o 3NPi @sdZddlZddlZddlmZmZddlmZddlZ ddl m Z e dZ ddeed ed ed ed ef d dZdedeedeee jfd eefddZdeed eefddZdS)z9Speaker clustering and merging with temporal constraints.N)ListDict) defaultdict)cosine_similarityzFastPipelineV5.Clustering皙?segmentsspk1spk2 min_overlapreturnc slfdd|D}fdd|D}|D]\}}|D]\}} t||} t|| } | | |kr2dSqqdS)zv Cannot-link constraint: temporal overlap = different speakers. Two speakers can't talk at the same time. c(g|]}|dkr|d|dfqSspeakerstartend.0s)rrT/home/ubuntu/.cursor/worktrees/maya3data__SSH__216.81.248.184_/zxg/src/clustering.py (z,speakers_overlap_in_time..cr r rr)r rrrrTF)maxmin) rrr r spk1_times spk2_timess1_starts1_ends2_starts2_end overlap_start overlap_endr)rr rspeakers_overlap_in_times     r" audio_path embeddingsc Cstd|jdt}|s|Stt}t|D]\}}||vr0||d||||fqi}i} |D]\} } dd| D} | rUt j | dd|| <t | | | <q9t| } t | dkrotd t | d |St | }t ||f}t|D]"}t|D]}||krt|| |g|| |gd |||f<qq~d d | D}i}ddddd}g}t|D]}t|d|D]}|| || ||||ffqq|jdddd|D]K\}}}|dd7<||jkr|dd7<qt|||r |dd7<q|}||vr||}||vs|||<|||<|dd7<qtt|}dd t|D}|D]}||d|d}||||d<q>t|}t tdd|D}tdt | d|d|ddtd||S) a Conservative speaker merging with cannot-link constraints. Strategy: 1. Compute speaker centroids from embeddings 2. Build similarity matrix 3. Merge high-similarity speakers (respecting temporal constraints) 4. Assign final speaker labels u!🔧 Merging speakers (threshold=z)...rcSsg|]}|dqS)r)ritemrrrr?sz"merge_speakers..r)axisu ✅ Only z speaker(s), no merging needed)rrcSsi|]}||qSrrrrrr Usz"merge_speakers..) attemptedblocked_thresholdblocked_overlap successfulcS|dS)Nr%rxrrrcz merge_speakers..T)keyreverser*r+r,r-cSsi|] \}}|d|dqS)SPEAKER_02dr)rirrrrr)scss|]}|dVqdS)rNr)rsegrrr sz!merge_speakers..u ✅ Merged:  → z speakers (z.1fzs)z Stats: )loggerinfocluster_merge_thresholdtimerlist enumerateappenditemsnpmeanlenkeyszerosrangersortr"sortedsetvaluesget)r#rr$configrspeaker_segmentsr7r8speaker_embeddingsspeaker_countsrrBembsspeakersn sim_matrixj merge_map merged_tostatspairsrr simroot1unique_speakersname_mapmergedelapsedfinal_speakersrrrmerge_speakerss         "    $rbc Cs|sgSt|ddd}g}d}|D]\}|dur#|}|ddq|d|dk}|d|d }|dddk}|dddk}|r_|r_|r_||jkr_|d |d <|d |d|d <q|||}|ddq|rv||td t|d t|d |S)a\ Merge adjacent same-speaker segments with small gaps. Rules: 1. Same speaker 2. Gap <= max_silence_gap Important semantics: - The "gap" is typically a VAD-derived non-speech region (silence/breath). If we MERGE, we do NOT "ignore" that non-speech; we *include* it inside the merged [start, end] span by extending `end` to the later segment. - This is intentionally conservative by default: larger allowed gaps reduce fragmentation, but also increase the amount of silence included in a clip and can (rarely) hide missed micro-interjections if upstream diarization failed to label them as speech. Status handling (v6.8 fix): - Preserve any existing `status` (especially 'unusable') on incoming segments. - If a segment has no `status` key, default it to 'usable' so downstream code (stats, sample generation, UI) can reliably filter by status. - Only merge two segments when BOTH are usable. Unusable segments act as hard boundaries so we never "smear" low-quality/too-short regions into an otherwise clean clip. cSr.)Nrrr/rrrr1r2z)merge_adjacent_segments..)r3Nstatususablerrrdurationz Adjacent merge: r:z segments) rJcopy setdefaultrMmax_silence_gaprAr;r<rE) rrNr_currentr8 same_speakergapcurrent_usable seg_usablerrrmerge_adjacent_segmentss<      rn)r)__doc__r>loggingtypingrr collectionsrnumpyrCsklearn.metrics.pairwiser getLoggerr;strfloatboolr"intndarrayrbrnrrrrs&   $  o