|
1
|
# YouTube Processing |
|
2
|
|
|
3
|
How to extract transcripts, metadata, and audio from YouTube videos for the podcast marketing pipeline. |
|
4
|
|
|
5
|
## Table of Contents |
|
6
|
|
|
7
|
- [URL Handling](#url-handling) — Recognize and normalize YouTube URLs |
|
8
|
- [Caption Extraction](#caption-extraction) — Fetch auto-generated or manual subtitles (preferred, no API key) |
|
9
|
|
|
10
|
- [Audio Download & Transcription](#audio-download--transcription) — Fallback when no captions exist |
|
11
|
- [Metadata Extraction](#metadata-extraction) — Title, channel, description, duration |
|
12
|
|
|
13
|
- [Integration with Content Pipeline](#integration-with-content-pipeline) — How YouTube input feeds into the standard pipeline |
|
14
|
|
|
15
|
## URL Handling |
|
16
|
|
|
17
|
YouTube URLs come in several formats. Normalize to a video ID before processing: |
|
18
|
|
|
19
|
- `https://www.youtube.com/watch?v=VIDEO_ID` |
|
20
|
- `https://youtu.be/VIDEO_ID` |
|
21
|
|
|
22
|
- `https://www.youtube.com/embed/VIDEO_ID` |
|
23
|
- `https://m.youtube.com/watch?v=VIDEO_ID` |
|
24
|
|
|
25
|
- URLs with extra parameters: `&t=120`, `&list=PLxxx`, `&si=xxx` |
|
26
|
|
|
27
|
```python |
|
28
|
|
|
29
|
import re |
|
30
|
|
|
31
|
def extract_video_id(url: str) -> str | None: |
|
32
|
|
|
33
|
patterns = [ |
|
34
|
|
|
35
|
r"(?:youtube\\com/watch\\v=|youtu\\be/|youtube\\com/embed/)([a-zA-Z0-9_-]{11})", |
|
36
|
|
|
37
|
] |
|
38
|
|
|
39
|
for pattern in patterns: |
|
40
|
|
|
41
|
match = re.search(pattern, url) |
|
42
|
|
|
43
|
if match: |
|
44
|
|
|
45
|
return match.group(1) |
|
46
|
|
|
47
|
return None |
|
48
|
|
|
49
|
``` |
|
50
|
|
|
51
|
## Caption Extraction |
|
52
|
|
|
53
|
The preferred approach — fast, free, and doesn't require downloading the full audio. |
|
54
|
|
|
55
|
### Using youtube-transcript-api |
|
56
|
|
|
57
|
```bash |
|
58
|
|
|
59
|
pip install youtube-transcript-api |
|
60
|
|
|
61
|
``` |
|
62
|
|
|
63
|
```python |
|
64
|
|
|
65
|
from youtube_transcript_api import YouTubeTranscriptApi |
|
66
|
|
|
67
|
def get_youtube_transcript(video_id: str, languages: list[str] | None = None) -> dict: |
|
68
|
|
|
69
|
"""Fetch transcript from YouTube's caption system. |
|
70
|
|
|
71
|
Tries manually uploaded captions first, then auto-generated. |
|
72
|
|
|
73
|
Returns a normalized transcript dict compatible with the content pipeline. |
|
74
|
|
|
75
|
Args: |
|
76
|
|
|
77
|
video_id: The 11-character YouTube video ID. |
|
78
|
|
|
79
|
languages: Preferred languages in order. Defaults to ["en"]. |
|
80
|
|
|
81
|
Returns: |
|
82
|
|
|
83
|
dict with "text" (full transcript string) and "segments" |
|
84
|
|
|
85
|
(list of {start, end, text} dicts). |
|
86
|
|
|
87
|
Raises: |
|
88
|
|
|
89
|
Exception if no captions are available in any language. |
|
90
|
|
|
91
|
""" |
|
92
|
|
|
93
|
if languages is None: |
|
94
|
|
|
95
|
languages = ["en"] |
|
96
|
|
|
97
|
try: |
|
98
|
|
|
99
|
transcript_list = YouTubeTranscriptApi.list_transcripts(video_id) |
|
100
|
|
|
101
|
transcript = None |
|
102
|
|
|
103
|
try: |
|
104
|
|
|
105
|
transcript = transcript_list.find_manually_created_transcript(languages) |
|
106
|
|
|
107
|
except Exception: |
|
108
|
|
|
109
|
try: |
|
110
|
|
|
111
|
transcript = transcript_list.find_generated_transcript(languages) |
|
112
|
|
|
113
|
except Exception: |
|
114
|
|
|
115
|
pass |
|
116
|
|
|
117
|
if transcript is None: |
|
118
|
|
|
119
|
available = list(transcript_list) |
|
120
|
|
|
121
|
if available: |
|
122
|
|
|
123
|
transcript = available[0] |
|
124
|
|
|
125
|
if transcript.language_code not in languages: |
|
126
|
|
|
127
|
transcript = transcript.translate("en") |
|
128
|
|
|
129
|
if transcript is None: |
|
130
|
|
|
131
|
raise Exception(f"No captions available for video {video_id}") |
|
132
|
|
|
133
|
entries = transcript.fetch() |
|
134
|
|
|
135
|
segments = [] |
|
136
|
|
|
137
|
for entry in entries: |
|
138
|
|
|
139
|
segments.append({ |
|
140
|
|
|
141
|
"start": entry.start, |
|
142
|
|
|
143
|
"end": entry.start + entry.duration, |
|
144
|
|
|
145
|
"text": entry.text.strip(), |
|
146
|
|
|
147
|
}) |
|
148
|
|
|
149
|
full_text = " ".join(seg["text"] for seg in segments) |
|
150
|
|
|
151
|
return {"text": full_text, "segments": segments} |
|
152
|
|
|
153
|
except Exception as e: |
|
154
|
|
|
155
|
raise Exception(f"Failed to get transcript for {video_id}: {e}") |
|
156
|
|
|
157
|
``` |
|
158
|
|
|
159
|
### Caption Quality Notes |
|
160
|
|
|
161
|
- **Manually uploaded captions** are highest quality — proper punctuation, speaker labels sometimes included. |
|
162
|
- **Auto-generated captions** are decent for clear speech but may have errors with technical terms, names, and acronyms. The content generation pipeline should handle minor transcription errors gracefully. |
|
163
|
|
|
164
|
- **No captions available** — some videos have captions disabled. Fall back to audio download + Whisper transcription. |
|
165
|
|
|
166
|
## Audio Download & Transcription |
|
167
|
|
|
168
|
Fallback when YouTube captions are unavailable. Uses yt-dlp to download the audio track. |
|
169
|
|
|
170
|
### Setup |
|
171
|
|
|
172
|
```bash |
|
173
|
|
|
174
|
pip install yt-dlp |
|
175
|
|
|
176
|
``` |
|
177
|
|
|
178
|
yt-dlp is actively maintained and handles YouTube's frequent changes. Avoid using the deprecated `youtube-dl`. |
|
179
|
|
|
180
|
### Audio Download |
|
181
|
|
|
182
|
```python |
|
183
|
|
|
184
|
import yt_dlp |
|
185
|
|
|
186
|
import os |
|
187
|
|
|
188
|
def download_youtube_audio(video_id: str, output_dir: str = "/tmp") -> str: |
|
189
|
|
|
190
|
"""Download audio from a YouTube video. |
|
191
|
|
|
192
|
Returns the path to the downloaded audio file. |
|
193
|
|
|
194
|
""" |
|
195
|
|
|
196
|
os.makedirs(output_dir, exist_ok=True) |
|
197
|
|
|
198
|
output_path = os.path.join(output_dir, f"{video_id}.mp3") |
|
199
|
|
|
200
|
ydl_opts = { |
|
201
|
|
|
202
|
"format": "bestaudio/best", |
|
203
|
|
|
204
|
"postprocessors": [{ |
|
205
|
|
|
206
|
"key": "FFmpegExtractAudio", |
|
207
|
|
|
208
|
"preferredcodec": "mp3", |
|
209
|
|
|
210
|
"preferredquality": "128", |
|
211
|
|
|
212
|
}], |
|
213
|
|
|
214
|
"outtmpl": os.path.join(output_dir, f"{video_id}.%(ext)s"), |
|
215
|
|
|
216
|
"quiet": True, |
|
217
|
|
|
218
|
"no_warnings": True, |
|
219
|
|
|
220
|
} |
|
221
|
|
|
222
|
with yt_dlp.YoutubeDL(ydl_opts) as ydl: |
|
223
|
|
|
224
|
ydl.download([f"https://www.youtube.com/watch?v={video_id}"]) |
|
225
|
|
|
226
|
return output_path |
|
227
|
|
|
228
|
``` |
|
229
|
|
|
230
|
After downloading, pass the audio file to `whisper_transcribe.transcribe_episode()` from the existing transcription script. |
|
231
|
|
|
232
|
## Metadata Extraction |
|
233
|
|
|
234
|
Extract video metadata without downloading the full video: |
|
235
|
|
|
236
|
```python |
|
237
|
|
|
238
|
import yt_dlp |
|
239
|
|
|
240
|
def get_youtube_metadata(video_id: str) -> dict: |
|
241
|
|
|
242
|
"""Extract metadata from a YouTube video. |
|
243
|
|
|
244
|
Returns title, channel, description, duration, publish date, |
|
245
|
|
|
246
|
and thumbnail URL. |
|
247
|
|
|
248
|
""" |
|
249
|
|
|
250
|
ydl_opts = { |
|
251
|
|
|
252
|
"quiet": True, |
|
253
|
|
|
254
|
"no_warnings": True, |
|
255
|
|
|
256
|
"skip_download": True, |
|
257
|
|
|
258
|
} |
|
259
|
|
|
260
|
with yt_dlp.YoutubeDL(ydl_opts) as ydl: |
|
261
|
|
|
262
|
info = ydl.extract_info( |
|
263
|
|
|
264
|
f"https://www.youtube.com/watch?v={video_id}", |
|
265
|
|
|
266
|
download=False, |
|
267
|
|
|
268
|
) |
|
269
|
|
|
270
|
return { |
|
271
|
|
|
272
|
"title": info.get("title", "Untitled"), |
|
273
|
|
|
274
|
"channel": info.get("channel", info.get("uploader", "Unknown")), |
|
275
|
|
|
276
|
"description": info.get("description", ""), |
|
277
|
|
|
278
|
"duration_seconds": info.get("duration", 0), |
|
279
|
|
|
280
|
"publish_date": info.get("upload_date", ""), |
|
281
|
|
|
282
|
"thumbnail": info.get("thumbnail", ""), |
|
283
|
|
|
284
|
"view_count": info.get("view_count", 0), |
|
285
|
|
|
286
|
"url": f"https://www.youtube.com/watch?v={video_id}", |
|
287
|
|
|
288
|
"tags": info.get("tags", []), |
|
289
|
|
|
290
|
} |
|
291
|
|
|
292
|
``` |
|
293
|
|
|
294
|
## Integration with Content Pipeline |
|
295
|
|
|
296
|
When processing a YouTube link, the output maps to the standard pipeline like this: |
|
297
|
|
|
298
|
1. **Episode metadata** comes from YouTube video metadata (title, channel as host/guest, description, publish date). |
|
299
|
2. **Transcript** comes from captions or Whisper transcription — same format as other input modes. |
|
300
|
|
|
301
|
3. **Episode link** in generated content should point to the YouTube video URL instead of a podcast player link. |
|
302
|
4. **Guest name** — try extracting from the video title (common patterns: "Interview with X", "X on Y", "feat. X") or from the video description. Falls back to the channel name. |
|
303
|
|
|
304
|
The content pipeline doesn't need to know the source was YouTube — once you have a transcript and metadata dict, everything works identically. |
|
305
|
|
|
306
|
### Recommended Workflow |
|
307
|
|
|
308
|
```python |
|
309
|
|
|
310
|
from youtube_extractor import extract_video_id, get_youtube_metadata |
|
311
|
|
|
312
|
video_id = extract_video_id(user_provided_url) |
|
313
|
|
|
314
|
metadata = get_youtube_metadata(video_id) |
|
315
|
|
|
316
|
try: |
|
317
|
|
|
318
|
transcript = get_youtube_transcript(video_id) |
|
319
|
|
|
320
|
except Exception: |
|
321
|
|
|
322
|
audio_path = download_youtube_audio(video_id) |
|
323
|
|
|
324
|
transcript = transcribe_episode(audio_path) |
|
325
|
|
|
326
|
# Now pass transcript + metadata to the content generation pipeline |
|
327
|
|
|
328
|
``` |