speech-to-text-transcription_skill

Overview

This skill transcribes audio and video files to text with speaker detection, timestamps, and format conversion. It supports local and cloud providers, handles long files by splitting, and stores transcripts in a local memory folder for later use. Designed for meetings, interviews, podcasts, voice memos, and lectures.

How this skill works

The agent detects the input type (local file, URL, or recording), preprocesses audio with ffmpeg as needed, and selects an appropriate provider based on context (local Whisper, OpenAI, AssemblyAI, or Deepgram). For long or noisy recordings it splits and cleans audio, runs transcription per chunk, merges results with timestamps and speaker labels, then offers output formats and save options.

When to use it

• Transcribing meeting recordings or multi-speaker interviews needing diarization
• Creating subtitles (.srt/.vtt) for video content
• Quick, private local transcription without API keys
• High-accuracy cloud transcription for noisy or critical audio
• Processing long files by splitting and batching to avoid timeouts

Best practices

• Detect the file type first and verify local paths or download URLs to a temp workspace
• Use local Whisper for privacy and quick jobs; use AssemblyAI for native speaker diarization
• Split files >25MB or >2 hours into chunks with ffmpeg, then merge transcripts with timestamps
• Preprocess noisy audio (noise reduction, resampling) before transcription
• Store transcripts in the local memory folder and ask the user before uploading audio to cloud providers

Example use cases

• Convert a Zoom meeting recording into a speaker-labeled transcript and action-item summary
• Generate .srt subtitles for a lecture video and a plain-text transcript for notes
• Transcribe a podcast episode with timestamps and export structured JSON including word-level timing
• Quickly transcribe a voice memo locally with Whisper and save a concise .md formatted version