j# "SrSSKrSSKJrJr SSKJr SSKJr SSK r \R"\ 5r SrSrSr\"S S 55r\"S S 55rS \ R&S\4SjrSS\S\S-S\S\4SjjrS\\S\S\\4Sjrg)avSpeaker diarization using Resemblyzer embedding-based clustering. Assigns speaker labels to transcript segments by: 1. Encoding sliding windows of audio into speaker embedding vectors 2. Estimating the optimal number of speakers via silhouette score 3. Clustering embeddings into speaker groups 4. Mapping each transcript segment to the dominant speaker in its time window N) dataclassfield)Path)Anyg?c8\rSrSr%Sr\\S'\\S'\\S'Srg)SpeakerSegmentz.A time segment attributed to a single speaker. start_timeend_timespeakerN) __name__ __module__ __qualname____firstlineno____doc__float__annotations__str__static_attributes__r^/home/mande/repo/quber/.venv/lib/python3.13/site-packages/docling/utils/speaker_diarization.pyr r s8O Lrr c`\rSrSr%Sr\"\S9r\\\ S'Sr \ \ S'\"\S9r \\ \ S'Srg ) DiarizationResult'zOutput of speaker diarization.)default_factorysegmentsr num_speakers speaker_idsrN)rrrrrrlistrr rr intr!rrrrrrr's6(%*4%@Hd>"@L#"48Kc8rr embeddingsreturnc jSSKJn SSKJn [SpC[ [[ [S-[U555H[nU"US9RU5n[[U55S:aM2U"X5n[RSXW5 Xt:dMWUnUnM] [RS X45 U$) zEstimate optimal speaker count via silhouette score. Args: embeddings: Per-window speaker embeddings. Returns: The speaker count with the highest silhouette score. rAgglomerativeClustering)silhouette_scoreg n_clustersrzN=%d silhouette=%.4fz'Estimated %d speakers (silhouette=%.4f))sklearn.clusterr(sklearn.metricsr) _MIN_SPEAKERSrangemin _MAX_SPEAKERSlen fit_predictset_logdebuginfo)r$r(r)best_n best_scorenlabelsscores r_estimate_num_speakersr>0s80&J =#ma&7Z"I J(A6BB:N s6{ a   4 )14  JFK II7L Mrwav_pathr accelerator_devicecSSKnSSKnSSKJn SSKJnJnJn SSKJ n [RSU5 UR[!U5SS 9upU R"S :aU R%S S 9n X:waUR'XUS 9n U"XS S9n [)U 5S:Xa[RS5 [5$SSKJn U "U5nU"US9nUn[/US-5n[/[0U-5n/n/nSnUU-[)U 5::aDUR3UU- 5 UR3U UUU-5 UU- nUU-[)U 5::aMDU(d[RS5 [5$[RS[)U55 SSKnUR75nUR9[;SU55 [<R>"UVs/sHnURAU5PM sn5nUR9U5 UbUO [CU5nU "US9REU5nU5Vs/sH nSUS3PM nn/n[)U5S:aUUSnUSnUS[0-n [IUS SUS S5HEun!n"UU"n#U#U:Xa U![0-n MUR3[KUU U55 U#nU!nU![0-n MG UR3[KU[)U 5U- U55 [UUUS9$![a" [RS5 [5s$f=fs snf!UR9U5 f=fs snf)aRun speaker diarization on a WAV file. Loads and resamples the audio to the rate Resemblyzer expects, encodes fixed-size sliding windows into speaker embeddings, clusters the embeddings by speaker, then merges consecutive same-speaker windows into contiguous segments. Args: wav_path: Path to a 16kHz mono WAV file. num_speakers: Number of speakers. None = auto-detect. accelerator_device: Device selector passed to decide_device(), e.g. "auto", "cpu", "cuda", "cuda:0", "mps". Returns: Per-segment speaker labels. rN) VoiceEncoder)audio_norm_target_dBFSnormalize_volume sampling_rater'zSpeaker diarization requires resemblyzer, soundfile, scikit-learn, and librosa. Speaker diarization disabled. Install with: pip install 'docling-slim[format-video]'z!Loading audio for diarization: %sfloat32)dtyper*)axis)orig_sr target_srT) increase_onlyu$Empty audio — skipping diarization) decide_device)deviceg?zAudio too short for diarizationzEncoding %d audio windowsr+SPEAKER_02d)rr r!)&librosa soundfile resemblyzerrBresemblyzer.audiorCrDrEr-r( ImportErrorr6warningrr8readrndimmeanresampler3docling.utils.accelerator_utilsrLr# _WINDOW_STEPappendtorchget_num_threadsset_num_threadsr1nparrayembed_utterancer>r4r0zipr )$r?r r@rQsfrBrCrD_RESEMBLYZER_SRr(rawfile_srwavrLrMencodersrwindow_samples step_samples timestamps wav_splitsir^ prev_threadswr$r;r<r!r cur_speaker cur_startcur_endtslabelspks$ rdiarizeryJsB*#,  < II18<773x= 7:LC xx!|hhAh!sO 3d KC 3x1} ;< ""= - .F&)G Bc]N|b()L J#%J A n C (!b&!#a!n"456 \ n C (  67 ""II)3z?; ((*L #a./,XX:N:aw66q9:NO  l+$0 6LZ6XA $ 2 > >z JF/4Qx8x!XaW%xK8%'H :!&), qM Q-,.Z^VABZ8IBe$Ck!|+y'; OP!  |+9 y#c(R-MN  m #  E !"" #tO l+ 9s5LL8L30L8:M)L0/L03L88M transcript_items diarizationc rUR(dU$UHnUR=(d SnUR=(d UnSnSnURHMn[S[ XGR5[X7R5- 5nX:dM?UnUR nMO U(dMXRlM U$)a^Assign speaker labels to transcript ConversationItems. For each transcript segment, find the diarization segment with the maximum time overlap and assign its speaker label. Args: transcript_items: List of ConversationItem from ASR transcriber. diarization: Result from diarize(). Returns: The same list with .speaker set on each item. gN)rr r maxr1r) rzr{itemstartend best_speaker best_overlapsegoverlaps rassign_speakersrs    &3mm$u  ''C#s3 5E>>8RRSG%& "{{ ( <'L! r)Nauto)rlogging dataclassesrrpathlibrtypingrnumpyra getLoggerrr6r/r2r\r rndarrayr#r>rryr"rrrrrs(" ; ; |   99 9rzzc8 $$yy*yy yx#3i#"# #Y#r