File size: 16,209 Bytes
de5da7a
 
228a065
de5da7a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6bf9bbb
de5da7a
6bf9bbb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
de0b3d5
6bf9bbb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
de5da7a
 
 
 
 
 
 
e441a1a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
de5da7a
e441a1a
de5da7a
 
 
 
 
 
 
 
 
e441a1a
 
 
 
 
 
 
 
de5da7a
 
 
 
 
 
 
 
 
 
 
 
 
 
e441a1a
 
 
 
 
de5da7a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e441a1a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
de5da7a
 
 
e441a1a
 
 
 
de5da7a
e441a1a
de5da7a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
"""
Diarisation Améliorée avec Clustering Adaptatif et Validation de Qualité
Vendored copy for importability from src/.
"""

import numpy as np
from sklearn.cluster import AgglomerativeClustering
from sklearn.metrics import silhouette_score
from typing import List, Dict, Tuple, Any
import logging

logger = logging.getLogger(__name__)

class ImprovedDiarization:
    """Diarisation améliorée avec clustering adaptatif et validation de qualité"""
    
    def __init__(self):
        self.min_speaker_duration = 3.0  # Durée minimum par locuteur (secondes)
        self.max_speakers = 10
        self.quality_threshold = 0.3  # Seuil de qualité minimum
        
    def adaptive_clustering(self, embeddings: np.ndarray) -> Tuple[int, float, np.ndarray]:
        """
        Détermine automatiquement le nombre optimal de locuteurs
        (version optimisée FAISS ; retombe sur sklearn si faiss absent)
        """
        try:
            import faiss
            HAS_FAISS = True
        except ImportError:
            HAS_FAISS = False

        if len(embeddings) < 2:
            return 1, 1.0, np.zeros(len(embeddings))

        if HAS_FAISS:
            return self._adaptive_faiss(embeddings)
        else:
            return self._adaptive_sklearn(embeddings)

    def _adaptive_faiss(self, embeddings: np.ndarray) -> Tuple[int, float, np.ndarray]:
        """Recherche du meilleur k via FAISS Kmeans (très rapide CPU)."""
        import faiss
        n_samples, dim = embeddings.shape
        best_score, best_k, best_labels = -1, 2, None
        max_k = min(8, max(2, n_samples // 10))  # Reduced for memory efficiency
        for k in range(2, max_k + 1):
            kmeans = faiss.Kmeans(dim, k, niter=20, verbose=False, seed=42)
            kmeans.train(embeddings.astype(np.float32))
            _, labels = kmeans.index.search(embeddings.astype(np.float32), 1)
            labels = labels.ravel()
            sil = silhouette_score(embeddings, labels) if len(set(labels)) > 1 else -1
            unique, counts = np.unique(labels, return_counts=True)
            balance = min(counts) / max(counts)
            adjusted = sil * (0.7 + 0.3 * balance)
            if adjusted > best_score:
                best_score, best_k, best_labels = adjusted, k, labels
        return best_k, best_score, best_labels

    def _adaptive_sklearn(self, embeddings: np.ndarray) -> Tuple[int, float, np.ndarray]:
        """Ancienne logique sklearn (conservée pour fallback)."""
        if len(embeddings) < 2:
            return 1, 1.0, np.zeros(len(embeddings))
        
        best_score = -1
        best_n_speakers = 2
        best_labels = None
        
        # Reduced configurations for faster processing on large datasets
        if len(embeddings) > 100:
            # For large datasets, use faster configurations only
            configurations = [
                ('euclidean', 'ward'),
                ('cosine', 'average'),
            ]
            max_test_speakers = min(6, len(embeddings) - 1)  # Limit search space
        else:
            # Full search for smaller datasets
            configurations = [
                ('euclidean', 'ward'),
                ('cosine', 'average'),
                ('cosine', 'complete'),
                ('euclidean', 'complete'),
            ]
            max_test_speakers = min(self.max_speakers, len(embeddings) - 1)
        
        for n_speakers in range(2, max_test_speakers + 1):
            for metric, linkage in configurations:
                try:
                    clustering = AgglomerativeClustering(
                        n_clusters=n_speakers,
                        metric=metric,
                        linkage=linkage
                    )
                    labels = clustering.fit_predict(embeddings)
                    
                    # Score de silhouette (with sampling for large datasets)
                    if len(embeddings) > 300:
                        # Sample for silhouette calculation to speed up
                        sample_size = min(300, len(embeddings))
                        indices = np.random.choice(len(embeddings), sample_size, replace=False)
                        score = silhouette_score(embeddings[indices], labels[indices], metric=metric)
                    else:
                        score = silhouette_score(embeddings, labels, metric=metric)
                    
                    # Bonus pour distribution équilibrée
                    unique, counts = np.unique(labels, return_counts=True)
                    balance_ratio = min(counts) / max(counts)
                    adjusted_score = score * (0.7 + 0.3 * balance_ratio)
                    
                    logger.debug(f"n_speakers={n_speakers}, metric={metric}, linkage={linkage}: "
                               f"score={score:.3f}, balance={balance_ratio:.3f}, "
                               f"adjusted={adjusted_score:.3f}")
                    
                    if adjusted_score > best_score:
                        best_score = adjusted_score
                        best_n_speakers = n_speakers
                        best_labels = labels.copy()
                    
                    # Early stopping for large datasets if score is decreasing
                    if len(embeddings) > 200 and n_speakers > 3 and adjusted_score < best_score * 0.9:
                        logger.debug(f"Early stopping at {n_speakers} speakers (score degrading)")
                        break
                        
                except Exception as e:
                    logger.warning(f"Clustering failed for n={n_speakers}, "
                                 f"metric={metric}, linkage={linkage}: {e}")
                    continue
        
        return best_n_speakers, best_score, best_labels
    
    def validate_clustering_quality(self, embeddings: np.ndarray, labels: np.ndarray) -> Dict[str, Any]:
        """Valide la qualité du clustering"""
        
        if len(np.unique(labels)) == 1:
            return {
                'silhouette_score': -1.0,
                'cluster_balance': 1.0,
                'quality': 'poor',
                'reason': 'single_cluster'
            }
        
        try:
            # Score de silhouette
            sil_score = silhouette_score(embeddings, labels)
            
            # Distribution des clusters
            unique, counts = np.unique(labels, return_counts=True)
            cluster_balance = min(counts) / max(counts)
            
            # Distance intra vs inter-cluster (optimized with vectorization)
            # Sample only 1000 pairs max for large datasets to avoid O(n²) complexity
            n_samples = len(embeddings)
            max_pairs = min(1000, (n_samples * (n_samples - 1)) // 2)
            
            if n_samples > 50:
                # Sample random pairs for large datasets
                np.random.seed(42)  # Reproducible sampling
                indices = np.random.choice(n_samples, size=min(100, n_samples), replace=False)
                sample_embeddings = embeddings[indices]
                sample_labels = labels[indices]
            else:
                sample_embeddings = embeddings
                sample_labels = labels
            
            # Vectorized distance calculation
            from scipy.spatial.distance import pdist, squareform
            distances = pdist(sample_embeddings, metric='euclidean')
            dist_matrix = squareform(distances)
            
            intra_distances = []
            inter_distances = []
            
            for i in range(len(sample_embeddings)):
                for j in range(i + 1, len(sample_embeddings)):
                    if sample_labels[i] == sample_labels[j]:
                        intra_distances.append(dist_matrix[i, j])
                    else:
                        inter_distances.append(dist_matrix[i, j])
            
            separation_ratio = np.mean(inter_distances) / np.mean(intra_distances) if intra_distances else 1.0
            
            # Évaluation globale
            quality = 'excellent' if sil_score > 0.7 and cluster_balance > 0.5 else \
                     'good' if sil_score > 0.5 and cluster_balance > 0.3 else \
                     'fair' if sil_score > 0.3 else 'poor'
            
            return {
                'silhouette_score': sil_score,
                'cluster_balance': cluster_balance,
                'separation_ratio': separation_ratio,
                'cluster_distribution': dict(zip(unique, counts)),
                'quality': quality,
                'reason': f"sil_score={sil_score:.3f}, balance={cluster_balance:.3f}"
            }
            
        except Exception as e:
            logger.error(f"Quality validation failed: {e}")
            return {
                'silhouette_score': -1.0,
                'cluster_balance': 0.0,
                'quality': 'error',
                'reason': str(e)
            }
    
    def refine_speaker_assignments(self, utterances: List[Dict], 
                                 min_duration: float = None) -> List[Dict]:
        """Affine les assignations de locuteurs"""
        
        if min_duration is None:
            min_duration = self.min_speaker_duration
        
        # Calcule la durée par locuteur
        speaker_durations = {}
        for utt in utterances:
            speaker = utt['speaker']
            duration = utt['end'] - utt['start']
            speaker_durations[speaker] = speaker_durations.get(speaker, 0) + duration
        
        logger.info(f"Speaker durations: {speaker_durations}")
        
        # Identifie les locuteurs avec durée insuffisante
        weak_speakers = {s for s, d in speaker_durations.items() if d < min_duration}
        
        if not weak_speakers:
            return utterances
        
        logger.info(f"Weak speakers to reassign: {weak_speakers}")
        
        # Réassigne les segments des locuteurs faibles
        refined_utterances = []
        for utt in utterances:
            if utt['speaker'] in weak_speakers:
                # Trouve le locuteur dominant adjacent
                new_speaker = self._find_dominant_adjacent_speaker(utt, utterances, weak_speakers)
                utt['speaker'] = new_speaker
                logger.debug(f"Reassigned segment [{utt['start']:.1f}-{utt['end']:.1f}s] "
                           f"to speaker {new_speaker}")
            
            refined_utterances.append(utt)
        
        return refined_utterances
    
    def _find_dominant_adjacent_speaker(self, target_utt: Dict, 
                                      all_utterances: List[Dict], 
                                      exclude_speakers: set) -> int:
        """Trouve le locuteur dominant adjacent pour réassignation"""
        
        # Trouve les segments adjacents
        target_start = target_utt['start']
        target_end = target_utt['end']
        
        candidates = []
        for utt in all_utterances:
            if utt['speaker'] in exclude_speakers:
                continue
            
            # Distance temporelle
            if utt['end'] <= target_start:
                distance = target_start - utt['end']
            elif utt['start'] >= target_end:
                distance = utt['start'] - target_end
            else:
                distance = 0  # Chevauchement
            
            candidates.append((utt['speaker'], distance))
        
        if not candidates:
            # Fallback: premier locuteur non exclu
            for utt in all_utterances:
                if utt['speaker'] not in exclude_speakers:
                    return utt['speaker']
            return 0  # Fallback ultime
        
        # Retourne le locuteur le plus proche
        return min(candidates, key=lambda x: x[1])[0]
    
    def merge_consecutive_same_speaker(self, utterances: List[Dict], 
                                     max_gap: float = 1.0) -> List[Dict]:
        """Fusionne les segments consécutifs du même locuteur"""
        
        if not utterances:
            return utterances
        
        merged = []
        current = utterances[0].copy()
        
        for next_utt in utterances[1:]:
            # Même locuteur et gap acceptable
            if (current['speaker'] == next_utt['speaker'] and 
                next_utt['start'] - current['end'] <= max_gap):
                
                # Fusionne les textes
                current['text'] = current['text'].strip() + ' ' + next_utt['text'].strip()
                current['end'] = next_utt['end']
                
                logger.debug(f"Merged segments: [{current['start']:.1f}-{current['end']:.1f}s] "
                           f"Speaker {current['speaker']}")
            else:
                # Finalise le segment actuel
                merged.append(current)
                current = next_utt.copy()
        
        # Ajoute le dernier segment
        merged.append(current)
        
        return merged
    
    def diarize_with_quality_control(self, embeddings: np.ndarray, 
                                   utterances: List[Dict]) -> Tuple[List[Dict], Dict[str, Any]]:
        """
        Diarisation complète avec contrôle qualité
        
        Returns:
            (utterances_with_speakers, quality_metrics)
        """
        
        if len(embeddings) < 2:
            # Cas trivial : un seul segment
            for utt in utterances:
                utt['speaker'] = 0
            return utterances, {'quality': 'trivial', 'n_speakers': 1}
        
        # Clustering adaptatif
        n_speakers, clustering_score, labels = self.adaptive_clustering(embeddings)
        
        # Validation de qualité
        quality_metrics = self.validate_clustering_quality(embeddings, labels)
        quality_metrics['n_speakers'] = n_speakers
        quality_metrics['clustering_score'] = clustering_score
        
        logger.info(f"Adaptive clustering: {n_speakers} speakers, "
                   f"score={clustering_score:.3f}, quality={quality_metrics['quality']}")
        
        # Applique les labels aux utterances
        for i, utt in enumerate(utterances):
            utt['speaker'] = int(labels[i])
        
        # Affinage des assignations
        if quality_metrics['quality'] not in ['error']:
            utterances = self.refine_speaker_assignments(utterances)
            utterances = self.merge_consecutive_same_speaker(utterances)
        
        return utterances, quality_metrics


def enhance_diarization_pipeline(embeddings: np.ndarray, 
                               utterances: List[Dict]) -> Tuple[List[Dict], Dict[str, Any]]:
    """
    Pipeline de diarisation amélioré - fonction principale
    
    Args:
        embeddings: Embeddings des segments audio (n_segments, 512)
        utterances: Liste des segments avec transcription
    
    Returns:
        (utterances_with_speakers, quality_report)
    """
    
    improved_diarizer = ImprovedDiarization()
    
    # Diarisation avec contrôle qualité
    diarized_utterances, quality_metrics = improved_diarizer.diarize_with_quality_control(
        embeddings, utterances
    )
    
    # Rapport de qualité détaillé
    quality_report = {
        'success': quality_metrics['quality'] not in ['error', 'poor'],
        'confidence': 'high' if quality_metrics['quality'] in ['excellent', 'good'] else 'low',
        'metrics': quality_metrics,
        'recommendations': []
    }
    
    # Recommandations basées sur la qualité
    if quality_metrics['quality'] == 'poor':
        quality_report['recommendations'].append(
            "Consider using single-speaker mode - clustering quality too low"
        )
    elif quality_metrics['silhouette_score'] < 0.3:
        quality_report['recommendations'].append(
            "Low speaker differentiation - verify audio quality"
        )
    elif quality_metrics['cluster_balance'] < 0.2:
        quality_report['recommendations'].append(
            "Unbalanced speaker distribution - check audio content"
        )
    
    return diarized_utterances, quality_report