Speech Transcriber

Speech Transcriber

This skill transcribes audio from videos into text with accurate timestamps, essential for finding highlights and generating captions.

When to Use This Skill

• After downloading a YouTube video
• Need to search for specific content in a video
• Preparing to select clips for shorts
• Generating captions/subtitles

What This Skill Does

1. Extract Audio: Separates audio track from video if needed
2. Transcribe Speech: Converts speech to text with timestamps
3. Generate Formats: Creates TXT, SRT, and JSON outputs
4. Detect Language: Auto-detects spoken language

How to Use

Basic Transcription

Transcribe the audio from: ./output/downloads/VIDEO_ID.mp4

With Language Specification

Transcribe in Chinese: ./output/downloads/VIDEO_ID.mp4

Instructions

When transcribing audio:

Step 1: Extract Audio (if from video)

ffmpeg -i "./output/downloads/VIDEO_ID.mp4" \
  -vn -acodec libmp3lame -q:a 2 \
  "./temp/VIDEO_ID_audio.mp3"

Step 2: Transcribe Using Whisper

Option A: Using OpenAI Whisper API (if available)

# This would use API - implementation depends on available tools

Option B: Using Local Whisper

whisper "./temp/VIDEO_ID_audio.mp3" \
  --model small \
  --output_format all \
  --output_dir "./output/transcripts/"

Option C: Using yt-dlp Auto-Subtitles (fallback)

# Download existing subtitles if available
yt-dlp --write-auto-sub --sub-lang en,zh \
  --skip-download \
  -o "./output/transcripts/%(id)s" \
  "[ORIGINAL_URL]"

Step 3: Convert to Standard Formats

If needed, convert VTT to SRT:

ffmpeg -i "./output/transcripts/VIDEO_ID.vtt" \
  "./output/transcripts/VIDEO_ID.srt"

Step 4: Create JSON with Timestamps

Parse the SRT and create a structured JSON:

{
  "video_id": "VIDEO_ID",
  "duration": 600,
  "language": "en",
  "segments": [
    {
      "id": 1,
      "start": 0.0,
      "end": 4.5,
      "text": "Hello everyone, welcome to today's video"
    },
    {
      "id": 2,
      "start": 4.5,
      "end": 8.2,
      "text": "Today we're going to talk about..."
    }
  ]
}

Step 5: Report to User

## ✅ Transcription Complete

**Language Detected**: English
**Total Duration**: 10:32
**Segments**: 156 segments transcribed

**Files Generated**:
- Text: ./output/transcripts/VIDEO_ID.txt
- Subtitles: ./output/transcripts/VIDEO_ID.srt
- Structured: ./output/transcripts/VIDEO_ID.json

**Preview** (first 500 characters):
[Show beginning of transcript]

Next steps:
1. Use `highlight-finder` to identify the best clips
2. Search for specific topics in the transcript
3. Tell me which segments interest you

Output

• Plain text: ./output/transcripts/{video_id}.txt
• SRT subtitles: ./output/transcripts/{video_id}.srt
• JSON with timestamps: ./output/transcripts/{video_id}.json

Tips

• For Chinese content, specify language for better accuracy
• Longer videos take more time to process
• Check transcript for errors before generating captions