o SiL@sxdZddlZddlZddlZddlZddlZddlmZmZddl m Z ddl m Z ddl mZmZddlmZmZmZmZmZddlmZdd lmZmZmZmZmZmZdd lm Z dd l!m"Z"m#Z#dd l$m%Z%m&Z&gd Z'gdZ(dZ)edgdZ* d5de+dee+de+de+de+de%dee,fddZ-  d6de%dee,defddZ.  d6de%dee,defd d!Z/    "d7d#ee%d$ee%d%ee%d&e0dee+ee+eeeffff d'd(Z1d)e%d&e0dee+ee+eeeffffd*d+Z2d,ee+ee+eeefffdee+eeefffd-d.Z3d)e%d&e0dee+ee+eeeffffd/d0Z4d1e%d2ee+e*fdeeeffd3d4Z5dS)8a The following description is taken from the official website: https://www.robots.ox.ac.uk/~vgg/data/voxceleb/ VoxCeleb is an audio-visual dataset consisting of short clips of human speech, extracted from interview videos uploaded to YouTube. VoxCeleb contains speech from speakers spanning a wide range of different ethnicities, accents, professions and ages. There are a total of 7000+ speakers and 1 million utterances. All speaking face-tracks are captured "in the wild", with background chatter, laughter, overlapping speech, pose variation and different lighting conditions. VoxCeleb consists of both audio and video, comprising over 2000 hours of speech. Each segment is at least 3 seconds long. The dataset consists of two versions, VoxCeleb1 and VoxCeleb2. Each version has it's own train/test split. For each version, the YouTube URLs, face detections and tracks, audio files, cropped face videos and speaker meta-data are provided. There is no overlap between the two versions. - VoxCeleb1: VoxCeleb1 contains over 100,000 utterances for 1,251 celebrities. http://www.robots.ox.ac.uk/~vgg/data/voxceleb/ - VoxCeleb2: VoxCeleb2 contains over a million utterances for 6,112 identities. http://www.robots.ox.ac.uk/~vgg/data/voxceleb2/ LICENSE: The VoxCeleb dataset is available to download for commercial/research purposes under a Creative Commons Attribution 4.0 International License. The copyright remains with the original owners of the video. This Lhotse recipe prepares the VoxCeleb1 and VoxCeleb2 datasets. N) defaultdict namedtuple) as_completed)ProcessPoolExecutor)PathPurePath)DictListOptionalTupleUnion)tqdm)CutSetMonoCut Recording RecordingSetSupervisionSegmentSupervisionSet)combine) fix_manifests$validate_recordings_and_supervisions)Pathlikeresumable_download)zIhttps://thor.robots.ox.ac.uk/~vgg/data/voxceleb/vox1a/vox1_dev_wav_partaazIhttps://thor.robots.ox.ac.uk/~vgg/data/voxceleb/vox1a/vox1_dev_wav_partabzIhttps://thor.robots.ox.ac.uk/~vgg/data/voxceleb/vox1a/vox1_dev_wav_partaczIhttps://thor.robots.ox.ac.uk/~vgg/data/voxceleb/vox1a/vox1_dev_wav_partadzGhttps://thor.robots.ox.ac.uk/~vgg/data/voxceleb/vox1a/vox1_test_wav.zipzAhttps://www.robots.ox.ac.uk/~vgg/data/voxceleb/meta/vox1_meta.csv) zIhttps://thor.robots.ox.ac.uk/~vgg/data/voxceleb/vox1a/vox2_dev_aac_partaazIhttps://thor.robots.ox.ac.uk/~vgg/data/voxceleb/vox1a/vox2_dev_aac_partabzIhttps://thor.robots.ox.ac.uk/~vgg/data/voxceleb/vox1a/vox2_dev_aac_partaczIhttps://thor.robots.ox.ac.uk/~vgg/data/voxceleb/vox1a/vox2_dev_aac_partadzIhttps://thor.robots.ox.ac.uk/~vgg/data/voxceleb/vox1a/vox2_dev_aac_partaezIhttps://thor.robots.ox.ac.uk/~vgg/data/voxceleb/vox1a/vox2_dev_aac_partafzIhttps://thor.robots.ox.ac.uk/~vgg/data/voxceleb/vox1a/vox2_dev_aac_partagzIhttps://thor.robots.ox.ac.uk/~vgg/data/voxceleb/vox1a/vox2_dev_aac_partahzGhttps://thor.robots.ox.ac.uk/~vgg/data/voxceleb/vox1a/vox2_test_aac.zipzAhttps://www.robots.ox.ac.uk/~vgg/data/voxceleb/meta/vox2_meta.csvz9http://www.openslr.org/resources/49/voxceleb1_test_v2.txtSpeakerMetadataidnamegender nationalitysplitF voxceleb_name part_urls part_suffix dev_zip_name test_zip_name target_dirforce_downloadc Cst|}|jddd||}|r |s td|d|St}|D]} ttj tj | j j } t|}|| } t| | |dq't||d0} t||dD]} t| d}t|| Wd n1sqwYqYWd n1swY|d D] } t| |t| j q|d D] } t| |t| j qWd n1swYtd t| }||Wd n1swYtd t||}||Wd |S1swY|S)aN Download and unzip a VoxCeleb dataset ;param voxceleb_name :str, dataset name. :param part_urls: List[str], list of downloable links to zip partials. ;param part_suffix: str, dataset partial suffix. ;param dev_zip_name: str, name of concatenated dev zip file. ;param test_zip_name: str, name of concatenated test zip file. :param target_dir: Pathlike, the path of the dir to store the dataset. :param force_download: bool, if True, download the archive even if it already exists. :return: the path to downloaded and extracted directory with data. T)parentsexist_okz Skipping z because file exists.)filenamer&wb*rbNz*.zipz*.csvzUnzipping dev...zUnzipping test...)rmkdirexistslogginginfotempfileTemporaryDirectoryrurllibparseunquoteurlparsepathrropensortedglobshutil copyfileobjmovezipfileZipFile extractall)r r!r"r#r$r%r&zip_pathtemp_dirurl url_filenametemp_target_fileoutFilefileinFilezfrJK/home/ubuntu/.local/lib/python3.10/site-packages/lhotse/recipes/voxceleb.py_download_voxcelebUsV         rL.returnc Ctdtddd||dS)a~ Download and unzip the VoxCeleb1 data. .. note:: A "connection refused" error may occur if you are downloading without a password. :param target_dir: Pathlike, the path of the dir to store the dataset. :param force_download: bool, if True, download the archive even if it already exists. :return: the path to downloaded and extracted directory with data. VoxCeleb1vox1_dev_wav_partzvox1_dev_wav.zipzvox1_test_wav.zipr r!r"r#r$r%r&)rLVOXCELEB1_PARTS_URLr%r&rJrJrKdownload_voxceleb1rUc CrO)a~ Download and unzip the VoxCeleb2 data. .. note:: A "connection refused" error may occur if you are downloading without a password. :param target_dir: Pathlike, the path of the dir to store the dataset. :param force_download: bool, if True, download the archive even if it already exists. :return: the path to downloaded and extracted directory with data. VoxCeleb2vox2_dev_aac_partz vox2_aac.zipzvox2_test_aac.ziprR)rLVOXCELEB2_PARTS_URLrTrJrJrKdownload_voxceleb2rVrZvoxceleb1_rootvoxceleb2_root output_dirnum_jobsc Cs|rt|nd}|rt|nd}|s|std|dur t|nd}tt}|r?td|t|||t|dntd|rwtdt ||}d|vrst |dd|d|dd<t |dd |d |dd <n||d<d D]E}||vrqy||d}||d }t ||\}}t |||||d<|||d <|dur| |d |d | |d |d qy|durd|vrt|dD]\} } | |d| dd qd|vrt|dD]\} } | |d| dd q|S)a Prepare manifests for the VoxCeleb v1 and v2 corpora. The manifests are created in a dict with three splits: train, dev and test, for each of the two versions. Each split contains a RecordingSet and SupervisionSet in a dict under keys 'recordings' and 'supervisions'. :param voxceleb1_root: Path to the VoxCeleb v1 dataset. :param voxceleb2_root: Path to the VoxCeleb v2 dataset. :param output_dir: Path to the output directory. :param num_jobs: Number of parallel jobs to run. :return: A dict with standard corpus splits ("train" and "test") containing the manifests. NOTE: We prepare the data using the Kaldi style split, i.e., the whole VoxCeleb2 ("dev" and "test") and the training portion ("dev") of VoxCeleb1 are put into the "train" split. The "test" split contains the "test" portion of VoxCeleb1. So if VoxCeleb1 is not provided, no "test" split is created in the output manifests. Example usage: .. code-block:: python >>> from lhotse.recipes.voxceleb import prepare_voxceleb >>> manifests = prepare_voxceleb(voxceleb_v1_root='/path/to/voxceleb1', ... voxceleb_v2_root='/path/to/voxceleb2', ... output_dir='/path/to/output', ... num_jobs=4) NOTE: If VoxCeleb1 is provided, we also prepare the trials file using the list provided in http://www.openslr.org/resources/49/voxceleb1_test_v2.txt. This file is used in the Kaldi recipes for VoxCeleb speaker verification. This is prepared as 2 tuples of the form (CutSet, CutSet) with identical id's, one for each of positive pairs and negative pairs. These are stored in the dict under keys 'pos_trials' and 'neg_trials', respectively. For evaluation purpose, the :class:`lhotse.dataset.sampling.CutPairsSampler` can be used to sample from this tuple. Nz4Either VoxCeleb1 or VoxCeleb2 path must be provided.zPreparing VoxCeleb1...testzGVoxCeleb1 not provided, no test split or trials file will be created...zPreparing VoxCeleb2...train recordings supervisions)rar`voxceleb_recordings_z .jsonl.gzvoxceleb_supervisions_ pos_trialszvoxceleb_pos-trials_uttr[ neg_trialszvoxceleb_neg-trials_utt)r ValueErrorrdictr/r0update_prepare_voxceleb_v1_prepare_voxceleb_trials_prepare_voxceleb_v2rrrto_file enumerate) r\r]r^r_ manifests v2_manifestsrrbrcicutsrJrJrKprepare_voxcelebsZ*         rt corpus_pathc si}t|dd(}t||D]}|d\}}}}t||||d||<qWdn1s4wYt|K} g} g} g} |ddD] } | | t | |qKt t | t | dd d D]}| \}}| || |qdt| }t| }Wdn1swYtt}d D]'|fd d |d<dd|dD|fdd |d<q|d|d<|S)z Prepare manifests for the VoxCeleb1 corpus. The manifests are created in a dict with 2 splits: train ("dev") and test. z vox1_meta.csvr rNwavz*.wavzProcessing VoxCeleb1Ftotaldescleave)devr`cs|jdkS)Nr)custom)s)rrJrKHsz&_prepare_voxceleb_v1..rccSsg|]}|jqSrJ recording_id).0rrJrJrK Jsz(_prepare_voxceleb_v1..cs |jvS)N)r)rv) split_idsrJrKrLs rbr}ra)r8nextstriprrrrglobappendsubmit _process_filer rlenresultrfrom_recordingsr from_segmentsrrifilterpop)rur_speaker_metadataflinespkidrrrexrbrcfuturespfuture recording supervision recording_setsupervision_setrprJ)rrrKrk#sN           rkrpc Cs|d}|d}ggggf\}}}}ttddtdd}t|D]\}} | d\} } } d| dd d } d| dd d } | |vsR| |vratd |d | d | q"| dkr| t d||| d || j || d d| t d||| d || j || d dq"| t d||| d || j || d d| t d||| d || j || d dq"Wdn1swYt |t |ft |t |fdS)z; Prepare the trials file for the VoxCeleb1 corpus. rbrczvoxceleb_trials.txt)r)rv -rMr/zTrial z contains unknown recording: z or 1ztrial-)rrstartdurationrcchannelN)rfrg)rVOXCELEB1_TRIALS_URLr8rorrjoinr/warningrrrr from_cuts) rprbrc cuts_utt1_pos cuts_utt2_pos cuts_utt1_neg cuts_utt2_negridxrtargetutt1utt2rJrJrKrlRs     7rlc Cs0i}t|dd)}t||D]}ttj|d\}}}}t|d|d|d||<qWdn1s5wYt|A} g} g} g} |dD] } | | t | |qJt | t | d|d d d D]}|\}}| || |qeWdn1swYt| }t| }||d }|S) z Prepare manifests for the VoxCeleb2 corpus. The manifests are created the same dict without any splits since the whole data is used in the final "train" split. z vox2_meta.csvrv,rNz*.m4azProcessing VoxCeleb2 z split...Fry)rbrc)r8rmapstrrrrrrrrrr rrrrrr)rur_rrrr_rrrrbrcrrrrrrrrprJrJrKrmsB          rm file_pathrc Csz|jjj}|jj}|j}|d|d|}tj||d}t|||||jd|j||j||j||j dd}||fS)zT Process a single wav file and return a Recording and a SupervisionSegment. rrg) speaker_namerr)rrspeakerrrrr~) parentstemr from_filerrrrrr)rr speaker_id session_iduttidrrrrJrJrKrs$  r)F)rMF)NNNr[)6__doc__r/r;r1r3r> collectionsrrconcurrent.futuresrconcurrent.futures.processrpathlibrrtypingrr r r r tqdm.autor lhotserrrrrrlhotse.manipulationr lhotse.qarr lhotse.utilsrrrSrYrrrboolrLrUrZintrtrkrlrmrrJrJrJrKs         :   c / K *