o }oi @sddlZddlZddlZddlZddlmZddlmZddlm Z m Z m Z ddl Z ddlmZddlmZddlmZddlmZmZmZmZmZdd lmZzddlZd ZWn eycd ZYnwd gZd e de!fddZ"de de fddZ#de de e e e ffddZ$de e e e%fde e e e ffddZ& d1de e e e%fde de'de e e e ffd d!Z(d"e e e e e%ffd#e e e e e%ffd$e e fd%d&Z)d'e d(e%d)e%d*e de f d+d,Z*d-e e de'fd.d/Z+Gd0d d Z,dS)2N) OrderedDict)datetime)DictListTuple)concat_perm_word_error_rate)word_error_rate)ClusteringDiarizer)audio_rttm_mapget_uniqname_from_filepathlabels_to_rttmfilerttm_to_labelswrite_rttm2manifest)loggingTFOfflineDiarWithASR file_pathsession_trans_dictcCs@t|d}tj||ddWddS1swYdS)a  Write a json file from the session_trans_dict dictionary. Args: file_path (str): Target filepath where json file is saved session_trans_dict (dict): Dictionary containing transcript, speaker labels and timestamps w)indentN)openjsondump)rroutfilerf/home/ubuntu/.local/lib/python3.10/site-packages/nemo/collections/asr/parts/utils/diarization_utils.pydump_json_to_file/s "rw_pathvalcCs>t|d}||dWddS1swYdS)z Write a text file from the string input. Args: w_path (str): Target path for saving a file val (str): String variable to be written r N)rwrite)rroutputrrr write_txt=s "r" ctm_file_pathreturncCsgi}}t|}|D]"}|}|d}||vrg||<|||d||dq dd|D}d|}||fS)a Convert ctm file into a list containing transcription (space seperated string) per each speaker. Args: ctm_file_path (str): Filepath to the reference CTM files. Returns: spk_reference (list): List containing the reference transcripts for each speaker. Example: >>> spk_reference = ["hi how are you well that's nice", "i'm good yeah how is your sister"] mix_reference (str): Reference transcript from CTM file. This transcript has word sequence in temporal order. Example: >>> mix_reference = "hi how are you i'm good well that's nice yeah how is your sister" rcSg|]}d|qS join.0 word_listrrr iz'convert_ctm_to_text..r()r readlinessplitappendvaluesr*)r# mix_referenceper_spk_ref_trans_dict ctm_contentctm_line ctm_splitspk spk_referencerrrconvert_ctm_to_textKs   r;word_dict_seq_listcCspgi}}|D]}|d}||vrg||<|||d||dqdd|D}d|}||fS)ac Convert word_dict_seq_list into a list containing transcription (space seperated string) per each speaker. Args: word_dict_seq_list (list): List containing words and corresponding word timestamps in dictionary format. Example: >>> word_dict_seq_list = >>> [{'word': 'right', 'start_time': 0.0, 'end_time': 0.04, 'speaker': 'speaker_0'}, {'word': 'and', 'start_time': 0.64, 'end_time': 0.68, 'speaker': 'speaker_1'}, ...], Returns: spk_hypothesis (list): Dictionary containing the hypothesis transcript for each speaker. A list containing the sequence of words is assigned for each speaker. Example: >>> spk_hypothesis= ["hi how are you well that's nice", "i'm good yeah how is your sister"] mix_hypothesis (str): Hypothesis transcript from ASR output. This transcript has word sequence in temporal order. Example: >>> mix_hypothesis = "hi how are you i'm good well that's nice yeah how is your sister" speakerwordcSr&r'r)r+rrrr.r/z1convert_word_dict_seq_to_text..r()r2r3r*)r<mix_hypothesisper_spk_hyp_trans_dict word_dictr9spk_hypothesisrrrconvert_word_dict_seq_to_textns  rCnulluniq_iddecimalsc Cstg}d}|D]1}|d}|d}t|d|d|}|d} |d|d|d|d| d| } || q|S)a Convert word_dict_seq_list into a list containing transcription in CTM format. Args: word_dict_seq_list (list): List containing words and corresponding word timestamps in dictionary format. Example: >>> word_dict_seq_list = >>> [{'word': 'right', 'start_time': 0.0, 'end_time': 0.34, 'speaker': 'speaker_0'}, {'word': 'and', 'start_time': 0.64, 'end_time': 0.81, 'speaker': 'speaker_1'}, ...], Returns: ctm_lines_list (list): List containing the hypothesis transcript in CTM format. Example: >>> ctm_lines_list= ["my_audio_01 speaker_0 0.0 0.34 right 0", my_audio_01 speaker_0 0.64 0.81 and 0", rr= start_timeend_timer>r()roundr2) r<rFrG ctm_lines confidencerAr9sttdurr> ctm_line_strrrrconvert_word_dict_seq_to_ctms& rP der_results wer_results csv_columnscCs~i}|D]0}|dkr qdd|D||<|||d<||vr)||||||vr6||||qt|}|S)a Merge WER results and DER results into a single dictionary variable. Args: der_results (dict): Dictionary containing FA, MISS, CER and DER values for both aggregated amount and each session. wer_results (dict): Dictionary containing session-by-session WER and cpWER. `wer_results` only exists when CTM files are provided. Returns: total_result_dict (dict): Dictionary containing both DER and WER results. This dictionary contains unique-IDs of each session and `total` key that includes average (cp)WER and DER/CER/Miss/FA values. totalcSsi|]}|dqS)-rr,xrrr sz)get_total_result_dict..rF)keysupdatelistr3)rQrRrStotal_result_dictrFtotal_result_jsonsrrrget_total_result_dicts   r^r>stt_secend_secr=cCs(|dd}|d|d|d|S)aA Get a string formatted line for Audacity label. Args: word (str): A decoded word stt_sec (float): Start timestamp of the word end_sec (float): End timestamp of the word Returns: speaker (str): Speaker label in string type _ z [] )r1)r>r_r`r=r9rrrget_audacity_labelsrelabelscCsdd|D}tt|S)at Count the number of speakers in a segment label list. Args: labels (list): List containing segment start and end timestamp and speaker labels. Example: >>> labels = ["15.25 21.82 speaker_0", "21.18 29.51 speaker_1", ... ] Returns: n_spk (int): The number of speakers in the list `labels` cSsg|] }|ddqS)r(rb)r1striprVrrrr.sz.get_num_of_spk_from_labels..)lenset)rfspk_setrrrget_num_of_spk_from_labelss rkc@s2eZdZdZddZedeeeffddZede efddZ d d Z d d Z d ede fddZddZdeeeee efffddZede edede efddZdeee effddZdfddZe dgd eeeeeffd!eeeeeffd"ed#e deeeeefff d$d%Z &dhd'ed(ejd)e defd*d+Zd,e edeee efdeee e effd-d.Zd/eee efd0eee efd1eee efdeeeeefffd2d3Z 4 5did6e ed7e ed8e e ed9e e ed#e deeeeefff d:d;Zd ed7e edZd?e d@e dee e ffdAdBZdCe edefdDdEZd. ctm_filepathNT) r rtAUDIO_RTTM_MAPitemsaudio_file_list ctm_file_list enumerater r2rhr)rkaudio_file_pathrFrrrres  z"OfflineDiarWithASR.make_file_listscCsJ|jd|jdf|_ddg|_td|jdt|jddS)zS Load ARPA language model for realigning speaker labels for words. min_number_of_wordsmax_number_of_wordszzzLoading LM for realigning: arpa_language_modelr)rsN_rangestt_end_tokensrinfoarpaloadfrrrr_load_realigning_LMzs  z&OfflineDiarWithASR._load_realigning_LMrFn_spkcCstd|d|ggdS)z Initialize json (in dictionary variable) formats for session level result and Gecko style json. Returns: (dict): Session level result dictionary variable initialized)status session_id transcription speaker_countwords sentencesod)rrFrrrr_init_session_trans_dictsz+OfflineDiarWithASR._init_session_trans_dictcCstdgdS)z Initialize a dictionary format for Gecko style json. Returns: (dict): Gecko style json dictionary. g@) schemaVersion monologuesrrrrr_init_session_gecko_dictsz+OfflineDiarWithASR._init_session_gecko_dict word_ts_dictc Csi|_t|D]6\}\}}|||j}||}tj|j d}tj |s.t |t |||}|j ||d|j|<q dS)z Take the non_speech labels from logit output. The logit output is obtained from `run_ASR` function. Args: word_ts_dict (dict): Dictionary containing word timestamps. pred_rttms)r rttm_filepathN) VAD_RTTM_MAPrr)get_speech_labels_from_decoded_predictionrvget_str_speech_labelsospathr*ryexistsmakedirsr r) rridxrFword_timestampsspeech_labels_float speech_labels output_pathfilenamerrr_save_VAD_labels_lists     z(OfflineDiarWithASR._save_VAD_labels_list input_word_tsrvcCsg}t|}|gkr |St|d}|dkrOt|dkrG||d||dd|krG||}||d}||d|d|dg|d8}|dks|S)a> Extract speech labels from the ASR output (decoded predictions) Args: input_word_ts (list): List containing word timestamps. Returns: word_ts (list): The ranges of the speech segments, which are merged ranges of input_word_ts. r%r)copydeepcopyrhpopinsert)rrvrword_tscounttrangeBtrangeArrrrs     zwY||j |<qdS)a? Read frame-level VAD outputs. Args: vad_processing_dir (str): Path to the directory where the VAD results are saved. smoothing_type (bool or str): [False, median, mean] type of smoothing applied softmax logits to smooth the predictions. frame.rN) isinstanceboolrrrr*rr0r2floatrgr) rrrext_typerF frame_vadframe_vad_float_listfplinerrrrs     z'OfflineDiarWithASR._get_frame_level_VADraudio_rttm_map_dicttrans_info_dictryrGc Csz|\}}}|j}i} d} |D]} | \} } d|| vr!|| d}n tj|d| d}t|}|| d}t|}t|}t|}| d| d| d| d | d| d | d| d | d| d f\}}}}t||t||t||t|||||| d | | <| t||k7} qt||d|d |d|d |d|d f\}}}}||||| t |jd | d <| S) a! Gather diarization evaluation results from pyannote DiarizationErrorRate metric object. Args: metric (DiarizationErrorRate metric): DiarizationErrorRate metric pyannote object trans_info_dict (dict): Dictionary containing word timestamps, speaker labels and words from all sessions. Each session is indexed by unique ID as a key. mapping_dict (dict): Dictionary containing speaker mapping labels for each audio file with key as unique name decimals (int): The number of rounding decimals for DER value Returns: der_results (dict): Dictionary containing scores for each audio file along with aggregated results rhyp_rttm_filepathrrr confusionz false alarmzmissed detectionrT)rrrrrrr)rrrrspk_counting_acc) results_rrr*r rkrJintabsrh) diar_scorerrryrGmetric mapping_dictraresultsrQcount_correct_spk_countingresultkeyrr pred_labelsref_rttm ref_labelsrr_DER_CER_FA_MISSrrrrrrrgather_eval_resultssP         z&OfflineDiarWithASR.gather_eval_results vad_index_word_end vad_framesoffsetcCsx||}td|j|}|t|kr*|||jkrn|d7}||kr$n|t|kstt|d|}t|dd}|S)aN Find the closest silence frame from the given starting position. Args: vad_index_word_end (float): The timestamp of the end of the current word. vad_frames (numpy.array): The numpy array containing frame-level VAD probability. params (dict): Contains the parameters for diarization and ASR decoding. Returns: cursor (float): A timestamp of the earliest start of a silence region from the given time point, vad_index_word_end. dr%gY@)rr{rhrzminrJ)rrrrcursorlimitrrr_get_the_closest_silence_startas  z1OfflineDiarWithASR._get_the_closest_silence_startrcCsi}t|D]\}\}}t|}g}t|D]o\} } | |dkr|t| d| dd} t|| dd| ddd} ||jvr^td| d} || |j|}t|| dd}n| }t|| }tt|j || }| | d| d|gq| | d| dgq|||<q|S)a Compensate the word timestamps based on the VAD output. The length of each word is capped by self.max_word_ts_length_in_sec. Args: audio_file_list (list): List containing audio file paths. word_ts_dict (dict): Dictionary containing timestamps of words. Returns: enhanced_word_ts_dict (dict): Dictionary containing the enhanced word timestamp values indexed by unique-IDs. r%rrg{Gz?r) rrrhrJrrrrmaxr{r2)rrrenhanced_word_ts_dictrrFword_ts_seq_listNenhanced_word_ts_bufferrrword_lenlen_to_next_wordrclosest_sil_stt vad_est_len min_candidatefixed_word_lenrrr_compensate_word_ts_lists* "    z+OfflineDiarWithASR._compensate_word_ts_listrword_hyp word_ts_hypcCsi}|jr|jikrtd||j|}n|}|jdr)ts$td| |_ g}t |jD]4\}}t |} || || } } || || } } |j | | | | d}|j r[||}|| | ||| <q0td|jd|S)a[ Match the diarization result with the ASR output. The words and the timestamps for the corresponding words are matched in a for loop. Args: diar_hyp (dict): Dictionary of the Diarization output labels in str. Indexed by unique IDs. Example: >>> diar_hyp['my_audio_01'] = ['0.0 4.375 speaker_1', '4.375 5.125 speaker_0', ...] word_hyp (dict): Dictionary of words from ASR inference. Indexed by unique IDs. Example: >>> word_hyp['my_audio_01'] = ['hi', 'how', 'are', ...] word_ts_hyp (dict): Dictionary containing the start time and the end time of each word. Indexed by unique IDs. Example: >>> word_ts_hyp['my_audio_01'] = [[0.0, 0.04], [0.64, 0.68], [0.84, 0.88], ...] Returns: trans_info_dict (dict): Dictionary containing word timestamps, speaker labels and words from all sessions. Each session is indexed by a unique ID. zkVAD timestamps are not provided. Fixing word timestamps without VAD. Please check the hydra configurations.rzbLM for realigning is provided but arpa is not installed. Install arpa using PyPI: pip install arpa)rr word_rfnd_ts diar_labelsz0Diarization with ASR output files are saved in: /pred_rttms)rwrrwarningr%rrsARPA ImportErrorrr~rr get_word_level_json_listrealign_words_with_lm_make_json_outputrry)rrr&r'rword_ts_refinedr<rrrFrr)rr(rrr"get_transcript_with_speaker_labelss6     z5OfflineDiarWithASR.get_transcript_with_speaker_labelsNrrr)rr(cCs|dur|}|d\}}}d\} } g} tt|||D]@\} \} }}||} | t|krE| d7} t| t|d} || \}}}t|d|}t|d|}| | |||dq| S)a Assign speaker labels to each word and save the hypothesis words and speaker labels to a dictionary variable for future use. Args: uniq_id (str): A unique ID (key) that identifies each input audio file. diar_labels (list): List containing the Diarization output labels in str. Indexed by unique IDs. Example: >>> diar_labels = ['0.0 4.375 speaker_1', '4.375 5.125 speaker_0', ...] words (list): Dictionary of words from ASR inference. Indexed by unique IDs. Example: >>> words = ['hi', 'how', 'are', ...] word_ts (list): Dictionary containing the start time and the end time of each word. Indexed by unique IDs. Example: >>> word_ts = [[0.0, 0.04], [0.64, 0.68], [0.84, 0.88], ...] word_ts_refined (list): Dictionary containing the refined (end point fixed) word timestamps based on hypothesis word timestamps. Indexed by unique IDs. Example: >>> word_rfnd_ts = [[0.0, 0.60], [0.64, 0.80], [0.84, 0.92], ...] Returns: word_dict_seq_list (list): List containing word by word dictionary containing word, timestamps and speaker labels. Example: >>> [{'word': 'right', 'start_time': 0.0, 'end_time': 0.04, 'speaker': 'speaker_0'}, {'word': 'and', 'start_time': 0.64, 'end_time': 0.68, 'speaker': 'speaker_1'}, {'word': 'i', 'start_time': 0.84, 'end_time': 0.88, 'speaker': 'speaker_1'}, ...] Nr)rrr%)r>rHrIr=) r1rzip_get_word_timestamp_anchorrrrhrJr2)rrr)rr(rG start_point end_pointr=word_posturn_idxr<word_idxr>word_ts_stt_endrefined_word_ts_stt_endr_r`rrrr.s3  z+OfflineDiarWithASR.get_word_level_json_listr<cCsgg}}|d\}}}|} gg} } |||dd} t|} td|d| d|j|| d}|}t|D]t\}}|d|d }}|||d |d }}|| krt| dkro|d d | d| dg} | d | d<| | |||dd} n|| d <||}}| |||dd| d| d7<|t |||||} q;||d<| d | d<| | |d d |d| dd ||d<| |d<| ||||| |S)a< Generate json output files and transcripts from the ASR and diarization results. Args: uniq_id (str): A unique ID (key) that identifies each input audio file. diar_labels (list): List containing the diarization hypothesis timestamps Example: >>> diar_hyp['my_audio_01'] = ['0.0 4.375 speaker_1', '4.375 5.125 speaker_0', ...] word_dict_seq_list (list): List containing words and corresponding word timestamps in dictionary format. Example: >>> [{'word': 'right', 'start_time': 0.0, 'end_time': 0.04, 'speaker': 'speaker_0'}, {'word': 'and', 'start_time': 0.64, 'end_time': 0.68, 'speaker': 'speaker_1'}, {'word': 'i', 'start_time': 0.84, 'end_time': 0.88, 'speaker': 'speaker_1'}, ...] Returns: session_result_dict (dict): A dictionary containing overall results of diarization and ASR inference. `session_result_dict` has following keys: `status`, `session_id`, `transcription`, `speaker_count`, `words`, `sentences`. Example: >>> session_trans_dict = { 'status': 'Success', 'session_id': 'my_audio_01', 'transcription': 'right and i really think ...', 'speaker_count': 2, 'words': [{'word': 'right', 'start_time': 0.0, 'end_time': 0.04, 'speaker': 'speaker_0'}, {'word': 'and', 'start_time': 0.64, 'end_time': 0.68, 'speaker': 'speaker_1'}, {'word': 'i', 'start_time': 0.84, 'end_time': 0.88, 'speaker': 'speaker_1'}, ... ] 'sentences': [{'sentence': 'right', 'start_time': 0.0, 'end_time': 0.04, 'speaker': 'speaker_0'}, {'sentence': 'and i really think ...', 'start_time': 0.92, 'end_time': 4.12, 'speaker': 'speaker_0'}, ... ] } rr)r=rHrItextzCreating results for Session: z n_spk: r()rFrr>r=rHrIrN)nameid)r=termsr<WORD)startendr<typerrr) r1rkrrrrrr2rhrgrer*_write_and_log)rrFr)r< word_seq_listaudacity_label_wordsr5r6r= prev_speakerr terms_listsentencerr gecko_dictrrAr>r_r`rrrr03sH 1      z$OfflineDiarWithASR._make_json_outputr word_seq_lencCs||jdkrt||jd}t|||jd}||fS|||jdkr but he's coming back to albuquerque hyp_latter: since i think like tuesday but he's coming back to albuquerque The joint probabilities of words in the sentence are computed for these two hypotheses. In addition, logprob_diff_threshold parameter is used for reducing the false positive realigning. Args: word_dict_seq_list (list): List containing words and corresponding word timestamps in dictionary format. Returns: realigned_list (list): List of dictionaries containing words, word timestamps and speaker labels. r>r=rr%r(Nrblogprob_diff_threshold)rhrr2rrrrNr~log_sr*rnpargsortrs)rr<rKhyp_w_dict_listspk_listr line_dictr> spk_labelrealigned_list org_spk_listrLrM hyp_former hyp_latterlog_pp_orderrrrr/s4    ((*2"   z(OfflineDiarWithASR.realign_words_with_lmhyp_trans_info_dicthyp_ctm_file_listref_ctm_file_listcCsi}|durgg}}gg}}gg} } tt||D]m\} \} } t| }| ||t| kr5td|durN|t|| krJt|| \}}ntd|dura||vrat||d\}}ntdt| \}}||||||||| t|g|gqt||\}}}i|d<t||d|dd<t||d|dd <t| || D]\}}}i||<|||d <|||d <q|S) a Evaluate the result transcripts based on the provided CTM file. WER and cpWER are calculated to assess the performance of ASR system and diarization at the same time. Args: audio_file_list (list): List containing file path to the input audio files. hyp_trans_info_dict (dict): Dictionary containing the hypothesis transcriptions for all sessions. hyp_ctm_file_list (list): List containing file paths of the hypothesis transcriptions in CTM format for all sessions. ref_ctm_file_list (list): List containing file paths of the reference transcriptions in CTM format for all sessions. Note: Either `hyp_trans_info_dict` or `hyp_ctm_file_list` should be provided. Returns: wer_results (dict): Session-by-session results including DER, miss rate, false alarm rate, WER and cpWER Nz:audio_file_list has mismatch in uniq_id with ctm_file_pathz;Hypothesis CTM files are provided but uniq_id is mismatchedrz=Hypothesis information is not provided in the correct format.rT) hypotheses references average_cpWER average_WERrr) rr3r r2 ValueErrorr;rCrr)rr`rarbrRspk_hypothesesspk_referencesmix_hypothesesmix_references WER_values uniq_id_listrrr#rFrBr?r:r4 cpWER_valueshyps_spkrefs_spkrrrrrevaluatesD             zOfflineDiarWithASR.evaluatercCs(g}|D] \}}|d||q|S)a Convert floating point speech labels list to a list containing string values. Args: speech_labels_float (list): List containing start and end timestamps of the speech segments in floating point type speech_labels (list): List containing start and end timestamps of the speech segments in string format z{:.3f} {:.3f} speech)r2format)rrrArBrrrr_s z(OfflineDiarWithASR.get_str_speech_labels ctm_eval.csvrQrRrS csv_file_namec Cs|d}tj|ddtd|d|t|||}z4t|d|d}tj||d}||D]} | | q6WdWdS1sJwYWdSt y`td YdSw) aQ This function is for development use when a CTM file is provided. Saves the session-level diarization and ASR result into a csv file. Args: wer_results (dict): Dictionary containing session-by-session results of ASR and diarization in terms of WER and cpWER. r*T)exist_okzWriting /r) fieldnamesNz0I/O error has occurred while writing a csv file.) rrrrr^rcsv DictWriter writeheaderwriterowIOError) rQrRryrSrt target_pathr]csvfilewriterdatarrr!write_session_level_result_in_csvos   & z4OfflineDiarWithASR.write_session_level_result_in_csvZ string_outmax_chars_in_linec Cs|jdr tdnd}|d}g}|D]G}g}t|||krV|dkr*|d|nd}t|t||D]} || | |} t| dkrO|| } || q4||q||qd|}|S)a[ Break the lines in the transcript. Args: string_out (str): Input transcript with speaker labels max_chars_in_line (int): Maximum characters in each line Returns: return_string_out (str): String variable containing line breaking colored_textzrrNr)rorhr1rangergr2extendr*) rrr color_str_lensplit_string_outreturn_string_out org_chunkbuffer color_stri trans_str c_trans_strrrr _break_liness"     zOfflineDiarWithASR._break_linesrrFrJrcCs||}|jdr||}d|d<t|d}t|jd|d|t|jd|d|t|jd|dd |t|jd|d |t|jd|d d |d S) ad Write output files and display logging messages. Args: uniq_id (str): A unique ID (key) that identifies each input audio file session_trans_dict (dict): Dictionary containing the transcription output for a session audacity_label_words (list): List containing word and word timestamp information in Audacity label format gecko_dict (dict): Dictionary formatted to be opened in Gecko software sentences (list): List containing sentence dictionary break_linessuccessrrz /pred_rttms/z.jsonz _gecko.jsonz.ctmrz.txtz.w.labelN) print_sentencesrorrPrryr"r*rg)rrFrrFrJrrctm_lines_listrrrrDs    "z!OfflineDiarWithASR._write_and_logc Csd|dddd|dddd|dddd |dd dd |dd d }|d urMt|dkrMt|d|dddd|dddd St|d S)a Print a slew of error metrics for ASR and Diarization. Args: der_results (dict): Dictionary containing FA, MISS, CER and DER values for both aggregated amount and each session. wer_results (dict): Dictionary containing session-by-session WER and cpWER. `wer_results` only exists when CTM files are provided. z DER : rTrz.4fz, FA : rz, MISS : rz, CER : rz, Spk. counting acc. : rNrz cpWER : rez, WER : rf)rhrr)rQrRDER_inforrr print_errorss$     zOfflineDiarWithASR.print_errorsc Csd}|D]x}|d}|d}|d}|d}|jdr#|j|d}nd}d} t|d kr0d } nd } tt|d tt|d }}t|| | d d} t|| | d d} |jdrld| d| d} nd} ||| |d|d7}q|S)a= Print a transcript with speaker labels and timestamps. Args: sentences (list): List containing sentence-level dictionaries. Returns: string_out (str): String variable containing transcript and the corresponding speaker label. rr=rHrIr<rriiz %H:%M:%S.%fz%M:%S.%frN print_time[z - rdz: r)rorgetrrr fromtimestampstrftime) rrrrIr=r5r6r<colordatetime_offsettime_strstart_point_str end_point_strrrrrs*    z"OfflineDiarWithASR.print_sentences)F)r)r)Nr)NN)rs)r)(__name__ __module__ __qualname____doc__r staticmethodrstrrrrrrrrrrrrrrrrTndarrayrr%r2r.r0rrNr4r/rqrrrrDrrrrrrrs( "   ( M ! * I   C k*"6 M    %4)rDrE)-rrxrr collectionsrrrtypingrrrnumpyrT nemo.collections.asr.metrics.derr nemo.collections.asr.metrics.werrnemo.collections.asr.modelsr .nemo.collections.asr.parts.utils.speaker_utilsr r r r r nemo.utilsrrr,r-__all__rdictrr"r;rrCrrPr^rerkrrrrrsX       *#+ & !