diff --git a/yt_dlp/extractor/slideslive.py b/yt_dlp/extractor/slideslive.py index 9a60a79e7..86c26a8a2 100644 --- a/yt_dlp/extractor/slideslive.py +++ b/yt_dlp/extractor/slideslive.py @@ -1,92 +1,176 @@ from .common import InfoExtractor from ..utils import ( - bool_or_none, smuggle_url, - try_get, + traverse_obj, + unified_timestamp, url_or_none, ) class SlidesLiveIE(InfoExtractor): _VALID_URL = r'https?://slideslive\.com/(?P[0-9]+)' - _WORKING = False _TESTS = [{ - # video_service_name = YOUTUBE + # service_name = yoda 'url': 'https://slideslive.com/38902413/gcc-ia16-backend', - 'md5': 'b29fcd6c6952d0c79c5079b0e7a07e6f', 'info_dict': { - 'id': 'LMtgR8ba0b0', + 'id': '38902413', 'ext': 'mp4', 'title': 'GCC IA16 backend', - 'description': 'Watch full version of this video at https://slideslive.com/38902413.', - 'uploader': 'SlidesLive Videos - A', - 'uploader_id': 'UC62SdArr41t_-_fX40QCLRw', - 'timestamp': 1597615266, - 'upload_date': '20170925', - } + 'timestamp': 1648189972, + 'upload_date': '20220325', + 'thumbnail': r're:^https?://.*\.jpg', + }, + 'params': { + 'skip_download': 'm3u8', + }, }, { - # video_service_name = yoda + # service_name = yoda 'url': 'https://slideslive.com/38935785', - 'md5': '575cd7a6c0acc6e28422fe76dd4bcb1a', 'info_dict': { - 'id': 'RMraDYN5ozA_', + 'id': '38935785', 'ext': 'mp4', 'title': 'Offline Reinforcement Learning: From Algorithms to Practical Challenges', + 'upload_date': '20211115', + 'timestamp': 1636996003, + 'thumbnail': r're:^https?://.*\.jpg', + }, + 'params': { + 'skip_download': 'm3u8', + }, + }, { + # service_name = yoda + 'url': 'https://slideslive.com/38973182/how-should-a-machine-learning-researcher-think-about-ai-ethics', + 'info_dict': { + 'id': '38973182', + 'ext': 'mp4', + 'title': 'How Should a Machine Learning Researcher Think About AI Ethics?', + 'upload_date': '20220201', + 'thumbnail': r're:^https?://.*\.jpg', + 'timestamp': 1643728135, + }, + 'params': { + 'skip_download': 'm3u8', }, }, { - # video_service_name = youtube + # service_name = youtube + 'url': 'https://slideslive.com/38897546/special-metaprednaska-petra-ludwiga-hodnoty-pro-lepsi-spolecnost', + 'md5': '8a79b5e3d700837f40bd2afca3c8fa01', + 'info_dict': { + 'id': 'jmg02wCJD5M', + 'display_id': '38897546', + 'ext': 'mp4', + 'title': 'SPECIÁL: Meta-přednáška Petra Ludwiga - Hodnoty pro lepší společnost', + 'description': 'Watch full version of this video at https://slideslive.com/38897546.', + 'channel_url': 'https://www.youtube.com/channel/UCZWdAkNYFncuX0khyvhqnxw', + 'channel': 'SlidesLive Videos - G1', + 'channel_id': 'UCZWdAkNYFncuX0khyvhqnxw', + 'uploader_id': 'UCZWdAkNYFncuX0khyvhqnxw', + 'uploader': 'SlidesLive Videos - G1', + 'uploader_url': 'http://www.youtube.com/channel/UCZWdAkNYFncuX0khyvhqnxw', + 'live_status': 'not_live', + 'upload_date': '20160710', + 'timestamp': 1618786715, + 'duration': 6827, + 'like_count': int, + 'view_count': int, + 'comment_count': int, + 'channel_follower_count': int, + 'age_limit': 0, + 'thumbnail': r're:^https?://.*\.jpg', + 'playable_in_embed': True, + 'availability': 'unlisted', + 'tags': [], + 'categories': ['People & Blogs'], + }, + }, { + # service_name = youtube 'url': 'https://slideslive.com/38903721/magic-a-scientific-resurrection-of-an-esoteric-legend', 'only_matching': True, }, { - # video_service_name = url + # service_name = url 'url': 'https://slideslive.com/38922070/learning-transferable-skills-1', 'only_matching': True, }, { - # video_service_name = vimeo + # service_name = vimeo 'url': 'https://slideslive.com/38921896/retrospectives-a-venue-for-selfreflection-in-ml-research-3', 'only_matching': True, }] + def _extract_custom_m3u8_info(self, m3u8_data): + m3u8_dict = {} + + lookup = { + 'PRESENTATION-TITLE': 'title', + 'PRESENTATION-UPDATED-AT': 'timestamp', + 'PRESENTATION-THUMBNAIL': 'thumbnail', + 'PLAYLIST-TYPE': 'playlist_type', + 'VOD-VIDEO-SERVICE-NAME': 'service_name', + 'VOD-VIDEO-ID': 'service_id', + 'VOD-VIDEO-SERVERS': 'video_servers', + 'VOD-SUBTITLES': 'subtitles', + } + + for line in m3u8_data.splitlines(): + if not line.startswith('#EXT-SL-'): + continue + tag, _, value = line.partition(':') + key = lookup.get(tag.lstrip('#EXT-SL-')) + if not key: + continue + m3u8_dict[key] = value + + # Some values are stringified JSON arrays + for key in ('video_servers', 'subtitles'): + if key in m3u8_dict: + m3u8_dict[key] = self._parse_json(m3u8_dict[key], None, fatal=False) or [] + + return m3u8_dict + def _real_extract(self, url): video_id = self._match_id(url) - video_data = self._download_json( - 'https://ben.slideslive.com/player/' + video_id, video_id) - service_name = video_data['video_service_name'].lower() + webpage = self._download_webpage(url, video_id) + player_token = self._search_regex(r'data-player-token="([^"]+)"', webpage, 'player token') + player_data = self._download_webpage( + f'https://ben.slideslive.com/player/{video_id}', video_id, + note='Downloading player info', query={'player_token': player_token}) + player_info = self._extract_custom_m3u8_info(player_data) + + service_name = player_info['service_name'].lower() assert service_name in ('url', 'yoda', 'vimeo', 'youtube') - service_id = video_data['video_service_id'] + service_id = player_info['service_id'] + subtitles = {} - for sub in try_get(video_data, lambda x: x['subtitles'], list) or []: - if not isinstance(sub, dict): - continue + for sub in traverse_obj(player_info, ('subtitles', ...), expected_type=dict): webvtt_url = url_or_none(sub.get('webvtt_url')) if not webvtt_url: continue - lang = sub.get('language') or 'en' - subtitles.setdefault(lang, []).append({ + subtitles.setdefault(sub.get('language') or 'en', []).append({ 'url': webvtt_url, + 'ext': 'vtt', }) + info = { 'id': video_id, - 'thumbnail': video_data.get('thumbnail'), - 'is_live': bool_or_none(video_data.get('is_live')), + 'title': player_info.get('title') or self._html_search_meta('title', webpage, default=''), + 'timestamp': unified_timestamp(player_info.get('timestamp')), + 'is_live': player_info.get('playlist_type') != 'vod', + 'thumbnail': url_or_none(player_info.get('thumbnail')), 'subtitles': subtitles, } + if service_name in ('url', 'yoda'): - info['title'] = video_data['title'] if service_name == 'url': info['url'] = service_id else: + cdn_hostname = player_info['video_servers'][0] formats = [] - _MANIFEST_PATTERN = 'https://01.cdn.yoda.slideslive.com/%s/master.%s' - # use `m3u8` entry_protocol until EXT-X-MAP is properly supported by `m3u8_native` entry_protocol formats.extend(self._extract_m3u8_formats( - _MANIFEST_PATTERN % (service_id, 'm3u8'), - service_id, 'mp4', m3u8_id='hls', fatal=False)) + f'https://{cdn_hostname}/{service_id}/master.m3u8', + video_id, 'mp4', m3u8_id='hls', fatal=False, live=True)) formats.extend(self._extract_mpd_formats( - _MANIFEST_PATTERN % (service_id, 'mpd'), service_id, - mpd_id='dash', fatal=False)) + f'https://{cdn_hostname}/{service_id}/master.mpd', + video_id, mpd_id='dash', fatal=False)) info.update({ - 'id': service_id, 'formats': formats, }) else: @@ -94,10 +178,11 @@ class SlidesLiveIE(InfoExtractor): '_type': 'url_transparent', 'url': service_id, 'ie_key': service_name.capitalize(), - 'title': video_data.get('title'), + 'display_id': video_id, }) if service_name == 'vimeo': info['url'] = smuggle_url( - 'https://player.vimeo.com/video/' + service_id, + f'https://player.vimeo.com/video/{service_id}', {'http_headers': {'Referer': url}}) + return info