Audio Transcriber Pro

Transform audio recordings into professional Markdown documentation with intelligent summaries using LLM integration

Security Scan

The skill generally matches its stated transcription purpose, but there are multiple inconsistencies and scope creeps (LLM/cloud integration despite '100% local' claims, meta.json requiring OpenAI, runtime auto-installs, and filesystem/CLI probing) that warrant caution before installing or running it.

Purpose & Capability

The skill's stated goal (local Whisper/Faster-Whisper transcription to Markdown) is reasonable and matches most of the code, but metadata is inconsistent: meta.json lists a requirement of "openai" (not used in code), README/SKILL.md claim "100% local Whisper processing, no cloud uploads" while the code and docs explicitly include optional LLM integration (Claude CLI, GitHub Copilot CLI) which will transmit transcripts to external services. These contradictions make it unclear whether sensitive audio will remain local.

Instruction Scope

Runtime instructions and scripts do more than pure transcription: they detect and invoke external CLIs (claude, gh copilot), call a prompt-engineer skill (via gh copilot or by reading ~/.copilot/skills/prompt-engineer/SKILL.md), accept remote URLs (download audio), and can send transcript text to third‑party CLIs/LLMs. The skill also probes other skill files in the user's home directory. Any LLM/CLI invocation will likely transmit transcript contents off-host—this behavior is not consistent with a strict "local only" privacy promise.

Install Mechanism

No registry install spec is declared (instruction-only), but the bundle includes an install script (scripts/install-requirements.sh) and the Python implementation will attempt to pip-install missing packages at runtime (rich, tqdm) using subprocess.run. The install script downloads/installs faster-whisper or openai-whisper and can optionally download models (over the network). This is moderately risky but typical for CLI tools; users should review the install script and consent before running.

Credentials

The registry metadata (_meta.json) claims an external requirement of "openai" and pricing info, but the code does not use the OpenAI HTTP API; instead it relies on local Whisper/Faster-Whisper and optional CLIs (claude, gh). The skill declares no required env vars in registry metadata, but the README and SKILL.md mention optional cloud providers and CLI tools. The mismatch between declared requirements and actual behavior is a red flag (unnecessary/misleading requirements).

Persistence & Privilege

The skill does not request always:true and doesn't modify other skills' configurations in the repo. However it will perform local installs (pip --user / break-system-packages) and may download large model files into the environment, and it reads other skill files (e.g., ~/.copilot/skills/prompt-engineer/SKILL.md) to detect integration—this is limited privilege but should be accepted consciously by the user.

What to consider before installing

This skill is not outright malicious, but proceed carefully: - Privacy: If you expect strictly local processing, disable/decline any LLM/CLI steps. The skill can send transcript text to external CLIs (Claude/Copilot), which may upload your audio/transcript to cloud services. - Review installs: Inspect scripts/install-requirements.sh and scripts/transcribe.py before running. They will run pip installs and may download Whisper/Faster-Whisper models and other packages automatically. - Metadata mismatches: _meta.json claims an OpenAI requirement and paid pricing that do not match the code. Treat those fields as unreliable and verify what services the skill actually uses. - Filesystem probing & integrations: The skill checks for and may read other skill files (prompt-engineer) and invokes gh/claude CLIs. If you do not want the skill to inspect or interact with other local skills or CLIs, avoid running it or run in an isolated environment (container or VM). - Least privilege: Run the installation and first transcriptions in a sandbox or virtualenv so pip --user or model downloads do not affect your system Python environment. What would change the assessment to benign: the author clarifies and removes misleading meta.json fields, documents and asks explicit, consent-based steps for any networked LLM calls, and provides a clear switch to enforce 100% local-only mode (disabling any external CLI/LLM invocation and not auto-installing packages).

Purpose

This skill automates audio-to-text transcription with professional Markdown output, extracting rich technical metadata (speakers, timestamps, language, file size, duration) and generating structured meeting minutes and executive summaries. It uses Faster-Whisper or Whisper with zero configuration, working universally across projects without hardcoded paths or API keys.

Inspired by tools like Plaud, this skill transforms raw audio recordings into actionable documentation, making it ideal for meetings, interviews, lectures, and content analysis.

When to Use

Invoke this skill when:

• User needs to transcribe audio/video files to text
• User wants meeting minutes automatically generated from recordings
• User requires speaker identification (diarization) in conversations
• User needs subtitles/captions (SRT, VTT formats)
• User wants executive summaries of long audio content
• User asks variations of "transcribe this audio", "convert audio to text", "generate meeting notes from recording"
• User has audio files in common formats (MP3, WAV, M4A, OGG, FLAC, WEBM)

Workflow

Step 0: Discovery (Auto-detect Transcription Tools)

Objective: Identify available transcription engines without user configuration.

Actions:

Run detection commands to find installed tools:

# Check for Faster-Whisper (preferred - 4-5x faster)
if python3 -c "import faster_whisper" 2>/dev/null; then
    TRANSCRIBER="faster-whisper"
    echo "✅ Faster-Whisper detected (optimized)"
# Fallback to original Whisper
elif python3 -c "import whisper" 2>/dev/null; then
    TRANSCRIBER="whisper"
    echo "✅ OpenAI Whisper detected"
else
    TRANSCRIBER="none"
    echo "⚠️  No transcription tool found"
fi

# Check for ffmpeg (audio format conversion)
if command -v ffmpeg &>/dev/null; then
    echo "✅ ffmpeg available (format conversion enabled)"
else
    echo "ℹ️  ffmpeg not found (limited format support)"
fi

If no transcriber found:

Offer automatic installation using the provided script:

echo "⚠️  No transcription tool found"
echo ""
echo "🔧 Auto-install dependencies? (Recommended)"
read -p "Run installation script? [Y/n]: " AUTO_INSTALL

if [[ ! "$AUTO_INSTALL" =~ ^[Nn] ]]; then
    # Get skill directory (works for both repo and symlinked installations)
    SKILL_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
    
    # Run installation script
    if [[ -f "$SKILL_DIR/scripts/install-requirements.sh" ]]; then
        bash "$SKILL_DIR/scripts/install-requirements.sh"
    else
        echo "❌ Installation script not found"
        echo ""
        echo "📦 Manual installation:"
        echo "  pip install faster-whisper  # Recommended"
        echo "  pip install openai-whisper  # Alternative"
        echo "  brew install ffmpeg         # Optional (macOS)"
        exit 1
    fi
    
    # Verify installation succeeded
    if python3 -c "import faster_whisper" 2>/dev/null || python3 -c "import whisper" 2>/dev/null; then
        echo "✅ Installation successful! Proceeding with transcription..."
    else
        echo "❌ Installation failed. Please install manually."
        exit 1
    fi
else
    echo ""
    echo "📦 Manual installation required:"
    echo ""
    echo "Recommended (fastest):"
    echo "  pip install faster-whisper"
    echo ""
    echo "Alternative (original):"
    echo "  pip install openai-whisper"
    echo ""
    echo "Optional (format conversion):"
    echo "  brew install ffmpeg  # macOS"
    echo "  apt install ffmpeg   # Linux"
    echo ""
    exit 1
fi

This ensures users can install dependencies with one confirmation, or opt for manual installation if preferred.

If transcriber found:

Proceed to Step 0b (CLI Detection).

Step 1: Validate Audio File

Objective: Verify file exists, check format, and extract metadata.

Actions:

1. Accept file path or URL from user:

   • Local file: meeting.mp3
   • URL: https://example.com/audio.mp3 (download to temp directory)

2. Verify file exists:

if [[ ! -f "$AUDIO_FILE" ]]; then
    echo "❌ File not found: $AUDIO_FILE"
    exit 1
fi

3. Extract metadata using ffprobe or file utilities:

# Get file size
FILE_SIZE=$(du -h "$AUDIO_FILE" | cut -f1)

# Get duration and format using ffprobe
DURATION=$(ffprobe -v error -show_entries format=duration \
    -of default=noprint_wrappers=1:nokey=1 "$AUDIO_FILE" 2>/dev/null)
FORMAT=$(ffprobe -v error -select_streams a:0 -show_entries \
    stream=codec_name -of default=noprint_wrappers=1:nokey=1 "$AUDIO_FILE" 2>/dev/null)

# Convert duration to HH:MM:SS
DURATION_HMS=$(date -u -r "$DURATION" +%H:%M:%S 2>/dev/null || echo "Unknown")

4. Check file size (warn if large for cloud APIs):

SIZE_MB=$(du -m "$AUDIO_FILE" | cut -f1)
if [[ $SIZE_MB -gt 25 ]]; then
    echo "⚠️  Large file ($FILE_SIZE) - processing may take several minutes"
fi

5. Validate format (supported: MP3, WAV, M4A, OGG, FLAC, WEBM):

EXTENSION="${AUDIO_FILE##*.}"
SUPPORTED_FORMATS=("mp3" "wav" "m4a" "ogg" "flac" "webm" "mp4")

if [[ ! " ${SUPPORTED_FORMATS[@]} " =~ " ${EXTENSION,,} " ]]; then
    echo "⚠️  Unsupported format: $EXTENSION"
    if command -v ffmpeg &>/dev/null; then
        echo "🔄 Converting to WAV..."
        ffmpeg -i "$AUDIO_FILE" -ar 16000 "${AUDIO_FILE%.*}.wav" -y
        AUDIO_FILE="${AUDIO_FILE%.*}.wav"
    else
        echo "❌ Install ffmpeg to convert formats: brew install ffmpeg"
        exit 1
    fi
fi

Step 3: Generate Markdown Output

Objective: Create structured Markdown with metadata, transcription, meeting minutes, and summary.

Output Template:

# Audio Transcription Report

## 📊 Metadata

| Field | Value |
|-------|-------|
| **File Name** | {filename} |
| **File Size** | {file_size} |
| **Duration** | {duration_hms} |
| **Language** | {language} ({language_code}) |
| **Processed Date** | {process_date} |
| **Speakers Identified** | {num_speakers} |
| **Transcription Engine** | {engine} (model: {model}) |


## 📋 Meeting Minutes

### Participants
- {speaker_1}
- {speaker_2}
- ...

### Topics Discussed
1. **{topic_1}** ({timestamp})
   - {key_point_1}
   - {key_point_2}

2. **{topic_2}** ({timestamp})
   - {key_point_1}

### Decisions Made
- ✅ {decision_1}
- ✅ {decision_2}

### Action Items
- [ ] **{action_1}** - Assigned to: {speaker} - Due: {date_if_mentioned}
- [ ] **{action_2}** - Assigned to: {speaker}


*Generated by audio-transcriber skill v1.0.0*  
*Transcription engine: {engine} | Processing time: {elapsed_time}s*

Implementation:

Use Python or bash with AI model (Claude/GPT) for intelligent summarization:

def generate_meeting_minutes(segments):
    """Extract topics, decisions, action items from transcription."""
    
    # Group segments by topic (simple clustering by timestamps)
    topics = cluster_by_topic(segments)
    
    # Identify action items (keywords: "should", "will", "need to", "action")
    action_items = extract_action_items(segments)
    
    # Identify decisions (keywords: "decided", "agreed", "approved")
    decisions = extract_decisions(segments)
    
    return {
        "topics": topics,
        "decisions": decisions,
        "action_items": action_items
    }

def generate_summary(segments, max_paragraphs=5):
    """Create executive summary using AI (Claude/GPT via API or local model)."""
    
    full_text = " ".join([s["text"] for s in segments])
    
    # Use Chain of Density approach (from prompt-engineer frameworks)
    summary_prompt = f"""
    Summarize the following transcription in {max_paragraphs} concise paragraphs.
    Focus on key topics, decisions, and action items.
    
    Transcription:
    {full_text}
    """
    
    # Call AI model (placeholder - user can integrate Claude API or use local model)
    summary = call_ai_model(summary_prompt)
    
    return summary

Output file naming:

# v1.1.0: Use timestamp para evitar sobrescrever
TIMESTAMP=$(date +%Y%m%d-%H%M%S)
TRANSCRIPT_FILE="transcript-${TIMESTAMP}.md"
ATA_FILE="ata-${TIMESTAMP}.md"

echo "$TRANSCRIPT_CONTENT" > "$TRANSCRIPT_FILE"
echo "✅ Transcript salvo: $TRANSCRIPT_FILE"

if [[ -n "$ATA_CONTENT" ]]; then
    echo "$ATA_CONTENT" > "$ATA_FILE"
    echo "✅ Ata salva: $ATA_FILE"
fi

SCENARIO A: User Provided Custom Prompt

Workflow:

1. Display user's prompt:

   📝 Prompt fornecido pelo usuário:
   ┌──────────────────────────────────┐
   │ [User's prompt preview]          │
   └──────────────────────────────────┘
   

2. Automatically improve with prompt-engineer (if available):

   🔧 Melhorando prompt com prompt-engineer...
   [Invokes: gh copilot -p "melhore este prompt: {user_prompt}"]
   

3. Show both versions:

   ✨ Versão melhorada:
   ┌──────────────────────────────────┐
   │ Role: Você é um documentador...  │
   │ Instructions: Transforme...      │
   │ Steps: 1) ... 2) ...             │
   │ End Goal: ...                    │
   └──────────────────────────────────┘
   
   📝 Versão original:
   ┌──────────────────────────────────┐
   │ [User's original prompt]         │
   └──────────────────────────────────┘
   

4. Ask which to use:

   💡 Usar versão melhorada? [s/n] (default: s):
   

5. Process with selected prompt:

   • If "s": use improved
   • If "n": use original

LLM Processing (Both Scenarios)

Once prompt is finalized:

from rich.progress import Progress, SpinnerColumn, TextColumn

def process_with_llm(transcript, prompt, cli_tool='claude'):
    full_prompt = f"{prompt}\n\n---\n\nTranscrição:\n\n{transcript}"
    
    with Progress(
        SpinnerColumn(),
        TextColumn("[progress.description]{task.description}"),
        transient=True
    ) as progress:
        progress.add_task(
            description=f"🤖 Processando com {cli_tool}...",
            total=None
        )
        
        if cli_tool == 'claude':
            result = subprocess.run(
                ['claude', '-'],
                input=full_prompt,
                capture_output=True,
                text=True,
                timeout=300  # 5 minutes
            )
        elif cli_tool == 'gh-copilot':
            result = subprocess.run(
                ['gh', 'copilot', 'suggest', '-t', 'shell', full_prompt],
                capture_output=True,
                text=True,
                timeout=300
            )
    
    if result.returncode == 0:
        return result.stdout.strip()
    else:
        return None

Progress output:

🤖 Processando com claude... ⠋
[After completion:]
✅ Ata gerada com sucesso!

Final Output

Success (both files):

💾 Salvando arquivos...

✅ Arquivos criados:
  - transcript-20260203-023045.md  (transcript puro)
  - ata-20260203-023045.md         (processado com LLM)

🧹 Removidos arquivos temporários: metadata.json, transcription.json

✅ Concluído! Tempo total: 3m 45s

Transcript only (user declined LLM):

💾 Salvando arquivos...

✅ Arquivo criado:
  - transcript-20260203-023045.md

ℹ️  Ata não gerada (processamento LLM recusado pelo usuário)

🧹 Removidos arquivos temporários: metadata.json, transcription.json

✅ Concluído!

Step 5: Display Results Summary

Objective: Show completion status and next steps.

Output:

echo ""
echo "✅ Transcription Complete!"
echo ""
echo "📊 Results:"
echo "  File: $OUTPUT_FILE"
echo "  Language: $LANGUAGE"
echo "  Duration: $DURATION_HMS"
echo "  Speakers: $NUM_SPEAKERS"
echo "  Words: $WORD_COUNT"
echo "  Processing time: ${ELAPSED_TIME}s"
echo ""
echo "📝 Generated:"
echo "  - $OUTPUT_FILE (Markdown report)"
[if alternative formats:]
echo "  - ${OUTPUT_FILE%.*}.srt (Subtitles)"
echo "  - ${OUTPUT_FILE%.*}.json (Structured data)"
echo ""
echo "🎯 Next steps:"
echo "  1. Review meeting minutes and action items"
echo "  2. Share report with participants"
echo "  3. Track action items to completion"

Example Usage

Example 1: Basic Transcription

User Input:

copilot> transcribe audio to markdown: meeting-2026-02-02.mp3

Skill Output:

✅ Faster-Whisper detected (optimized)
✅ ffmpeg available (format conversion enabled)

📂 File: meeting-2026-02-02.mp3
📊 Size: 12.3 MB
⏱️  Duration: 00:45:32

🎙️  Processing...
[████████████████████] 100%

✅ Language detected: Portuguese (pt-BR)
👥 Speakers identified: 4
📝 Generating Markdown output...

✅ Transcription Complete!

📊 Results:
  File: meeting-2026-02-02.md
  Language: pt-BR
  Duration: 00:45:32
  Speakers: 4
  Words: 6,842
  Processing time: 127s

📝 Generated:
  - meeting-2026-02-02.md (Markdown report)

🎯 Next steps:
  1. Review meeting minutes and action items
  2. Share report with participants
  3. Track action items to completion

Example 3: Batch Processing

User Input:

copilot> transcreva estes áudios: recordings/*.mp3

Skill Output:

📦 Batch mode: 5 files found
  1. team-standup.mp3
  2. client-call.mp3
  3. brainstorm-session.mp3
  4. product-demo.mp3
  5. retrospective.mp3

🎙️  Processing batch...

[1/5] team-standup.mp3 ✅ (2m 34s)
[2/5] client-call.mp3 ✅ (15m 12s)
[3/5] brainstorm-session.mp3 ✅ (8m 47s)
[4/5] product-demo.mp3 ✅ (22m 03s)
[5/5] retrospective.mp3 ✅ (11m 28s)

✅ Batch Complete!
📝 Generated 5 Markdown reports
⏱️  Total processing time: 6m 15s

Example 5: Large File Warning

User Input:

copilot> transcribe audio to markdown: conference-keynote.mp3

Skill Output:

✅ Faster-Whisper detected (optimized)

📂 File: conference-keynote.mp3
📊 Size: 87.2 MB
⏱️  Duration: 02:15:47
⚠️  Large file (87.2 MB) - processing may take several minutes

Continue? [Y/n]:

User: Y

🎙️  Processing... (this may take 10-15 minutes)
[████░░░░░░░░░░░░░░░░] 20% - Estimated time remaining: 12m

This skill is platform-agnostic and works in any terminal context where GitHub Copilot CLI is available. It does not depend on specific project configurations or external APIs, following the zero-configuration philosophy.