1
0
mirror of https://github.com/l1ving/youtube-dl synced 2024-11-22 06:49:31 +08:00
youtube-dl/youtube_dl/extractor/radiobremen.py

64 lines
2.4 KiB
Python

# -*- coding: utf-8 -*-
from __future__ import unicode_literals
import re
from .common import InfoExtractor
from ..utils import parse_duration
class RadioBremenIE(InfoExtractor):
_VALID_URL = r'http?://(?:www\.)?radiobremen\.de/mediathek/(?:index\.html)?\?id=(?P<id>[0-9]+)'
IE_NAME = 'radiobremen'
_TEST = {
'url': 'http://www.radiobremen.de/mediathek/index.html?id=114720',
'info_dict': {
'id': '114720',
'ext': 'mp4',
'duration': 1685,
'width': 512,
'title': 'buten un binnen vom 22. Dezember',
'thumbnail': 're:https?://.*\.jpg$',
'description': 'Unter anderem mit diesen Themen: 45 Flüchtlinge sind in Worpswede angekommen +++ Freies Internet für alle: Bremer arbeiten an einem flächendeckenden W-Lan-Netzwerk +++ Aktivisten kämpfen für das Unibad +++ So war das Wetter 2014 +++',
},
}
def _real_extract(self, url):
video_id = self._match_id(url)
meta_url = "http://www.radiobremen.de/apps/php/mediathek/metadaten.php?id=%s" % video_id
meta_doc = self._download_webpage(
meta_url, video_id, 'Downloading metadata')
title = self._html_search_regex(
r"<h1.*>(?P<title>.+)</h1>", meta_doc, "title")
description = self._html_search_regex(
r"<p>(?P<description>.*)</p>", meta_doc, "description", fatal=False)
duration = parse_duration(self._html_search_regex(
r"L&auml;nge:</td>\s+<td>(?P<duration>[0-9]+:[0-9]+)</td>",
meta_doc, "duration", fatal=False))
page_doc = self._download_webpage(
url, video_id, 'Downloading video information')
mobj = re.search(
r"ardformatplayerclassic\(\'playerbereich\',\'(?P<width>[0-9]+)\',\'.*\',\'(?P<video_id>[0-9]+)\',\'(?P<secret>[0-9]+)\',\'(?P<thumbnail>.+)\',\'\'\)",
page_doc)
video_url = (
"http://dl-ondemand.radiobremen.de/mediabase/%s/%s_%s_%s.mp4" %
(video_id, video_id, mobj.group("secret"), mobj.group('width')))
formats = [{
'url': video_url,
'ext': 'mp4',
'width': int(mobj.group("width")),
}]
return {
'id': video_id,
'title': title,
'description': description,
'duration': duration,
'formats': formats,
'thumbnail': mobj.group('thumbnail'),
}