o ei0@sdZddlZddlZddlZddlZddlZddlmZddlm Z ddlm Z ddl m Z ejdzddlZddlmZdd lmZdd lmZWn-eyyd Zed 7Zed 7Zed7Zed7Zed7Zed7Zed7Zed7ZeewddZddZddZddZddZddZd d!Zd"d#Z d$d%Z!d&d'Z"d(d)Z#d*d+Z$d,d-Z% . / /dAd0d1Z& .   2dBd3d4Z'Gd5d6d6eZ(Gd7d8d8Z)d9d:Z* Z+dCd?d@Z,dS)Du This script contains basic functions used for speaker diarization. This script has an optional dependency on open source scikit-learn (sklearn) library. A few scikit-learn functions are modified in this script as per requirement. Reference --------- This code is written using the following: - Von Luxburg, U. A tutorial on spectral clustering. Stat Comput 17, 395–416 (2007). https://doi.org/10.1007/s11222-007-9033-z - https://github.com/scikit-learn/scikit-learn/blob/0fb307bf3/sklearn/cluster/_spectral.py - https://github.com/tango4j/Auto-Tuning-Spectral-Clustering/blob/master/spectral_opt.py Authors * Nauman Dawalatabad 2020 N)sparse)connected_components) laplacian)eigsh)SpectralClustering)k_means)kneighbors_graphzFThe optional dependency scikit-learn (sklearn) is used in this module zCannot import scikit-learn. z%Please follow the below instructions z============================= z Using pip: zpip install scikit-learn z"================================ z Using conda: zconda install scikit-learncCsXg}t|ddd}|D] }|dd}||q Wd|S1s%wY|S)zReads and returns RTTM in list format. Arguments --------- rttm_file_path : str Path to the RTTM file to be read. Returns ------- rttm : list List containing rows of RTTM file. rutf-8encodingN)openappend)rttm_file_pathrttmflineentryr`/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/speechbrain/processing/diarization.py read_rttm4s    rc Cst|}ttdd|}g}d}t|dddR}|D]/}|d} | d} | |vrJ| tt||d g} || d| } | d | |d7}qd tt||d g} d| } | d | Wd d S1smwYd S) zWrite the final DERs for individual recording. Arguments --------- ref_rttm : str Reference RTTM file. DER : array Array containing DER values of each recording. out_der_file : str File to write the DERs. cSs |dS)Nz SPKR-INFO) startswithxrrrVs z!write_ders_file..rwr r  %s zOVERALL N) rlistfilterrsplitstrroundrjoinwrite) ref_rttmDER out_der_filer spkr_info rec_id_listcountrrowarec_idr line_strrrrwrite_ders_fileIs&     "r3c Cs|dg}g}|D]}|d|r||q ||}t|dddd}tj|ddtjd}|D]} || q1Wd d S1sDwYd S) aPrepares csv for a given recording ID. Arguments --------- full_diary_csv : csv Full csv containing all the recordings rec_id : str The recording ID for which csv has to be prepared out_csv_file : str Path of the output csv file. rrr )modenewliner ,") delimiter quotecharquotingN)rrrcsvwriter QUOTE_MINIMALwriterow) full_diary_csvr1 out_csv_file out_csv_headrr/out_csvcsv_file csv_writerr rrrprepare_subset_csvjs    "rFcCs||krdSdS)aReturns True if segments are overlapping. Arguments --------- end1 : float End time of the first segment. start2 : float Start time of the second segment. Returns ------- overlapped : bool True of segments overlapped else False. Example ------- >>> from speechbrain.processing import diarization as diar >>> diar.is_overlapped(5.5, 3.4) True >>> diar.is_overlapped(5.5, 6.4) False FTr)end1start2rrr is_overlappedsrIcCsg}|d}d}tdt|D]4}||}t|d|dr<|d|dkr<|d|d<|t|dkr;d}||q|||}q|durO||d|S)aMerge adjacent sub-segs from the same speaker. Arguments --------- lol : list of list Each list contains [rec_id, sseg_start, sseg_end, spkr_id]. Returns ------- new_lol : list of list new_lol contains adjacent segments merged from the same speaker ID. Example ------- >>> from speechbrain.processing import diarization as diar >>> lol=[['r1', 5.5, 7.0, 's1'], ... ['r1', 6.5, 9.0, 's1'], ... ['r1', 8.0, 11.0, 's1'], ... ['r1', 11.5, 13.0, 's2'], ... ['r1', 14.0, 15.0, 's2'], ... ['r1', 14.5, 15.0, 's1']] >>> diar.merge_ssegs_same_speaker(lol) [['r1', 5.5, 11.0, 's1'], ['r1', 11.5, 13.0, 's2'], ['r1', 14.0, 15.0, 's2'], ['r1', 14.5, 15.0, 's1']] rFrr TrrangelenrIr)lolnew_lolssegflagi next_ssegrrrmerge_ssegs_same_speakers "   rTcCsg}|d}tdt|D]^}||}t|d|drR|d|d}|d|d|d<|d|d|d<t|dkrD||n |d|krO|||}q t|dkr^||n |d|kri|||}q |||S)a Distributes the overlapped speech equally among the adjacent segments with different speakers. Arguments --------- lol : list of list It has each list structure as [rec_id, sseg_start, sseg_end, spkr_id]. Returns ------- new_lol : list of list It contains the overlapped part equally divided among the adjacent segments with different speaker IDs. Example ------- >>> from speechbrain.processing import diarization as diar >>> lol = [['r1', 5.5, 9.0, 's1'], ... ['r1', 8.0, 11.0, 's2'], ... ['r1', 11.5, 13.0, 's2'], ... ['r1', 12.0, 15.0, 's1']] >>> diar.distribute_overlap(lol) [['r1', 5.5, 8.5, 's1'], ['r1', 8.5, 11.0, 's2'], ['r1', 11.5, 12.5, 's2'], ['r1', 12.5, 15.0, 's1']] rrr g@rrK)rNrOrPrRrSoverlaprrrdistribute_overlaps(         rVc Csg}|dd}|D]'}d|dtt|ddtt|d|dddd|dddg }||q t|d d d }|D]}d |}|d |qrJrr r rr!N)r%r&rrr'r() segs_list out_rttm_filerr1segnew_rowrr/r2rrr write_rttm%s*   "r^c Cs|jd}t|r|}tj|td}tj|td}d||<t|D]B}|}tj |||d||kr;|St |d}| d|D]}t|rY|| } n||} tj || |dqIq$|S)a?Find the largest graph connected components that contains one given node. Arguments --------- graph : array-like, shape: (n_samples, n_samples) Adjacency matrix of the graph, non-zero weight means an edge between the nodes. node_id : int The index of the query node of the graph. Returns ------- connected_components_matrix : array-like shape - (n_samples,). An array of bool value indicating the indexes of the nodes belonging to the largest connected components of the given query node. r)dtypeT)outF)shaperissparsetocsrnpzerosboolrLsum logical_orwherefilltoarrayravel) graphnode_idn_nodeconnected_nodesnodes_to_explore_last_num_componentindicesrR neighborsrrr_graph_connected_componentJs*       rvcCs6t|rt|\}}|dkSt|d|jdkS)arReturn whether the graph is connected (True) or Not (False) Arguments --------- graph : array-like or sparse matrix, shape: (n_samples, n_samples) Adjacency matrix of the graph, non-zero weight means an edge between the nodes. Returns ------- is_connected : bool True means the graph is fully connected and False means not. rr)r isspmatrixrrvrgra)rmn_connected_componentsrrrrr_graph_is_connectedts rycCs|jd}t|s|r||jdd|d<|S|}|r)|j|jk}||j|<t |j|jj }|dkr=| }|S| }|S)aC Set the diagonal of the laplacian matrix and convert it to a sparse format well suited for eigenvalue decomposition. Arguments --------- laplacian : array or sparse matrix The graph laplacian. value : float The value of the diagonal. norm_laplacian : bool Whether the value of the diagonal should be changed or not. Returns ------- laplacian : array or sparse matrix An array of matrix in a form that is well suited to fast eigenvalue decomposition, depending on the bandwidth of the matrix. rNr) rarrwflattocoor/coldatarduniquesizetodiarc)rvaluenorm_laplaciann_nodesdiag_idxn_diagsrrr _set_diags    rcCsJtjt|dd}t|t|jd|f}||ddtjf9}|S)aModify the sign of vectors for reproducibility. Flips the sign of elements of all the vectors (rows of u) such that the absolute maximum element of each vector is positive. Arguments --------- u : ndarray Array with vectors as its rows. Returns ------- u_flipped : ndarray Array with the sign flipped vectors as its rows. The same shape as `u`. raxisrN)rdargmaxabssignrLranewaxis)u max_abs_rowssignsrrr_deterministic_vector_sign_flipsrcCsR|dus |tjurtjjjSt|tjrtj|St|tjjr#|Std|)aTurn seed into a np.random.RandomState instance. Arguments --------- seed : None | int | instance of RandomState If seed is None, return the RandomState singleton used by np.random. If seed is an int, return a new RandomState instance seeded with seed. If seed is already a RandomState instance, return it. Otherwise raise ValueError. Returns ------- np.random.RandomState Nz:%r cannot be used to seed a np.random.RandomState instance) rdrandommtrand_rand isinstancenumbersIntegral RandomState ValueError)seedrrr_check_random_states   rcCs"d}|D] }||vr|d7}q|S)au Returns actual number of speakers in a recording from the ground-truth. This can be used when the condition is oracle number of speakers. Arguments --------- rec_id : str Recording ID for which the number of speakers have to be obtained. spkr_info : list Header of the RTTM file. Starting with `SPKR-INFO`. Returns ------- num_spkrs : int Example ------- >>> from speechbrain.processing import diarization as diar >>> spkr_info = ['SPKR-INFO ES2011a 0 unknown ES2011a.A ', ... 'SPKR-INFO ES2011a 0 unknown ES2011a.B ', ... 'SPKR-INFO ES2011a 0 unknown ES2011a.C ', ... 'SPKR-INFO ES2011a 0 unknown ES2011a.D ', ... 'SPKR-INFO ES2011b 0 unknown ES2011b.A ', ... 'SPKR-INFO ES2011b 0 unknown ES2011b.B ', ... 'SPKR-INFO ES2011b 0 unknown ES2011b.C '] >>> diar.get_oracle_num_spkrs('ES2011a', spkr_info) 4 >>> diar.get_oracle_num_spkrs('ES2011b', spkr_info) 3 rrr)r1r, num_spkrsrrrrget_oracle_num_spkrss rTc Cs|r|d}t|stdt||dd\}}t|d|}|d9}t||ddd\}}|j|d d}|r:||}t|}|rG|d|jS|d |jS) aReturns spectral embeddings. Arguments --------- adjacency : array-like or sparse graph shape - (n_samples, n_samples) The adjacency matrix of the graph to embed. n_components : int The dimension of the projection subspace. norm_laplacian : bool If True, then compute normalized Laplacian. drop_first : bool Whether to drop the first eigenvector. Returns ------- embedding : array Spectral embeddings for each sample. Example ------- >>> import numpy as np >>> from speechbrain.processing import diarization as diar >>> affinity = np.array([[1, 1, 1, 0.5, 0, 0, 0, 0, 0, 0.5], ... [1, 1, 1, 0, 0, 0, 0, 0, 0, 0], ... [1, 1, 1, 0, 0, 0, 0, 0, 0, 0], ... [0.5, 0, 0, 1, 1, 1, 0, 0, 0, 0], ... [0, 0, 0, 1, 1, 1, 0, 0, 0, 0], ... [0, 0, 0, 1, 1, 1, 0, 0, 0, 0], ... [0, 0, 0, 0, 0, 0, 1, 1, 1, 1], ... [0, 0, 0, 0, 0, 0, 1, 1, 1, 1], ... [0, 0, 0, 0, 0, 0, 1, 1, 1, 1], ... [0.5, 0, 0, 0, 0, 0, 1, 1, 1, 1]]) >>> embs = diar.spectral_embedding_sb(affinity, 3) >>> # Notice similar embeddings >>> print(np.around(embs , decimals=3)) [[ 0.075 0.244 0.285] [ 0.083 0.356 -0.203] [ 0.083 0.356 -0.203] [ 0.26 -0.149 0.154] [ 0.29 -0.218 -0.11 ] [ 0.29 -0.218 -0.11 ] [-0.198 -0.084 -0.122] [-0.198 -0.084 -0.122] [-0.198 -0.084 -0.122] [-0.167 -0.044 0.316]] rzJGraph is not fully connected, spectral embedding may not work as expected.T)normed return_diagrg?LM)ksigmawhichN)rywarningswarncsgraph_laplacianrrTr) adjacency n_componentsr drop_firstrddvals diffusion_map embeddingrrrspectral_embedding_sbs06  r cCs@t|}|dur |n|}t||dd}t||||d\}}}|S)aPerforms spectral clustering. Arguments --------- affinity : matrix Affinity matrix. n_clusters : int Number of clusters for kmeans. n_components : int Number of components to retain while estimating spectral embeddings. random_state : int A pseudo random number generator used by kmeans. n_init : int Number of time the k-means algorithm will be run with different centroid seeds. Returns ------- labels : array Cluster label for each sample. Example ------- >>> import numpy as np >>> from speechbrain.processing import diarization as diar >>> affinity = np.array([[1, 1, 1, 0.5, 0, 0, 0, 0, 0, 0.5], ... [1, 1, 1, 0, 0, 0, 0, 0, 0, 0], ... [1, 1, 1, 0, 0, 0, 0, 0, 0, 0], ... [0.5, 0, 0, 1, 1, 1, 0, 0, 0, 0], ... [0, 0, 0, 1, 1, 1, 0, 0, 0, 0], ... [0, 0, 0, 1, 1, 1, 0, 0, 0, 0], ... [0, 0, 0, 0, 0, 0, 1, 1, 1, 1], ... [0, 0, 0, 0, 0, 0, 1, 1, 1, 1], ... [0, 0, 0, 0, 0, 0, 1, 1, 1, 1], ... [0.5, 0, 0, 0, 0, 0, 1, 1, 1, 1]]) >>> labs = diar.spectral_clustering_sb(affinity, 3) >>> # print (labs) # [2 2 2 1 1 1 0 0 0 0] NF)rr) random_staten_init)rrr)affinity n_clustersrrrmapsrrlabelsrrrspectral_clustering_sbks, rc@seZdZdZdddZdS) Spec_Clusterz9Performs spectral clustering using sklearn on embeddings.rcCs4t||dd}d||j|_t|j|jd|_|S)a Performs spectral clustering using sklearn on embeddings. Arguments --------- X : array (n_samples, n_features) Embeddings to be clustered. n_neighbors : int Number of neighbors in estimating affinity matrix. Returns ------- Spec_Cluster Reference --------- https://github.com/scikit-learn/scikit-learn/blob/0fb307bf3/sklearn/cluster/_spectral.py T) n_neighbors include_self?)r)r raffinity_matrix_rrlabels_)selfXr connectivityrrr perform_scszSpec_Cluster.perform_scN)r)__name__ __module__ __qualname____doc__rrrrrrsrc@sTeZdZdZdddZddZdd Zd d Zd d ZdddZ ddZ ddZ dS)Spec_Clust_unormu This class implements the spectral clustering with unnormalized affinity matrix. Useful when affinity matrix is based on cosine similarities. Arguments --------- min_num_spkrs : int Minimum number of expected speakers. max_num_spkrs : int Maximum number of expected speakers. Reference --------- Von Luxburg, U. A tutorial on spectral clustering. Stat Comput 17, 395–416 (2007). https://doi.org/10.1007/s11222-007-9033-z Example ------- >>> from speechbrain.processing import diarization as diar >>> clust = diar.Spec_Clust_unorm(min_num_spkrs=2, max_num_spkrs=10) >>> emb = [[ 2.1, 3.1, 4.1, 4.2, 3.1], ... [ 2.2, 3.1, 4.2, 4.2, 3.2], ... [ 2.0, 3.0, 4.0, 4.1, 3.0], ... [ 8.0, 7.0, 7.0, 8.1, 9.0], ... [ 8.1, 7.1, 7.2, 8.1, 9.2], ... [ 8.3, 7.4, 7.0, 8.4, 9.0], ... [ 0.3, 0.4, 0.4, 0.5, 0.8], ... [ 0.4, 0.3, 0.6, 0.7, 0.8], ... [ 0.2, 0.3, 0.2, 0.3, 0.7], ... [ 0.3, 0.4, 0.4, 0.4, 0.7],] >>> # Estimating similarity matrix >>> sim_mat = clust.get_sim_mat(emb) >>> print (np.around(sim_mat[5:,5:], decimals=3)) [[1. 0.957 0.961 0.904 0.966] [0.957 1. 0.977 0.982 0.997] [0.961 0.977 1. 0.928 0.972] [0.904 0.982 0.928 1. 0.976] [0.966 0.997 0.972 0.976 1. ]] >>> # Pruning >>> pruned_sim_mat = clust.p_pruning(sim_mat, 0.3) >>> print (np.around(pruned_sim_mat[5:,5:], decimals=3)) [[1. 0. 0. 0. 0. ] [0. 1. 0. 0.982 0.997] [0. 0.977 1. 0. 0.972] [0. 0.982 0. 1. 0.976] [0. 0.997 0. 0.976 1. ]] >>> # Symmetrization >>> sym_pruned_sim_mat = 0.5 * (pruned_sim_mat + pruned_sim_mat.T) >>> print (np.around(sym_pruned_sim_mat[5:,5:], decimals=3)) [[1. 0. 0. 0. 0. ] [0. 1. 0.489 0.982 0.997] [0. 0.489 1. 0. 0.486] [0. 0.982 0. 1. 0.976] [0. 0.997 0.486 0.976 1. ]] >>> # Laplacian >>> laplacian = clust.get_laplacian(sym_pruned_sim_mat) >>> print (np.around(laplacian[5:,5:], decimals=3)) [[ 1.999 0. 0. 0. 0. ] [ 0. 2.468 -0.489 -0.982 -0.997] [ 0. -0.489 0.975 0. -0.486] [ 0. -0.982 0. 1.958 -0.976] [ 0. -0.997 -0.486 -0.976 2.458]] >>> # Spectral Embeddings >>> spec_emb, num_of_spk = clust.get_spec_embs(laplacian, 3) >>> print(num_of_spk) 3 >>> # Clustering >>> clust.cluster_embs(spec_emb, num_of_spk) >>> # print (clust.labels_) # [0 0 0 2 2 2 1 1 1 1] >>> # Complete spectral clustering >>> clust.do_spec_clust(emb, k_oracle=3, p_val=0.3) >>> # print(clust.labels_) # [0 0 0 2 2 2 1 1 1 1] r rcCs||_||_dS)N min_num_spkrs max_num_spkrs)rrrrrr__init__s zSpec_Clust_unorm.__init__c CsN||}|||}d||j}||}|||\}} ||| dS)a_Function for spectral clustering. Arguments --------- X : array (n_samples, n_features). Embeddings extracted from the model. k_oracle : int Number of speakers (when oracle number of speakers). p_val : float p percent value to prune the affinity matrix. rN) get_sim_mat p_pruningr get_laplacian get_spec_embs cluster_embs) rrk_oraclep_valsim_matpruned_sim_matsym_pruned_sim_matremb num_of_spkrrr do_spec_clust s   zSpec_Clust_unorm.do_spec_clustcCstjj||}|S)aReturns the similarity matrix based on cosine similarities. Arguments --------- X : array (n_samples, n_features). Embeddings extracted from the model. Returns ------- M : array (n_samples, n_samples). Similarity matrix with cosine similarities between each pair of embedding. )sklearnmetricspairwisecosine_similarity)rrMrrrr?szSpec_Clust_unorm.get_sim_matcCs\td||jd}t|jdD]}t||ddf}|d|}d|||f<q|S)aRefine the affinity matrix by zeroing less similar values. Arguments --------- A : array (n_samples, n_samples). Affinity matrix. pval : float p-value to be retained in each row of the affinity matrix. Returns ------- A : array (n_samples, n_samples). pruned affinity matrix based on p_val. rrN)intrarLrdargsort)rApvaln_elemsrR low_indexesrrrrRs  zSpec_Clust_unorm.p_pruningcCs>d|t|jd<tjt|dd}t|}||}|S)a6Returns the un-normalized laplacian for the given affinity matrix. Arguments --------- M : array (n_samples, n_samples) Affinity matrix. Returns ------- L : array (n_samples, n_samples) Laplacian matrix. rrr)rd diag_indicesrargrdiag)rrDLrrrros  zSpec_Clust_unorm.get_laplacianrYcCstj|\}}|dur|}n'||d|j}|r*t|dt|jt|ndd}||j kr6|j }|ddd|f}||fS)aReturns spectral embeddings and estimates the number of speakers using maximum Eigen gap. Arguments --------- L : array (n_samples, n_samples) Laplacian matrix. k_oracle : int Number of speakers when the condition is oracle number of speakers, else None. Returns ------- emb : array (n_samples, n_components) Spectral embedding for each sample with n Eigen components. num_of_spk : int Estimated number of speakers. If the condition is set to the oracle number of speakers then returns k_oracle. Nrrr ) scipylinalgeigh getEigenGapsrrdrminrMr)rrrlambdaseig_vecsrlambda_gap_listrrrrrs" zSpec_Clust_unorm.get_spec_embscCst||\}|_}dS)aClusters the embeddings using kmeans. Arguments --------- emb : array (n_samples, n_components) Spectral embedding for each sample with n Eigen components. k : int Number of clusters to kmeans. N)rr)rrrrrrrrrs zSpec_Clust_unorm.cluster_embscCsDg}tt|dD]}t||dt||}||q |S)a&Returns the difference (gaps) between the Eigen values. Arguments --------- eig_vals : list List of eigen values Returns ------- eig_vals_gap_list : list List of differences (gaps) between adjacent Eigen values. r)rLrMfloatr)reig_valseig_vals_gap_listrRgaprrrrs   zSpec_Clust_unorm.getEigenGapsN)r r)rY) rrrrrrrrrrrrrrrrrs J - rcCs|dkrtddd}|}||j|||j} nt|dddd}||j||j} |j} g} t| jd D]3} |d t | | } | | }| d d}t |d }t |d }t |d}|||| g}| |q6| j d d dt| } t| } t| |dS)aPerforms spectral clustering on embeddings. This function calls specific clustering algorithms as per affinity. Arguments --------- diary_obj : StatObject_SB type Contains embeddings in diary_obj.stat1 and segment IDs in diary_obj.segset. out_rttm_file : str Path of the output RTTM file. rec_id : str Recording ID for the recording under processing. k : int Number of speaker (None, if it has to be estimated). pval : float `pval` for pruning affinity matrix. affinity_type : str Type of similarity to be used to get affinity matrix (cos or nn). n_neighbors : int Number of neighbors to use for clustering cosr rrkmeansrnearest_neighbors)r assign_labelsrrrrrrcS t|dSNrrrrrrr z$do_spec_clustering..keyN)rrstat1rrrsegsetrLrar%rsplitrrsortrTrVr^) diary_objr[r1rr affinity_typer clust_objrr subseg_idsrNrRspkr_idsub_segsplitted sseg_startsseg_endr0rrrdo_spec_clusterings8       r rY333333?cCs|dur|}n&tddd}||j}|||}d||j} || } || |\} }t|j|\} } } |j} g}t | j dD]3}|dt | |}| |}| dd}t |d}t |d}t |d}||||g}||qB|jd d d t|}t|}t||dS) axPerforms kmeans clustering on embeddings. Arguments --------- diary_obj : StatObject_SB type Contains embeddings in diary_obj.stat1 and segment IDs in diary_obj.segset. out_rttm_file : str Path of the output RTTM file. rec_id : str Recording ID for the recording under processing. k_oracle : int Number of speaker (None, if it has to be estimated). p_val : float `pval` for pruning affinity matrix. Used only when number of speakers are unknown. Note that this is just for experiment. Prefer Spectral clustering for better clustering results. Nr rrrrrrrcSrrrrrrrr^rz&do_kmeans_clustering..r)rrrrrrrrrrLrar%rrrrrTrVr^)rr[r1rrrrrrrrrrrrrNrRrrrr r r0rrrdo_kmeans_clusterings0          r cCs ddlm}||dur|}||ddd|j}|j}n|ddd|d|j}|j}|j} g} t|jdD]3} |dt || } | | } | dd }t |d}t |d }t |d }|||| g}| |q:| j d d d t| } t| } t| |dS)aPerforms Agglomerative Hierarchical Clustering on embeddings. Arguments --------- diary_obj : StatObject_SB type Contains embeddings in diary_obj.stat1 and segment IDs in diary_obj.segset. out_rttm_file : str Path of the output RTTM file. rec_id : str Recording ID for the recording under processing. k_oracle : int Number of speaker (None, if it has to be estimated). p_val : float `pval` for pruning affinity matrix. Used only when number of speakers are unknown. Note that this is just for experiment. Prefer Spectral clustering for better clustering results. r)AgglomerativeClusteringNcosineward)rrlinkage)rrrdistance_thresholdrrr rcSrrrrrrrrrzdo_AHC..r)sklearn.clusterr norm_stat1fitrrrrLrar%rrrrrTrVr^)rr[r1rrrr clusteringrrrNrRrrrr r r0rrrdo_AHClsH       r)rTT)rNNr)rYr )-rr<rrnumpyrdrrscipy.sparse.csgraphrrrscipy.sparse.linalgrrrrrrsklearn.cluster._kmeansrsklearn.neighborsr ImportErrorerr_msgrr3rFrIrTrVr^rvryrrrrrrrrr r rrrrrst         !4N%*-* \ <)G P