Implementation of automatic transcription of lectures and webinars. Transcription of educational content - lecture notes, text versions of webinars, search through course recordings. Specifics: one main speaker (lecturer), slides and screen sharing are possible, academic vocabulary. ### A simple solution for a single lecturer
from faster_whisper import WhisperModel
from openai import AsyncOpenAI
model = WhisperModel("large-v3", device="cuda")
client = AsyncOpenAI()
async def transcribe_lecture(
video_path: str,
lecture_topic: str = None
) -> dict:
# Извлекаем аудио
audio_path = extract_audio(video_path)
# Транскрибируем
segments, info = model.transcribe(
audio_path,
language="ru",
initial_prompt=f"Лекция на тему: {lecture_topic}. " if lecture_topic else None,
vad_filter=True
)
full_text = " ".join(seg.text for seg in segments)
# Структурируем через LLM
structure = await client.chat.completions.create(
model="gpt-4o",
messages=[{
"role": "system",
"content": """Обработай транскрипт лекции:
1. Исправь очевидные ошибки распознавания
2. Раздели на логические разделы с заголовками H2
3. Выдели ключевые термины жирным
4. Добавь список ключевых понятий в конце
Формат: Markdown."""
}, {
"role": "user",
"content": full_text[:8000] # ограничение контекста
}]
)
return {
"raw_transcript": full_text,
"structured_notes": structure.choices[0].message.content,
"duration_minutes": info.duration / 60,
"language": info.language
}
```### Processing long lectures (2+ hours) We break them into 20-30 minute sections, process them in parallel, and glue them together taking into account the context:```python
async def process_long_lecture(audio_path: str, chunk_minutes: int = 25) -> str:
chunks = split_audio(audio_path, chunk_minutes * 60)
transcripts = await asyncio.gather(
*[transcribe_chunk(chunk) for chunk in chunks]
)
return merge_transcripts(transcripts)
```### Uploading to platforms - Notion API — automatic creation of a page with notes - Google Docs API — export to Drive - LMS (Moodle, Canvas) — upload as course material Timeframe: transcription + structuring of one lecture — 1 day. Automated pipeline for the series — 1 week.