lattifai 1.3.1__py3-none-any.whl → 1.3.2__py3-none-any.whl

lattifai/alignment/lattice1_aligner.py

@@ -165,10 +165,15 @@ class Lattice1Aligner(object):
                 return_details=return_details,
                 start_margin=self.config.start_margin,
                 end_margin=self.config.end_margin,
-                check_sanity=True,
+                check_sanity=self.config.check_sanity,
             )
             if verbose:
                 safe_print(colorful.green(f"         ✓ Successfully aligned {len(alignments)} segments"))
+            if not self.config.check_sanity:
+                # Find and report low-score segments
+                low_score_segments = _find_low_score_segments(alignments)
+                if low_score_segments:
+                    safe_print(colorful.yellow(_format_low_score_warning(low_score_segments)))
         except LatticeDecodingError as e:
             safe_print(colorful.red("         x Failed to decode lattice alignment results"))
             _alignments = self.tokenizer.detokenize(
@@ -180,18 +185,17 @@ class Lattice1Aligner(object):
                 end_margin=self.config.end_margin,
                 check_sanity=False,
             )
-            # Check for score anomalies (media-text mismatch)
-            anomaly = _detect_score_anomalies(_alignments)
-            if anomaly:
-                anomaly_str = _format_anomaly_warning(anomaly)
-                del _alignments
+            # Find low-score segments to provide helpful error context
+            low_score_segments = _find_low_score_segments(_alignments)
+            del _alignments
+            if low_score_segments:
+                warning_str = _format_low_score_warning(low_score_segments)
                 raise LatticeDecodingError(
                     lattice_id,
-                    message=colorful.yellow("Score anomaly detected - media and text mismatch:\n" + anomaly_str),
-                    skip_help=True,  # anomaly info is more specific than default help
+                    message=colorful.yellow("Media-text mismatch detected:\n" + warning_str),
+                    skip_help=True,
                 )
             else:
-                del _alignments
                 raise e
         except Exception as e:
             safe_print(colorful.red("         x Failed to decode lattice alignment results"))
@@ -204,6 +208,12 @@ class Lattice1Aligner(object):
         self.worker.profile()
 
 
+def _is_event_segment(text: str) -> bool:
+    """Check if text is an event marker like [MUSIC], [Applause], [Writes equation]."""
+    text = text.strip()
+    return text.startswith("[") and text.endswith("]")
+
+
 def _detect_score_anomalies(
     alignments: List[Supervision],
     drop_threshold: float = 0.08,
@@ -215,6 +225,9 @@ def _detect_score_anomalies(
     When the drop is significant, it indicates the audio doesn't match
     the text starting at that position.
 
+    Event segments like [MUSIC], [Applause] are excluded from scoring as they
+    naturally have low alignment scores.
+
     Args:
         alignments: List of aligned supervisions with scores
         drop_threshold: Minimum drop between before/after averages to trigger
@@ -223,10 +236,16 @@ def _detect_score_anomalies(
     Returns:
         Dict with anomaly info if found, None otherwise
     """
-    scores = [s.score for s in alignments if s.score is not None]
-    if len(scores) < window_size * 2:
+    # Build (original_index, score) pairs, excluding events and None scores
+    indexed_scores = [
+        (i, s.score) for i, s in enumerate(alignments) if s.score is not None and not _is_event_segment(s.text)
+    ]
+    if len(indexed_scores) < window_size * 2:
         return None
 
+    scores = [score for _, score in indexed_scores]
+    orig_indices = [idx for idx, _ in indexed_scores]
+
     for i in range(window_size, len(scores) - window_size):
         before_avg = np.mean(scores[i - window_size : i])
         after_avg = np.mean(scores[i : i + window_size])
@@ -236,12 +255,15 @@ def _detect_score_anomalies(
         if drop > drop_threshold:
             # Find the exact mutation point (largest single-step drop)
             max_drop = 0
-            mutation_idx = i
+            filtered_mutation_idx = i
             for j in range(i - 1, min(i + window_size, len(scores) - 1)):
                 single_drop = scores[j] - scores[j + 1]
                 if single_drop > max_drop:
                     max_drop = single_drop
-                    mutation_idx = j + 1
+                    filtered_mutation_idx = j + 1
+
+            # Map back to original alignments index
+            mutation_idx = orig_indices[filtered_mutation_idx]
 
             # Segments: last normal + anomaly segments
             last_normal = alignments[mutation_idx - 1] if mutation_idx > 0 else None
@@ -290,3 +312,35 @@ def _format_anomaly_warning(anomaly: Dict[str, Any]) -> str:
     lines.append("")
     lines.append("    Possible causes: Transcription error, missing content, or wrong audio region")
     return "\n".join(lines)
+
+
+def _find_low_score_segments(
+    alignments: List[Supervision],
+    threshold: float = 0.7,
+) -> List[Tuple[int, Supervision]]:
+    """Find segments with scores below threshold, excluding event markers.
+
+    Args:
+        alignments: List of aligned supervisions with scores
+        threshold: Score threshold (segments below this are considered low)
+
+    Returns:
+        List of (index, supervision) tuples for low-score segments
+    """
+    return [
+        (i, s)
+        for i, s in enumerate(alignments)
+        if s.score is not None and s.score < threshold and not _is_event_segment(s.text)
+    ]
+
+
+def _format_low_score_warning(low_score_segments: List[Tuple[int, Supervision]]) -> str:
+    """Format low-score segments as warning message."""
+    lines = [
+        f"⚠️  Found {len(low_score_segments)} low-score segments (potential mismatches):",
+        "",
+    ]
+    for idx, seg in low_score_segments:
+        text_preview = seg.text[:50] + "..." if len(seg.text) > 50 else seg.text
+        lines.append(f'    #{idx} [{seg.start:.2f}s-{seg.end:.2f}s] score={seg.score:.4f} "{text_preview}"')
+    return "\n".join(lines)

lattifai/cli/__init__.py

@@ -1,5 +1,7 @@
 """CLI module for LattifAI with nemo_run entry points."""
 
+import lattifai._init  # noqa: F401 # isort: skip  # Suppress warnings early
+
 import nemo_run as run  # noqa: F401
 
 # Import and re-export entrypoints at package level so NeMo Run can find them

lattifai/config/alignment.py

@@ -21,7 +21,7 @@ class AlignmentConfig:
     model_name: str = "LattifAI/Lattice-1"
     """Model identifier or path to local model directory (e.g., 'LattifAI/Lattice-1')."""
 
-    model_hub: Literal["huggingface", "modelscope"] = "huggingface"
+    model_hub: Literal["huggingface", "modelscope"] = "modelscope"
     """Which model hub to use when resolving remote model names: 'huggingface' or 'modelscope'."""
 
     device: Literal["cpu", "cuda", "mps", "auto"] = "auto"
@@ -107,6 +107,13 @@ class AlignmentConfig:
     Default: 0.0 (no penalty). Typical range: -1.0 to 0.0 (e.g., -0.5).
     """
 
+    check_sanity: bool = True
+    """Whether to perform sanity checks on alignment results.
+    When True, raises an error if media and caption content don't match.
+    When False, outputs a warning instead of raising an error on mismatch.
+    Default: True.
+    """
+
     client_wrapper: Optional["SyncAPIClient"] = field(default=None, repr=False)
     """Reference to the SyncAPIClient instance. Auto-set during client initialization."""
 

{lattifai-1.3.1.dist-info → lattifai-1.3.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lattifai
-Version: 1.3.1
+Version: 1.3.2
 Summary: Lattifai Python SDK: Seamless Integration with Lattifai's Speech and Video AI Services
 Author-email: Lattifai Technologies <tech@lattifai.com>
 Maintainer-email: Lattice <tech@lattifai.com>
@@ -53,8 +53,8 @@ License-File: LICENSE
 Requires-Dist: python-dotenv
 Requires-Dist: colorful>=0.5.6
 Requires-Dist: lattifai-run>=1.0.1
-Requires-Dist: lattifai-core>=0.6.4
-Requires-Dist: lattifai-captions[splitting]>=0.1.6
+Requires-Dist: lattifai-core>=0.6.5
+Requires-Dist: lattifai-captions[splitting]>=0.1.7
 Requires-Dist: g2p-phonemizer>=0.4.0
 Requires-Dist: error-align-fix>=0.1.4
 Requires-Dist: lhotse>=1.26.0

{lattifai-1.3.1.dist-info → lattifai-1.3.2.dist-info}/RECORD

@@ -7,21 +7,21 @@ lattifai/mixin.py,sha256=_d-kzu2w5Y4mZTFMEumh30baWzKsQ9n_IS432woi5vQ,26295
 lattifai/types.py,sha256=JK7KVaZhX89BiKPm4okY0DWLHY1S8aj-YiZXoVH1akw,667
 lattifai/utils.py,sha256=5LeunAN0OQ1jWoKMIThpXSEOxFYD2dCRTdsglosodUU,7963
 lattifai/alignment/__init__.py,sha256=aOyC1P5DqESNLpDh6Gu6LyUZAVMba-IKI7Ugz7v9G4w,344
-lattifai/alignment/lattice1_aligner.py,sha256=MvBpPnhfF8NYtH2ANhQszKNRQROUiSYrBz3aN1bdT3U,11120
+lattifai/alignment/lattice1_aligner.py,sha256=2tLk215a3E8mvmQWNW-mgjEd2zE0myYRXXtAF_Epb9c,13255
 lattifai/alignment/lattice1_worker.py,sha256=Z7hxaS-nucNsUmrphbD8tgBBYPkJOgQb-85nFON94_I,13041
 lattifai/alignment/phonemizer.py,sha256=fbhN2DOl39lW4nQWKzyUUTMUabg7v61lB1kj8SKK-Sw,1761
 lattifai/alignment/punctuation.py,sha256=qLcvuXhBzoEa6bznWZiAB5TAxR6eLr_ZV-PnnCY90UA,1218
 lattifai/alignment/segmenter.py,sha256=JTbBYEXn8hkFwy0tITORy7nKoUPiNYLfi3w1DJNeHZ0,6303
 lattifai/alignment/text_align.py,sha256=sF-6Tsf863BhJcii3joeNa6Auv-7l3SiOhh9j8oPGME,14935
 lattifai/alignment/tokenizer.py,sha256=OIpMGHg1rJ7n97zncDMPpXy32uGOSt1yXiNO4sO6eP0,18839
-lattifai/cli/__init__.py,sha256=PdqoCTqRSFSWrqL3FjBTa5VzJy_e6Rq0OzyT7YkyHpc,541
+lattifai/cli/__init__.py,sha256=oNIgyYFGdG0gsVTrPULMwJxTchee-5h08tYiAPDqh_k,619
 lattifai/cli/alignment.py,sha256=rqg6wU2vf6RJ058yWVoXft_UJfOCrEpmE-ye5fhTphg,6129
 lattifai/cli/caption.py,sha256=jkMme73sJ16dkVpRh7O6qjbr14SUeBif00vCTBn7ed0,10339
 lattifai/cli/diarization.py,sha256=GTd2vnTm6cJN6Q3mFP-ShY9bZBl1_zKzWFu-4HHcMzk,4075
 lattifai/cli/transcribe.py,sha256=vZIV0TCbZG_IL2F_Mg49cCGSCBinOOFAtROajVTpNWE,7853
 lattifai/cli/youtube.py,sha256=FJwDl48-cuacP1sdPvX19vdszXdT7EoOZgGYzJpoLeM,6360
 lattifai/config/__init__.py,sha256=nJUVk03JRj4rujoEmkCkQ8akZF7kqIj7ci3XphU9uVA,1249
-lattifai/config/alignment.py,sha256=3JUtgHBueIK_lH9PgeBPjuHGL4VvDEYVs9fvylir6bc,5392
+lattifai/config/alignment.py,sha256=Bc7_5Sp-5pKVaQ90BRKva2xc9e3gck61JFtwtEplCWc,5659
 lattifai/config/caption.py,sha256=OMLsW8QKDWM6A3G5V3Gf-9bgB3D1PC5gO8LiiNNeOwM,7195
 lattifai/config/client.py,sha256=qqHKFPV4iEjVHCDOuGx7kj-tYFtgZZAszOQRFsNFbO8,2359
 lattifai/config/diarization.py,sha256=cIkwCfsYqfMns3i6tKWcwBBBkdnhhmB_Eo0TuOPCw9o,2484
@@ -49,9 +49,9 @@ lattifai/workflow/file_manager.py,sha256=yc29Vb7JNUMJ9rwM_YjkAHfDInl8HMVAl9A7z7X
 lattifai/youtube/__init__.py,sha256=_uO3KCx-t6I-JaYFpcYLYpvkbmEOOni3xBqGEbExg68,1587
 lattifai/youtube/client.py,sha256=VU8FC1N7YYpbc4LeJNAsahNAI1R7e3_7Yjmb1rz7tyI,52878
 lattifai/youtube/types.py,sha256=80RgBmvM4tRbxqyNv9GU6hr9vPp_yhKrK0RJ_vG2h4E,472
-lattifai-1.3.1.dist-info/licenses/LICENSE,sha256=xGMLmdFJy6Jkz3Hd0znyQLmcxC93FSZB5isKnEDMoQQ,1066
-lattifai-1.3.1.dist-info/METADATA,sha256=4AuEcvs82Jg7m_kjUcexm3XJ6Du2QCyqvWFAMaCivYA,23564
-lattifai-1.3.1.dist-info/WHEEL,sha256=SmOxYU7pzNKBqASvQJ7DjX3XGUF92lrGhMb3R6_iiqI,91
-lattifai-1.3.1.dist-info/entry_points.txt,sha256=MfoqXNjXrhD7VMApHgaHmAECTcGVUMUiR0uqnTg7Ads,502
-lattifai-1.3.1.dist-info/top_level.txt,sha256=tHSoXF26r-IGfbIP_JoYATqbmf14h5NrnNJGH4j5reI,9
-lattifai-1.3.1.dist-info/RECORD,,
+lattifai-1.3.2.dist-info/licenses/LICENSE,sha256=xGMLmdFJy6Jkz3Hd0znyQLmcxC93FSZB5isKnEDMoQQ,1066
+lattifai-1.3.2.dist-info/METADATA,sha256=kVJb9QV958OrxRytDeIDIb0vyu8xHKUjda4HH-cxlqU,23564
+lattifai-1.3.2.dist-info/WHEEL,sha256=SmOxYU7pzNKBqASvQJ7DjX3XGUF92lrGhMb3R6_iiqI,91
+lattifai-1.3.2.dist-info/entry_points.txt,sha256=MfoqXNjXrhD7VMApHgaHmAECTcGVUMUiR0uqnTg7Ads,502
+lattifai-1.3.2.dist-info/top_level.txt,sha256=tHSoXF26r-IGfbIP_JoYATqbmf14h5NrnNJGH4j5reI,9
+lattifai-1.3.2.dist-info/RECORD,,