:j( ddlZddlZddlmZddlmZmZm Z m Z ddl Z ddl m Z mZddlmZmZmZddlmZeeZGddZGd d Z dd ej2d eeefd eeeeefdedeeeff dZGddZ y)N)Pipeline)OptionalUnionListTuple) load_audio SAMPLE_RATE)TranscriptionResultAlignedTranscriptionResultProgressCallback) get_loggerc heZdZdZdeeeeeffdZdededeeeeffdZ dede efd Z y ) IntervalTreez Simple interval tree for fast overlap queries using sorted array + binary search. Uses O(n) space and provides O(log n) query time instead of O(n) linear scan. This achieves ~228x speedup for speaker assignment in long-form content. intervalsc|sz'IntervalTree.__init__..$s 1Q4keyr)dtype)nparraystartsendsspeakerssortedfloat64)selfrsorted_intervalsis r__init__zIntervalTree.__init__s((2,DK DI')DM ")@hh.>?!?rzzR HH,<=qad=RZZP '78!18  @=8s C C! C&startendreturnct|jdk(rgStj|j|d}|dk(rgSt d|}|j||k|j ||kDz}g}tj |dD][}t|j ||t|j||z }|dkDs<|j|j||f]|S)a) Find all intervals that overlap with [start, end] and compute intersection. Args: start: Query interval start time end: Query interval end time Returns: List of (speaker, intersection_duration) tuples for overlapping segments rleft)side) lenr r searchsortedslicer!whereminmaxappendr") r%r)r* right_idx candidatesoverlapsresultsidx intersections rqueryzIntervalTree.query)s t{{ q IOODKK6B >I1i( KK +c1dii 6Ke6ST88H%a( CCtyy~s3c$++c:JE6RRLa c 2LAB Crtimect|jdk(ry|j|jzdz }tjtj ||z }|j |S)z Find the speaker of the nearest segment to a given time point. Args: time: Time point to find nearest segment for Returns: Speaker ID of nearest segment, or None if no segments exist rNr)r/r r!rargminabsr")r%r=mids nearest_idxs r find_nearestzIntervalTree.find_nearestHs[ t{{ q  dii'1,iitd{ 34 }}[))rN) __name__ __module__ __qualname____doc__rrfloatstrr(r<rrCrrrrrsc9$uUE3->'?"@9&5ueCJ6G1H>**8C=*rrceZdZ d deeeejffdZ d deee jfdee dee dee de d e d eeej eeeeeffej ffd Zy)DiarizationPipelineNdevicect|trtj|}|xsd}tj d|t j|||j||_ y)Nz(pyannote/speaker-diarization-community-1zLoading diarization model: )token cache_dir) isinstancerItorchrLloggerinforfrom_pretrainedtomodel)r% model_namerNrLrO model_configs rr(zDiarizationPipeline.__init__\s^ fc "\\&)F!O%O  1,@A--l%S\]``agh raudio num_speakers min_speakers max_speakersreturn_embeddingsprogress_callbackr+ct|tr t|}tj|dddft d}d}ddddgdfd }|j |f|||d|d |ini} d | j} |r | jnd} tj| jd gd } | djd| d<| djd| d<|rC| At| jD cic]\} }|| | j}} }| |fS|r| dfS| Scc}} w)a Perform speaker diarization on audio. Args: audio: Path to audio file or audio array num_speakers: Exact number of speakers (if known) min_speakers: Minimum number of speakers to detect max_speakers: Maximum number of speakers to detect return_embeddings: Whether to return speaker embeddings progress_callback: Optional callable receiving a float (0-100) with progress percentage Returns: If return_embeddings is True: Tuple of (diarization dataframe, speaker embeddings dictionary) Otherwise: Just the diarization dataframe N)waveform sample_rate)I@)rcX@) segmentation embeddingsrbc|M|J|dkDrDj|d\}}|t||z d||z zz}| dkDr| d< |yyyyy)Nr)rbrdg?)getr3) step_name step_artifactfiletotal completedoffsetr*pct _STEP_RANGESlast_pctr^s rhookz*DiarizationPipeline.__call__..hooksx$)>519"."2"29k"JKFC 3y5'8##>#,#OOCXa[(&) )#.)DM)>$r)rZr[r\rrgY@T) yield_label)segmentlabelspeaker)columnsrtc|jSN)r)rs rrz.DiarizationPipeline.__call__..s AGGrr)c|jSry)r*rs rrz.DiarizationPipeline.__call__..s !%%rr*)NNN)rPrIrrQ from_numpyr rVspeaker_diarizationspeaker_embeddingspd DataFrame itertracksapply enumeratelabelstolist)r%rYrZr[r\r]r^ audio_datarroutput diarizationrf diarize_dfsrvr}rprqs ` @@r__call__zDiarizationPipeline.__call__is4 eS !u%E((tQw8&   ( !,*LuH /  %%%  "&!1~r    ( e $00 2CV.. \\+"8"8T"8"JTst (399:KL 7&y177H 5 !7T]^i^p^p^rTs!tja':a=+?+?+A"A!t !t11 1 t# # "usD?)NNcpuN)NNNFN)rDrErFrrrIrQrLr(rndarrayintboolr tupler~rdictlistrHrrrrrKrK[s5: isELL012 i '+&*&*"'.2MS"**_%MsmMsm M sm M  M,M uR\\8Dd5k1A,B#CCDbllR SMrrKrtranscript_resultr} fill_nearestr+c|jdg}|r|t|dk(r|S|jDcgc]\}}|d|d|df}}}t|}|D]N} | jdd} | jdd} |j | | } | rGi} | D]\}}| j|d|z| |<t | j dd| d<n"|r | | zd z }|j|}|r|| d<d | vs| d D]}d|vr|d}|jd|}|j ||}|rGi} |D]\}}| j|d|z| |<t | j d d|d<z|s}||zd z }|j|}|s||d<Q|||d <|Scc}}w) a Assign speakers to words and segments in the transcript. Uses an interval tree for O(log n) overlap queries instead of O(n) linear scan, achieving ~228x speedup for long-form content (3+ hour podcasts). Args: diarize_df: Diarization dataframe from DiarizationPipeline transcript_result: Transcription result to augment with speaker labels speaker_embeddings: Optional dictionary mapping speaker IDs to embedding vectors fill_nearest: If True, assign speakers even when there's no direct time overlap Returns: Updated transcript_result with speaker assignments and optionally embeddings segmentsrr)r*rvrbc |dSNrrrs rrz&assign_word_speakers..s aPQdrrrwordsc |dSrrrs rrz&assign_word_speakers..s WXYZW[rr})rhr/iterrowsrr<r4itemsrC)rrr}rtranscript_segments_rowrtreeseg seg_startseg_endr8speaker_intersectionsrvr;seg_midnearest_speakerword word_startword_end word_overlapsword_mids rassign_word_speakersrsF*,// B? *"4J18L  !))+ As Ws5z3y>2I  "D"(:GGGS) ''%%::i1 68 !)1 h%1F1J1J7TW1X[g1g%g. h !6!NSTUVC N  7*a/G"//8O!0I c>G  :$&!'] 88E:6 $ :x @  ,.)1>p-9N9R9RSZ\_9`co9o-g6p&)*?*E*E*G^&\]^&_DO! *X 5:H&*&7&7&AO&*9Y% :-(:V%2D./ isF=c(eZdZddededeefdZy)SegmentNr)r*rvc.||_||_||_yry)r)r*rv)r%r)r*rvs rr(zSegment.__init__ s  rry)rDrErFrrrIr(rrrrr s Sc8C=rr)NF)!numpyrpandasr~pyannote.audiortypingrrrrrQwhisperx.audiorr whisperx.schemar r r whisperx.log_utilsr rDrRrrKrrrIrrHrrrrrrrs#// 2]]) H J*J*Z[[B<@ N N79LLMN!c4;&6!78N N  %': :; Nbr