transcribe

Transcribe - Audio/Video Processing Pipeline

CLI: npx @krasnoperov/transcribe <command> [args] [options]

Core Commands

Four operations for audio/video content processing:

transcribe <input>           # Audio/Video -> VTT transcript with speaker diarization
summarize <input>            # Text/VTT -> Markdown summary
infographic <input>          # Text -> Visual infographic image
process <input>              # All-in-one pipeline: video -> transcript -> summary -> infographic

When to Use

• Transcribe meeting recordings, podcasts, or interviews
• Generate speaker-attributed transcripts (who said what)
• Create summaries from long transcripts or documents
• Generate visual infographics from text content
• Process entire videos end-to-end in one command

Prerequisites

# Required: ffmpeg for audio processing
brew install ffmpeg  # macOS
apt install ffmpeg   # Linux

# API keys
export OPENAI_API_KEY="your-key"        # Get at https://platform.openai.com/api-keys
export GOOGLE_AI_STUDIO_KEY="your-key"  # Get at https://ai.google.dev/

Quick Examples

# Transcribe a video with speaker diarization (OpenAI, default)
npx @krasnoperov/transcribe transcribe meeting.mp4 -o transcript.vtt

# Transcribe with Gemini (good for long audio, up to ~8 hours)
npx @krasnoperov/transcribe transcribe podcast.mp3 --model gemini-3 -o transcript.vtt

# Generate summary from transcript
npx @krasnoperov/transcribe summarize transcript.vtt -o summary.md

# Create infographic from summary
npx @krasnoperov/transcribe infographic summary.md --style "modern minimal" -o visual.png

# All-in-one: process entire video
npx @krasnoperov/transcribe process recording.mp4 --language en --output-dir ./output

Transcription Options

--model <model>              Transcription model:
                             OpenAI models:
                             - gpt-4o-transcribe-diarize (default, with speakers)
                             - gpt-4o-transcribe (no speakers)
                             - whisper-1 (legacy)
                             Google models:
                             - gemini-3 (with speakers, handles long audio up to ~8h)
--language <lang>            Language code (e.g., en, es, ru, de)
-o, --output <file>          Output VTT file path

Summarization Options

--prompt <text>              Custom summarization instructions
-o, --output <file>          Output markdown file path

Infographic Options

--style <text>               Style instructions (e.g., "artistic", "corporate", "playful")
--reference <image>          Reference image for visual style
-o, --output <file>          Output image file path

Process (All-in-One) Options

--output-dir <dir>           Output directory for all files
--language <lang>            Language code for transcription
--model <model>              Transcription model
--style <text>               Style for infographic

Output Formats

• Transcription: WebVTT (.vtt) with speaker tags <v Speaker 1>text</v>
• Summary: Markdown (.md) with headers and bullet points
• Infographic: PNG image with visual summary

Audio Processing Notes

• Supports video formats: MP4, MKV, MOV, AVI, WebM
• Supports audio formats: WAV, MP3, M4A, OGG, FLAC, AAC
• Long files (>23 min) are automatically split into chunks
• Temporary files are cleaned up after processing

AI Models Used

• Transcription (OpenAI): gpt-4o-transcribe-diarize (speaker identification)
• Transcription (Google): gemini-2.5-flash via gemini-3 model option (speaker identification, long audio support)
• Summarization: OpenAI gpt-5.1 with reasoning
• Infographics: Google Gemini gemini-3-pro-image-preview