o Si.$@sddlZddlmZddlmZddlmZmZmZddl Z ddl m Z ddl Z ddl m Z mZmZddlmZmZmZddlmZdd lmZdd lmZdd lmZmZmZmZGd d d eZdede dedefddZ!dS)N) defaultdict)partial)ListOptionalUnion)tqdm) RecordingSetSupervisionSet dill_enabled)CutSetMixedCutMixTrack)mix) parallel_map)uuid4)MAX_TASKS_WAITINGBaseMeetingSimulatorMeetingSamplerreverberate_cutscseZdZdZd'dedeffdd Zdd Zed d(d ee d d fddZ de e de e jd efddZed      d)de deedeedeee efdee edeedeed ed!ed e fd"d#Zde d$ed e fd%d&ZZS)*"SpeakerIndependentMeetingSimulatora This simulator uses the simulation method used in the end-to-end neural diarization (EEND) paper: https://arxiv.org/abs/1909.06247 (Algorithm 1). It samples segments of each speaker from the input CutSet, and concatenates them into speaker-specific channels, with pauses sampled from an exponential distribution. The speaker channels are then mixed together (possibly after adding room impulse responses) to create the simulated meeting. Since the speakers are simulated independently, the resulting mixtures can contain more overlap than is usually present in real meetings. In the paper, a single hyper-parameter `beta` is used which is equivalent to the scale parameter of the exponential distribution. Here, we use both `loc` and `scale`, where `loc` would mean the minimum silence duration between two consecutive utterances from the same speaker. These parameters can be either provided in initialization, or learned from a dataset using the `fit()` method. @locscalecst||_||_dS)aW :param loc: the minimum silence duration between two consecutive utterances from the same speaker. [Default: 0.0] :param scale: the scale parameter of the exponential distribution used to sample the silence duration between two consecutive utterances from a speaker. [Default: 2.0] N)super__init__rr)selfrr __class__k/home/ubuntu/.local/lib/python3.10/site-packages/lhotse/workflows/meeting_simulation/speaker_independent.pyr(s  z+SpeakerIndependentMeetingSimulator.__init__cCs|jjd|jd|jdS)Nz(loc=, scale=))r__name__rr)rrrr __repr__4sz+SpeakerIndependentMeetingSimulator.__repr__TNmeetingsreturnc Cs|durtd|jd|jdSt|tsJdddlm}ddlm }|dd t |d d d }g}| D]&}t |d d d }t d t |D]}|td||j||d jqLq;||\|_|_td|jdd|jddS)z Learn the distribution of the meeting parameters from a given dataset. :param meetings: a SupervisionSet containing the meetings to be used for Nz4No meetings provided, using default parameters: loc=r!z2The meetings must be provided as a SupervisionSet.r)groupby)exponcS |j|jfSN recording_idspeakersrrr L z8SpeakerIndependentMeetingSimulator.fit..cSr)r*r+r.rrr r0Mr1)keycSs|jSr*)startr.rrr r0SszLearned parameters: loc=z.2f)logginginforr isinstancer cytoolz.itertoolzr' scipy.statsr(sortedvaluesrangelenappendmaxr3endfitprint)rr%r'r(speaker_segmentsinter_speech_intervalssegmentsirrr rA7s4    z&SpeakerIndependentMeetingSimulator.fit utterancessilence_durationsc Csg}tt||D]A\}\}}t|}|d}t|dd|ddD]\}} t|| |j|dd}q$t|t||dkr?dn|dd}||q tt t |dS)a Create a MixedCut object from a list of speaker-wise MonoCuts and silence intervals. Each `track` in the resulting MixedCut represents a different speaker. Each `track` itself can be a MonoCut or a MixedCut (if the speaker has multiple utterances). rr4NT)offset allow_padding)cuttyperI)idtracks) enumerateziplistrdurationr rLr>r strr) rrGrHrNrFspk_utterances spk_silencestracksiluttrrr _create_mixture^s" z2SpeakerIndependentMeetingSimulator._create_mixture4@rr4cuts num_meetings num_repeatsnum_speakers_per_meetingspeaker_count_probsmax_duration_per_speakermax_utterances_per_speakerseednum_jobsc  Cs|dur |dur td|durd}t|tr|g}|dur)dt|gt|}t||||||||d} t| } tt||d} g} | dkrXtt | | |dD]}| |qOntt | | | | t d|d d D]}| |qgt | S) a Simulate the desired number of multi-speaker meetings. :param cuts: CutSet containing the MonoCut objects to be used for simulation. :param num_meetings: the number of meetings to simulate. [Default: None] :param num_repeats: the number of times to repeat the provided cuts. This means that the number of simulated meetings depends on how many cuts are available. :param num_speakers_per_meeting: the number of speakers per meeting. If a list is provided, the number of speakers per meeting is sampled from this list. [Default: 2] :param speaker_count_probs: the probability of each number of speakers per meeting. [Default: None] :param max_duration_per_speaker: the maximum duration of a speaker's utterances. [Default: 20.0] :param max_utterances_per_speaker: the maximum number of utterances per speaker. [Default: 5] :param seed: the random seed to be used for simulation. [Default: 0] :param num_jobs: the number of jobs to use for simulation. Use more jobs to speed up simulation when you have large number of source utterances. [Default: 1] Nz4Either num_meetings or num_repeats must be provided.g?)r_r^rbrcr`rard)rd simulatorr4)total)re queue_sizezSimulating meetings)rgdesc) ValueErrorr7intr=riterr_simulate_workerrmapr>rrr from_cuts)rr]r^r_r`rarbrcrdresampler sampler_iterworkmixturesmixturerrr simulatevsN!    z+SpeakerIndependentMeetingSimulator.simulaterirscGst|g|RSr*)r)rr]rvrrr reverberatesz.SpeakerIndependentMeetingSimulator.reverberate)rrr*)NNrZNr[r\rr4)r# __module__ __qualname____doc__floatrr$r rr rArr nparrayr rYrkrrurrw __classcell__rrrr rsX &     QrrGrdrfr&csttj|tt}D] }||jdj|q dd|Dfddt t D} |}|S)NrcSsg|]}t|qSr)r ro).0r]rrr sz$_simulate_worker..cs*g|]}jjjt|dqS))rsize)r exponentialrr=)rrFnprrfrGrr rs) r|random RandomStaterrQ supervisionsr-r>r;r<r=rY)rGrdrfutts_by_speakerrXrHrtrrr rms   rm)"r5 collectionsr functoolsrtypingrrrnumpyr|rlhotserr r lhotse.cutr r r lhotse.cut.setrlhotse.parallelr lhotse.utilsr(lhotse.workflows.meeting_simulation.baserrrrrrkrmrrrr s0      6