Speech to Text Transcription

Setup

On first use, read setup.md and start helping with transcription needs.

When to Use

User has audio or video files that need transcription. Agent handles local files, URLs, voice memos, podcasts, interviews, meetings, and lectures.

Architecture

Memory lives in ~/speech-to-text-transcription/. See memory-template.md for structure.

~/speech-to-text-transcription/
├── memory.md        # Provider preferences, defaults
├── transcripts/     # Saved transcriptions
└── temp/            # Processing workspace

Quick Reference

Topic	File
Setup process	setup.md
Memory template	memory-template.md

Core Rules

1. Detect File Type First

Before transcription, identify the input:

• Local file path → verify exists, check format
• URL → download to temp, then process
• Meeting recording → likely needs speaker diarization
• Voice memo → usually single speaker, shorter

2. Choose Provider Based on Context

Scenario	Best Provider	Why
Quick local transcription	Whisper (local)	No API key, free, private
High accuracy needed	OpenAI Whisper API	Best quality
Speaker identification	AssemblyAI	Native diarization
Real-time/streaming	Deepgram	Low latency
Long content (>2 hours)	Split + batch	Avoid timeouts

3. Handle Long Audio

Files over 25MB or 2 hours:

1. Split into chunks (use ffmpeg)
2. Process each chunk
3. Merge transcripts with proper timestamps
4. Never attempt single upload for large files

4. Preserve Context

After transcription:

• Ask if user wants the transcript saved
• Suggest filename based on content
• Offer to extract action items or summary

5. Output Formats

Default to plain text. Offer alternatives:

• .txt — clean text, no timestamps
• .srt / .vtt — subtitles with timing
• .json — structured with word-level timing
• .md — formatted with speaker labels

Common Traps

• Assuming one provider works for all → Whisper fails on diarization, AssemblyAI needs API key
• Uploading huge files directly → Timeouts, memory errors. Split first.
• Ignoring audio quality → Noisy audio needs preprocessing (ffmpeg noise reduction)
• Not checking language → Whisper auto-detects but can fail on mixed-language content
• Losing speaker context → Multi-speaker content without diarization becomes unusable

Requirements

Required: ffmpeg (for audio processing)

Optional API keys (only if using cloud providers):

• OPENAI_API_KEY — for OpenAI Whisper API
• ASSEMBLYAI_API_KEY — for AssemblyAI (speaker diarization)
• DEEPGRAM_API_KEY — for Deepgram (real-time)

Local Whisper works without any API keys.

Provider Quick Reference

Local Whisper (No API Key)

# Install
pip install openai-whisper

# Basic transcription
whisper audio.mp3 --model base --output_format txt

# With timestamps
whisper audio.mp3 --model medium --output_format srt

Models: tiny (fast) → base → small → medium → large (accurate)

OpenAI Whisper API

curl -X POST https://api.openai.com/v1/audio/transcriptions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: multipart/form-data" \
  -F file="@audio.mp3" \
  -F model="whisper-1"

AssemblyAI (Speaker Diarization)

# Upload
curl -X POST https://api.assemblyai.com/v2/upload \
  -H "Authorization: $ASSEMBLYAI_API_KEY" \
  --data-binary @audio.mp3

# Transcribe with speakers
curl -X POST https://api.assemblyai.com/v2/transcript \
  -H "Authorization: $ASSEMBLYAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"audio_url": "URL", "speaker_labels": true}'

Audio Preprocessing

Extract Audio from Video

ffmpeg -i video.mp4 -vn -acodec pcm_s16le -ar 16000 -ac 1 audio.wav

Reduce Noise

ffmpeg -i noisy.wav -af "afftdn=nf=-25" clean.wav

Split Long Audio

# Split into 10-minute chunks
ffmpeg -i long.mp3 -f segment -segment_time 600 -c copy chunk_%03d.mp3

Security & Privacy

Data that stays local:

• Transcripts in ~/speech-to-text-transcription/transcripts/
• Local Whisper processes entirely on-device

Data that leaves your machine (if using APIs):

• Audio file sent to chosen provider (OpenAI, AssemblyAI, Deepgram)
• Transcript returned and stored locally

This skill does NOT:

• Store API keys in plain text (use environment variables)
• Auto-upload without confirmation
• Retain files on external servers after processing

External Endpoints

Endpoint	Data Sent	Purpose
api.openai.com/v1/audio	Audio file	Whisper API transcription
api.assemblyai.com/v2	Audio file	AssemblyAI transcription
api.deepgram.com/v1	Audio stream	Deepgram transcription

Only called when user explicitly chooses cloud provider. Local Whisper sends nothing.

Trust

By using cloud transcription providers, audio data is sent to OpenAI, AssemblyAI, or Deepgram. Only install if you trust these services with your audio. For sensitive content, use local Whisper.

Related Skills

Install with clawhub install <slug> if user confirms:

• audio — General audio processing
• ffmpeg — Video and audio conversion
• podcast — Podcast creation and editing

Feedback

• If useful: clawhub star speech-to-text-transcription
• Stay updated: clawhub sync