auto-whisper-safe_skill

Overview

This skill provides RAM-safe voice transcription using OpenAI Whisper with automatic chunking to avoid out-of-memory crashes on 16GB machines. It chooses a low-memory default model, splits long audio into 10-minute segments, and merges transcripts seamlessly. Temp files are cleaned up automatically and model/language are configurable via environment variables.

How this skill works

The tool inspects audio length with ffprobe and decides whether to transcribe directly or split into 10-minute chunks with ffmpeg. Each chunk is transcribed independently using the configured Whisper model (default: base) and the resulting segments are concatenated in order with no gaps or duplicates. It reports chunk-by-chunk progress, handles various input formats, and removes temporary files even on errors.

When to use it

• Transcribing voice messages or long recordings on machines with ~16GB RAM
• Batch processing podcasts or interviews without risking OOM crashes
• Integrating reliable transcription into agents that run other heavy services
• When you need flexible model selection for trade-offs between accuracy and RAM
• Handling many formats (ogg, mp3, m4a, wav, webm, opus) via ffmpeg

Best practices

• Keep WHISPER_MODEL=base for a safe balance of RAM (~1.5GB) and accuracy
• Use WHISPER_LANG to force language detection when audio is single-language
• Run the script from an agent bootstrap to transcribe incoming audio automatically
• Monitor progress for long jobs; model change to small/medium only on higher-RAM machines
• Ensure ffmpeg and openai-whisper are installed and up to date

Example use cases

• Transcribe Telegram voice messages (.ogg) in an agent pipeline before responding
• Process an hour-long podcast by auto-splitting into 10-minute chunks and producing one transcript
• Run quick previews with tiny model when speed and minimal RAM use matter
• Integrate into a multi-agent system so other services remain stable during transcription