o piA2@sddlZddlmZmZmZmZmZmZddlZ ddl m Z ddl m Z mZmZmZddlmZddlmZddlmZmZd Zd Zd Zd Zd ZdZGdddeZdS)N)UnionIterableOptionalTupleListDict) ArrayLike)Segment AnnotationSlidingWindowFeatureTimeline) BaseMetric) det_curve)MetricComponentsDetailstargetspeaker_latency spk_scoreabsolute_latency abs_scorescorec s(eZdZdZedefddZdeeeffddZ  d'de e d e e ffd d Z d e fd dZdedeedeedefddZdeeefdeedeedefddZdeeefdeeeeeeffdefddZeddZeddZ    d(d!ed"ed#ed$efd%d&ZZS))LowLatencySpeakerSpottingaEvaluation of low-latency speaker spotting (LLSS) systems LLSS systems can be evaluated in two ways: with fixed or variable latency. * When latency is fixed a priori (default), only scores reported by the system within the requested latency range are considered. Varying the detection threshold has no impact on the actual latency of the system. It only impacts the detection performance. * In variable latency mode, the whole stream of scores is considered. Varying the detection threshold will impact both the detection performance and the detection latency. Each trial will result in the alarm being triggered with a different latency. In case the alarm is not triggered at all (missed detection), the latency is arbitrarily set to the value one would obtain if it were triggered at the end of the last target speech turn. The reported latency is the average latency over all target trials. Parameters ---------- latencies : float iterable, optional Switch to fixed latency mode, using provided `latencies`. Defaults to [1, 5, 10, 30, 60] (in seconds). thresholds : float iterable, optional Switch to variable latency mode, using provided detection `thresholds`. Defaults to fixed latency mode. returncCsdS)NzLow-latency speaker spotting)clsrrM/home/ubuntu/.local/lib/python3.10/site-packages/pyannote/metrics/spotting.py metric_nameLsz%LowLatencySpeakerSpotting.metric_namecCsddiS)Nrgr)selfrrrmetric_componentsPsz+LowLatencySpeakerSpotting.metric_componentsN thresholds latenciescsjt|dur|durgd}|dur|durtd|dur't||_|dur0t|}||_dS)N)r  <z3One must choose between fixed and variable latency.)super__init__ ValueErrornpsortr r!)rr r! __class__rrr'Ss    z"LowLatencySpeakerSpotting.__init__detailcCsdS)Nr)rr-rrrcompute_metricgsz(LowLatencySpeakerSpotting.compute_metric reference timestampsscoresc Csf|sd}t|tt|jdf}|}nd}tdd|D}g}g}tt||j|jD]c\}} z&||j||| } t|| } | dkrQt j j } n t|d| } Wnt yit|} Ynw| | |dj | } t|| } | dkrt j j } n t|d| } | | q1t|d}t|d}t|t|jt|t|jt|iS)NFr TcSsg|]}|jqSr)duration).0segmentrrr vsz.rr )r)maxoneslenr!cumsumzip searchsortedendsys float_info IndexErrorappendstartarrayreshapeSPOTTING_TARGETSPOTTING_SPK_LATENCYSPOTTING_SCORESPOTTING_ABS_LATENCYSPOTTING_ABS_SCORE) rr/r0r1 target_trialrrtotalilatencytup_tosrrr_fixed_latencyjsH        z(LowLatencySpeakerSpotting._fixed_latencyc sNtjtt|df}tjtt|df}t|tr#|jdd}|rI|dj}t|D]\}} t || } | |  ||<t d| j ||<q.tj |d} | |jktfddt|jD} |rd} tj|| d d }tj|| d d }d ddf}|j ||<| ||<nd} tj}tj}t| t|t|tt |iS) Nr F)copyrr6cs(g|]\}}tdd|fdqS)NT)r)r=)r3rM_ triggeredrrr5sz?LowLatencySpeakerSpotting._variable_latency..Tclipmoder7)r)nanr9r: isinstancer get_timelinerC enumerater cropr2r8maximum accumulaterEr rDtakeextentrFrIrGrH)rr/r0r1kwargsrr first_timerMrOso_farmaxcumindicesrKpositiverrUr_variable_latencys<      z+LowLatencySpeakerSpotting._variable_latency hypothesiscKsJt|tr dd|D}t|\}}|jdur||||S||||S)z Parameters ---------- reference : Timeline or Annotation hypothesis : SlidingWindowFeature or (time, score) iterable cSsg|] \}}|j|fqSr)r>)r3windowvaluerrrr5sz@LowLatencySpeakerSpotting.compute_components..N)r[r r<r!rirR)rr/rjrcr0r1rrrcompute_componentss   z,LowLatencySpeakerSpotting.compute_componentscCdd|D}tj|ddS)NcS g|] \}}|tr|tqSr)rFrIr3rTtrialrrrr5 z>LowLatencySpeakerSpotting.absolute_latency..raxisr)nanmeanrr!rrrrz*LowLatencySpeakerSpotting.absolute_latencycCrn)NcSror)rFrGrprrrr5rrz=LowLatencySpeakerSpotting.speaker_latency..rrsrurwrrrrrxz)LowLatencySpeakerSpotting.speaker_latencydr {Gz?F cost_misscost_fa prior_targetreturn_latencycCs|jdurtdd|D}tdd|D}t||dd\}}} } |ddd|ddd| ddd}}} |||||d|} |rtj| |jd d } tj| | d d } tj|| d d }tj|| d d }tj| | d d } | ||| | |j|jfS| ||| | fStd d|D}tdd|D} tdd|D}i}| |d D]S\}}i||<t |jD]E\}}t||dd|fdd\}}}} |ddd|ddd|ddd}}}|||||d|} |||| | f|||<qq|S)aDET curve Parameters ---------- cost_miss : float, optional Cost of missed detections. Defaults to 100. cost_fa : float, optional Cost of false alarms. Defaults to 1. prior_target : float, optional Target trial prior. Defaults to 0.5. return_latency : bool, optional Set to True to return latency. Has no effect when latencies are given at initialization time. Returns ------- thresholds : numpy array Detection thresholds fpr : numpy array False alarm rate fnr : numpy array False rejection rate eer : float Equal error rate cdet : numpy array Cdet cost function speaker_latency : numpy array absolute_latency : numpy array Speaker and absolute latency when return_latency is set to True. NcSg|]\}}|tqSrrFrprrrr5$z7LowLatencySpeakerSpotting.det_curve..cSrr)rHrprrrr5%rF) distancesr7g?left)siderWrXcSrrrrprrrr5<rcSrr)SPOTTING_SPK_SCORErprrrr5=rcSrr)rJrprrrr5>r)speakerabsolute) r!r)rDrr=r rarritemsr])rr{r|r}r~y_truer1fprfnrr eercdetrg spk_scores abs_scoresresultkeyrMrNthetarrrrsN $.   . z#LowLatencySpeakerSpotting.det_curve)NN)ryr rzF) __name__ __module__ __qualname____doc__ classmethodstrrrfloatrrrr'rr.r rrrRrr rir rrrmpropertyrrboolr __classcell__rrr+rr0sf : 7   r)r?typingrrrrrrnumpyr) numpy.typingr pyannote.corer r r r baserbinary_classificationrtypesrrrFrGrrIrJrHrrrrrs