feat: 优化学习视频转写与字幕清洗

tests/test_hallucination_mask.py

@@ -33,6 +33,7 @@ from nodes.subtitle_cleanup import (
     HALLUCINATION_TOKENS,
     clean_japanese_hallucinations,
     clean_srt_text,
+    remove_short_moan_entries,
 )
 
 
@@ -163,3 +164,99 @@ def test_jp_middle_removal_resequences() -> None:
     assert out.count("-->") == 2
     assert out.startswith("1\n00:00:00,000 --> 00:00:02,000\nあ")
     assert "2\n00:00:41,000 --> 00:00:43,000\nい" in out
+
+
+# ---------------------------------------------------------------------------
+# 短呻吟/喘息碎片过滤（decode_full 救回弱语音的去噪）—— user 2026-09 决策
+# ---------------------------------------------------------------------------
+#
+# 背景实测（savr-1054-2 前 600s）：decode_full 无 VAD 解码会把呻吟/BGM 混叠
+# 的弱语音也救回来，但其中含大量**纯语气词碎片**（あ…/ん？/はぁ…/あ!あ!/んふふ
+# 等，有效假名 ≤3 个），这类内容放进字幕是噪声，影响观看。真实数据对照：
+#   - 要删（纯呻吟/喘息）：あ、ん、ん？、あ〜、あ…、はぁ…、あぁ…、あ!あ!、
+#     ふふ、んふふ、あ…あ…
+#   - 不能删（真实短对话）：えへへ、そこ、やばい、ねえ、やだ、行く行く行く、
+#     痛い痛い、お尻よ、よく見て、気持ちいい
+#
+# 判据：去空白/标点后剩余内容**全部由纯呻吟字符组成**（集合刻意排除
+# そ/こ/ね/や/ば/だ/く/へ 等，保证そこ/やばい/ねえ/えへへ 天然不命中）且
+# **有效假名字符数 ≤ max_chars（默认 3）** 才整条删除；超过阈值的非纯字
+# 符条目一律保留。仅 decode_full=true 时由 whisper 节点调用（用户决策）。
+
+
+def test_remove_short_moan_pure_moans_deleted() -> None:
+    """纯呻吟/喘息碎片（あ/ん/ん？/あ〜/はぁ…/あ!あ!/んふふ）整条删除。"""
+    srt = (
+        "1\n00:00:00,000 --> 00:00:02,000\n気持ちいい\n\n"
+        "2\n00:00:10,000 --> 00:00:11,000\nあ\n\n"              # 删
+        "3\n00:00:12,000 --> 00:00:13,000\nん？\n\n"            # 删
+        "4\n00:00:14,000 --> 00:00:15,000\nあ〜\n\n"            # 删
+        "5\n00:00:16,000 --> 00:00:17,000\nはぁ…\n\n"          # 删
+        "6\n00:00:18,000 --> 00:00:19,000\nあ!あ!\n\n"          # 删
+        "7\n00:00:20,000 --> 00:00:21,000\nんふふ\n\n"          # 删
+    )
+    out = remove_short_moan_entries(srt)
+    for frag in ("あ\n", "ん？", "あ〜", "はぁ", "あ!あ!", "んふふ"):
+        assert frag not in out
+    # 真实内容保留、序号重编号为 1。
+    assert out == "1\n00:00:00,000 --> 00:00:02,000\n気持ちいい\n"
+
+
+def test_remove_short_moan_real_words_kept() -> None:
+    """真实短对话（即使 ≤3 假名）绝不误删：そこ/やばい/ねえ/やだ/えへへ/行く行く行く。"""
+    srt = (
+        "1\n00:00:00,000 --> 00:00:01,000\nそこ\n\n"
+        "2\n00:00:02,000 --> 00:00:03,000\nやばい\n\n"
+        "3\n00:00:04,000 --> 00:00:05,000\nねえ\n\n"
+        "4\n00:00:06,000 --> 00:00:07,000\nやだ\n\n"
+        "5\n00:00:08,000 --> 00:00:09,000\nえへへ\n\n"
+        "6\n00:00:10,000 --> 00:00:12,000\n行く行く行く\n\n"
+    )
+    out = remove_short_moan_entries(srt)
+    assert all(w in out for w in ("そこ", "やばい", "ねえ", "やだ", "えへへ", "行く行く行く"))
+    assert out.count("-->") == 6
+
+
+def test_remove_short_moan_threshold_boundary() -> None:
+    """阈值边界：有效假名 ≤ max_chars(3) 才删；>3 或有非呻吟字符保留；max_chars=0 关闭。"""
+    # 4 个あ（>3）超出阈值 → 保留；3 个あ（=3）→ 删。
+    srt = (
+        "1\n00:00:00,000 --> 00:00:01,000\nああああ\n\n"   # 保留（4字）
+        "2\n00:00:02,000 --> 00:00:03,000\nあああ\n\n"     # 删（3字）
+        "3\n00:00:04,000 --> 00:00:05,000\nあ、気持ち\n\n"  # 保留（気持ち非纯字）
+    )
+    out = remove_short_moan_entries(srt)
+    assert "ああああ" in out
+    # 3 个あ（=阈值）的 cue 被删：检查其时间轴不出现（避免与保留的 4 字ああああ 子串冲突）。
+    assert "00:00:02,000 --> 00:00:03,000" not in out
+    assert "気持ち" in out
+    assert out.count("-->") == 2
+    # max_chars=0 关闭过滤：什么都不删。
+    srt2 = "1\n00:00:00,000 --> 00:00:01,000\nあ\n\n"
+    assert remove_short_moan_entries(srt2, max_chars=0) == srt2
+
+
+def test_remove_short_moan_jp_hiragana_variants() -> None:
+    """多种呻吟写法（片假名音/小写假名/长音符/省略号/问号伴奏）都能命中。"""
+    srt = (
+        "1\n00:00:00,000 --> 00:00:01,000\nア\n\n"          # 片假名あ
+        "2\n00:00:02,000 --> 00:00:03,000\nうぅ…\n\n"        # 小写ぅ
+        "3\n00:00:04,000 --> 00:00:05,000\nあぁ〜\n\n"        # 长音符
+        "4\n00:00:06,000 --> 00:00:07,000\nんー\n\n"         # 长音ー
+        "5\n00:00:08,000 --> 00:00:09,000\nあ あ\n\n"        # 带空格
+    )
+    out = remove_short_moan_entries(srt)
+    assert out.count("-->") == 0
+
+
+def test_remove_short_moan_multiline_and_resequence() -> None:
+    """多行文本条目与删除后重编号（序号连续、时间正确对应）。"""
+    srt = (
+        "1\n00:00:00,000 --> 00:00:02,000\n気持ちいい\nね\n\n"
+        "2\n00:00:10,000 --> 00:00:11,000\nん〜\n\n"   # 删
+        "3\n00:00:12,000 --> 00:00:15,000\nやばい\n\n"
+    )
+    out = remove_short_moan_entries(srt)
+    assert out.count("-->") == 2
+    assert out.startswith("1\n00:00:00,000 --> 00:00:02,000\n気持ちいい\nね")
+    assert "2\n00:00:12,000 --> 00:00:15,000\nやばい" in out

tests/test_nodes.py

@@ -292,13 +292,13 @@ def test_resolve_incomplete_candidate_skipped(tmp_path) -> None:
     empty = tmp_path / "empty"
     empty.mkdir()
     resolved = resolve_model_path({}, env={}, candidates=[empty])
-    assert resolved == "large-v3"
+    assert resolved == "large-v2"
 
 
 def test_resolve_fallback_remote() -> None:
-    """验证全部本地候选缺失时回退到远端 large-v3 作为最后兜底。"""
+    """验证全部本地候选缺失时回退到远端 large-v2 作为最后兜底。"""
     resolved = resolve_model_path({}, env={}, candidates=[])
-    assert resolved == "large-v3"
+    assert resolved == "large-v2"
 
 
 def test_format_timestamp() -> None:
@@ -1497,5 +1497,68 @@ def test_whisper_decode_full_cleanup_jp_hallucination(tmp_path, monkeypatch) ->
     assert all(t != "-" for t in text_lines)
 
 
+def test_whisper_decode_full_filters_short_moan(tmp_path, monkeypatch) -> None:
+    """decode_full=true 时纯呻吟碎片整条删除，真实短词保留（2026-09 决策）。"""
+    class MoanModel:
+        def __init__(self, *args, **kwargs):
+            pass
+
+        def transcribe(self, path, **kwargs):
+            # 混合：正常句 + 纯呻吟碎片 + 真实短词（都应保留）。
+            return (
+                [
+                    FakeSegment(0, 2, "気持ちいい"),   # 正常
+                    FakeSegment(2, 3, "あ〜"),         # 纯呻吟 → 删
+                    FakeSegment(3, 4, "ん？"),         # 纯呻吟 → 删
+                    FakeSegment(4, 5, "そこ"),         # 真实短词 → 保留
+                    FakeSegment(5, 6, "やばい"),       # 真实短词 → 保留
+                ],
+                None,
+            )
+
+    monkeypatch.setitem(
+        sys.modules, "faster_whisper", types.SimpleNamespace(WhisperModel=lambda *a, **k: MoanModel())
+    )
+    _make_wav(tmp_path / "audio.wav", 5)
+    resp = whisper_invoke(
+        _whisper_request(tmp_path, params={"language": "ja", "decode_full": True})
+    )
+    assert resp.status == "completed"
+    content = Path(resp.outputs["srt_uri"]).read_text(encoding="utf-8")
+    # 纯呻吟碎片被整条删除；真实短词与正常句保留。
+    assert "あ〜" not in content
+    assert "ん？" not in content
+    assert "気持ちいい" in content
+    assert "そこ" in content
+    assert "やばい" in content
+    assert content.count("-->") == 3  # 5 段 - 2 段呻吟 = 3 条
+
+
+def test_whisper_decode_full_short_moan_can_disable(tmp_path, monkeypatch) -> None:
+    """short_moan_max_chars=0 时关闭短呻吟过滤，所有内容原样保留。"""
+    class MoanModel2:
+        def __init__(self, *args, **kwargs):
+            pass
+
+        def transcribe(self, path, **kwargs):
+            return ([FakeSegment(0, 1, "あ〜"), FakeSegment(1, 2, "そこ")], None)
+
+    monkeypatch.setitem(
+        sys.modules, "faster_whisper", types.SimpleNamespace(WhisperModel=lambda *a, **k: MoanModel2())
+    )
+    _make_wav(tmp_path / "audio.wav", 5)
+    resp = whisper_invoke(
+        _whisper_request(
+            tmp_path,
+            params={"language": "ja", "decode_full": True, "short_moan_max_chars": 0},
+        )
+    )
+    assert resp.status == "completed"
+    content = Path(resp.outputs["srt_uri"]).read_text(encoding="utf-8")
+    assert "あ〜" in content
+    assert "そこ" in content
+    assert content.count("-->") == 2
+
+
 
 

tests/test_seed.py

@@ -28,7 +28,7 @@ def test_seed_default_workflows_idempotent(tmp_path) -> None:
     demo = db.get_latest_workflow_version("demo")["definition"]
     assert demo["name"] == "视频字幕生成"
     demo_asr = next(node for node in demo["nodes"] if node["id"] == "asr")
-    assert demo_asr["params"]["model_path"] == "faster-whisper-large-v3"
+    assert demo_asr["params"]["model_path"] == "faster-whisper-large-v2"
     assert demo_asr["params"]["condition_on_previous_text"] is False
 
     # zh-direct：使用中文直出模型并开启翻译任务。