + } |
| 70 | + // 2) 模型缺失 → 未启用（前端提示——诚实降级，不阻断） |
| 71 | + let Some(model_path) = speaker_engine::speaker_model_path(&state.model_dir) else { |
| 72 | + return Ok(SpeakerAnalysisResult { enabled: false, changes: Vec::new() }); |
| 73 | + }; |
| 74 | + // 3) 音频 + 段边界（端点断句产出 = 语音段） |
| 75 | + let wav_path = state |
| 76 | + .data_dir |
| 77 | + .join("session-audio") |
| 78 | + .join(format!("{}.wav", session_id)); |
| 79 | + let wave = sherpa_onnx::Wave::read(&wav_path.to_string_lossy()) |
| 80 | + .ok_or_else(|| "会话音频缺失（实时捕获未落盘或已清理）".to_string())?; |
| 81 | + let segments = state.db.list_segments(session_id).map_err(|e| e.to_string())?; |
| 82 | + if segments.is_empty() { |
| 83 | + return Ok(SpeakerAnalysisResult { enabled: true, changes: Vec::new() }); |
| 84 | + } |
| 85 | + // 4) 逐段提取音色向量（<MIN_SEGMENT_MS 或越界跳过——短段向量不稳） |
| 86 | + let sr = wave.sample_rate() as u32; |
| 87 | + let samples = wave.samples(); |
| 88 | + let mut engine = SpeakerEmbeddingEngine::load(&model_path) |
| 89 | + .ok_or_else(|| "说话人模型加载失败（模型文件损坏？请重新下载）".to_string())?; |
| 90 | + let mut speech: Vec<SpeechSegment> = Vec::new(); |
| 91 | + for seg in &segments { |
| 92 | + if seg.end_ms.saturating_sub(seg.start_ms) < speaker_engine::MIN_SEGMENT_MS { |
| 93 | + continue; |
| 94 | + } |
| 95 | + let Some((s, e)) = |
| 96 | + speaker_engine::segment_sample_range(seg.start_ms, seg.end_ms, sr, samples.len()) |
| 97 | + else { |
| 98 | + continue; |
| 99 | + }; |
| 100 | + let Some(emb) = engine.embedding_of(&samples[s..e], wave.sample_rate()) else { |
| 101 | + continue; |
| 102 | + }; |
| 103 | + speech.push(SpeechSegment { start_ms: seg.start_ms, embedding: emb }); |
| 104 | + } |
| 105 | + // 5) 相邻余弦判定（纯函数）→ 落库 → 返回 |
| 106 | + let changes = detect_speaker_changes(&speech); |
| 107 | + for c in &changes { |
| 108 | + let _ = state.db.add_event(&NewSessionEvent { |
| 109 | + session_id, |
| 110 | + kind: EventKind::SpeakerChange, |
| 111 | + timestamp_ms: c.time_ms, |
| 112 | + payload: serde_json::json!({ "confidence": c.confidence }), |
| 113 | + }); |
| 114 | + } |
| 115 | + Ok(SpeakerAnalysisResult { |
| 116 | + enabled: true, |
| 117 | + changes: changes |
| 118 | + .iter() |
| 119 | + .map(|c| SpeakerChangeOut { time_ms: c.time_ms, confidence: c.confidence }) |
| 120 | + .collect(), |
| 121 | + }) |
| 122 | +} |
0 commit comments