add-voice-transcription

name	add-voice-transcription
description	Add voice message transcription to BastionClaw using OpenAI's Whisper API. Automatically transcribes WhatsApp voice notes so the agent can read and respond to them.
allowed-tools	Bash(*), Read, Edit, Write, Glob, Grep, AskUserQuestion

Add Voice Message Transcription

This skill adds automatic voice message transcription using OpenAI's Whisper API. When users send voice notes in WhatsApp, they'll be transcribed and the agent can read and respond to the content.

UX Note: When asking the user questions, prefer using the AskUserQuestion tool instead of just outputting text. This integrates with Claude's built-in question/answer system for a better experience.

Prerequisites

USER ACTION REQUIRED

Use the AskUserQuestion tool to present this:

You'll need an OpenAI API key for Whisper transcription.

Get one at: https://platform.openai.com/api-keys

Cost: $0.006 per minute of audio ($0.003 per typical 30-second voice note)

Once you have your API key, we'll configure it securely.

Wait for user to confirm they have an API key before continuing.

---

Implementation

Step 1: Add OpenAI Dependency

Read package.json and add the openai package to dependencies:

"dependencies": {
  ...existing dependencies...
  "openai": "^4.77.0"
}

Then install it. IMPORTANT: The OpenAI SDK requires Zod v3 as an optional peer dependency, but BastionClaw uses Zod v4. This conflict is guaranteed, so always use --legacy-peer-deps:

npm install --legacy-peer-deps

Step 2: Create Transcription Configuration

Create a configuration file for transcription settings (without the API key):

Write to .transcription.config.json:

{
  "provider": "openai",
  "openai": {
    "apiKey": "",
    "model": "whisper-1"
  },
  "enabled": true,
  "fallbackMessage": "[Voice Message - transcription unavailable]"
}

Add this file to .gitignore to prevent committing API keys:

echo ".transcription.config.json" >> .gitignore

Use the AskUserQuestion tool to confirm:

I've created .transcription.config.json in the project root. You'll need to add your OpenAI API key to it manually:

1. Open .transcription.config.json
2. Replace the empty "apiKey": "" with your key: "apiKey": "sk-proj-..."
3. Save the file

Let me know when you've added it.

Wait for user confirmation.

Step 3: Create Transcription Module

Create src/transcription.ts:

import { downloadMediaMessage } from '@whiskeysockets/baileys';
import { WAMessage, WASocket } from '@whiskeysockets/baileys';
import fs from 'fs';
import path from 'path';
import { fileURLToPath } from 'url';
import { dirname } from 'path';

// Get __dirname equivalent in ES modules
const __filename = fileURLToPath(import.meta.url);
const __dirname = dirname(__filename);

// Configuration interface
interface TranscriptionConfig {
  provider: string;
  openai?: {
    apiKey: string;
    model: string;
  };
  enabled: boolean;
  fallbackMessage: string;
}

// Load configuration
function loadConfig(): TranscriptionConfig {
  const configPath = path.join(__dirname, '../.transcription.config.json');
  try {
    const configData = fs.readFileSync(configPath, 'utf-8');
    return JSON.parse(configData);
  } catch (err) {
    console.error('Failed to load transcription config:', err);
    return {
      provider: 'openai',
      enabled: false,
      fallbackMessage: '[Voice Message - transcription unavailable]'
    };
  }
}

// Transcribe audio using OpenAI Whisper API
async function transcribeWithOpenAI(audioBuffer: Buffer, config: TranscriptionConfig): Promise<string | null> {
  if (!config.openai?.apiKey || config.openai.apiKey === '') {
    console.warn('OpenAI API key not configured');
    return null;
  }

  try {
    // Dynamic import of openai
    const openaiModule = await import('openai');
    const OpenAI = openaiModule.default;
    const toFile = openaiModule.toFile;

    const openai = new OpenAI({
      apiKey: config.openai.apiKey
    });

    // Use OpenAI's toFile helper to create a proper file upload
    const file = await toFile(audioBuffer, 'voice.ogg', {
      type: 'audio/ogg'
    });

    // Call Whisper API
    const transcription = await openai.audio.transcriptions.create({
      file: file,
      model: config.openai.model || 'whisper-1',
      response_format: 'text'
    });

    // Type assertion needed: OpenAI SDK types response_format='text' as Transcription object,
    // but it actually returns a plain string when response_format is 'text'
    return transcription as unknown as string;
  } catch (err) {
    console.error('OpenAI transcription failed:', err);
    return null;
  }
}

// Main transcription function
export async function transcribeAudioMessage(
  msg: WAMessage,
  sock: WASocket
): Promise<string | null> {
  const config = loadConfig();

  // Check if transcription is enabled
  if (!config.enabled) {
    console.log('Transcription disabled in config');
    return config.fallbackMessage;
  }

  try {
    // Download the audio message
    const buffer = await downloadMediaMessage(
      msg,
      'buffer',
      {},
      {
        logger: console as any,
        reuploadRequest: sock.updateMediaMessage
      }
    ) as Buffer;

    if (!buffer || buffer.length === 0) {
      console.error('Failed to download audio message');
      return config.fallbackMessage;
    }

    console.log(`Downloaded audio message: ${buffer.length} bytes`);

    // Transcribe based on provider
    let transcript: string | null = null;

    switch (config.provider) {
      case 'openai':
        transcript = await transcribeWithOpenAI(buffer, config);
        break;
      default:
        console.error(`Unknown transcription provider: ${config.provider}`);
        return config.fallbackMessage;
    }

    if (!transcript) {
      return config.fallbackMessage;
    }

    return transcript.trim();
  } catch (err) {
    console.error('Transcription error:', err);
    return config.fallbackMessage;
  }
}

// Helper to check if a message is a voice note
export function isVoiceMessage(msg: WAMessage): boolean {
  return msg.message?.audioMessage?.ptt === true;
}

Step 4: Update Database to Handle Transcribed Content

Read src/db.ts and find the storeMessage function. Update its signature and implementation to accept transcribed content:

Change the function signature from:

export function storeMessage(msg: proto.IWebMessageInfo, chatJid: string, isFromMe: boolean, pushName?: string): void

To:

export function storeMessage(msg: proto.IWebMessageInfo, chatJid: string, isFromMe: boolean, pushName?: string, transcribedContent?: string): void

Update the content extraction to use transcribed content if provided:

const content = transcribedContent ||
  msg.message?.conversation ||
  msg.message?.extendedTextMessage?.text ||
  msg.message?.imageMessage?.caption ||
  msg.message?.videoMessage?.caption ||
  (msg.message?.audioMessage?.ptt ? '[Voice Message]' : '') ||
  '';

Step 5: Integrate Transcription into Message Handler

Note: Voice messages are transcribed for all messages in registered groups, regardless of the trigger word. This is because:

1. Voice notes can't easily include a trigger word
2. Users expect voice notes to work the same as text messages
3. The transcribed content is stored in the database for context, even if it doesn't trigger the agent

Read src/index.ts and find the sock.ev.on('messages.upsert', ...) event handler.

Change the callback from synchronous to async:

sock.ev.on('messages.upsert', async ({ messages }) => {

Inside the loop where messages are stored, add voice message detection and transcription:

// Only store full message content for registered groups
if (registeredGroups[chatJid]) {
  // Check if this is a voice message
  if (msg.message.audioMessage?.ptt) {
    try {
      // Import transcription module
      const { transcribeAudioMessage } = await import('./transcription.js');
      const transcript = await transcribeAudioMessage(msg, sock);

      if (transcript) {
        // Store with transcribed content
        storeMessage(msg, chatJid, msg.key.fromMe || false, msg.pushName || undefined, `[Voice: ${transcript}]`);
        logger.info({ chatJid, length: transcript.length }, 'Transcribed voice message');
      } else {
        // Store with fallback message
        storeMessage(msg, chatJid, msg.key.fromMe || false, msg.pushName || undefined, '[Voice Message - transcription unavailable]');
      }
    } catch (err) {
      logger.error({ err }, 'Voice transcription error');
      storeMessage(msg, chatJid, msg.key.fromMe || false, msg.pushName || undefined, '[Voice Message - transcription failed]');
    }
  } else {
    // Regular message, store normally
    storeMessage(msg, chatJid, msg.key.fromMe || false, msg.pushName || undefined);
  }
}

Step 6: Fix Orphan Container Cleanup (CRITICAL)

This step is essential. When the BastionClaw service restarts (e.g., launchctl kickstart -k), the running container is detached but NOT killed. The new service instance spawns a fresh container, but the orphan keeps running and shares the same IPC mount directory. Both containers race to read IPC input files, causing the new container to randomly miss messages — making it appear like the agent doesn't respond.

The existing cleanup code in ensureContainerSystemRunning() in src/index.ts uses container ls --format {{.Names}} which silently fails on Apple Container (only json and table are valid format options). The catch block swallows the error, so orphans are never cleaned up.

Find the orphan cleanup block in ensureContainerSystemRunning() (the section starting with // Kill and clean up orphaned BastionClaw containers from previous runs) and replace it with:

  // Kill and clean up orphaned BastionClaw containers from previous runs
  try {
    const listJson = execSync('container ls -a --format json', {
      stdio: ['pipe', 'pipe', 'pipe'],
      encoding: 'utf-8',
    });
    const containers = JSON.parse(listJson) as Array<{ configuration: { id: string }; status: string }>;
    const bastionclawContainers = containers.filter(
      (c) => c.configuration.id.startsWith('bastionclaw-'),
    );
    const running = bastionclawContainers
      .filter((c) => c.status === 'running')
      .map((c) => c.configuration.id);
    if (running.length > 0) {
      execSync(`container stop ${running.join(' ')}`, { stdio: 'pipe' });
      logger.info({ count: running.length }, 'Stopped orphaned containers');
    }
    const allNames = bastionclawContainers.map((c) => c.configuration.id);
    if (allNames.length > 0) {
      execSync(`container rm ${allNames.join(' ')}`, { stdio: 'pipe' });
      logger.info({ count: allNames.length }, 'Cleaned up stopped containers');
    }
  } catch {
    // No containers or cleanup not supported
  }

Step 7: Build and Restart

npm run build

Before restarting the service, kill any orphaned containers manually to ensure a clean slate:

container ls -a --format json | python3 -c "
import sys, json
data = json.load(sys.stdin)
nc = [c['configuration']['id'] for c in data if c['configuration']['id'].startswith('bastionclaw-')]
if nc: print(' '.join(nc))
" | xargs -r container stop 2>/dev/null
container ls -a --format json | python3 -c "
import sys, json
data = json.load(sys.stdin)
nc = [c['configuration']['id'] for c in data if c['configuration']['id'].startswith('bastionclaw-')]
if nc: print(' '.join(nc))
" | xargs -r container rm 2>/dev/null
echo "Orphaned containers cleaned"

Now restart the service:

launchctl kickstart -k gui/$(id -u)/com.bastionclaw

Verify it started with exactly one (or zero, before first message) bastionclaw container:

sleep 3 && launchctl list | grep bastionclaw
container ls -a --format json | python3 -c "
import sys, json
data = json.load(sys.stdin)
nc = [c for c in data if c['configuration']['id'].startswith('bastionclaw-')]
print(f'{len(nc)} bastionclaw container(s)')
for c in nc: print(f'  {c[\"configuration\"][\"id\"]} - {c[\"status\"]}')
"

Step 8: Test Voice Transcription

Tell the user:

Voice transcription is ready! Test it by:

1. Open WhatsApp on your phone
2. Go to a registered group chat
3. Send a voice note using the microphone button
4. The agent should receive the transcribed text and respond

In the database and agent context, voice messages appear as: [Voice: <transcribed text here>]

Watch for transcription in the logs:

tail -f logs/bastionclaw.log | grep -i "voice\|transcri"

---

Configuration Options

Enable/Disable Transcription

To temporarily disable without removing code, edit .transcription.config.json:

{
  "enabled": false
}

Change Fallback Message

Customize what's stored when transcription fails:

{
  "fallbackMessage": "[🎤 Voice note - transcription unavailable]"
}

Switch to Different Provider (Future)

The architecture supports multiple providers. To add Groq, Deepgram, or local Whisper:

1. Add provider config to .transcription.config.json
2. Implement provider function in src/transcription.ts (similar to transcribeWithOpenAI)
3. Add case to the switch statement

---

Troubleshooting

Agent doesn't respond to voice messages (or any messages after a voice note)

Most likely cause: orphaned containers. When the service restarts, the previous container keeps running and races to consume IPC messages. Check:

container ls -a --format json | python3 -c "
import sys, json
data = json.load(sys.stdin)
nc = [c for c in data if c['configuration']['id'].startswith('bastionclaw-')]
print(f'{len(nc)} bastionclaw container(s):')
for c in nc: print(f'  {c[\"configuration\"][\"id\"]} - {c[\"status\"]}')
"

If you see more than one running container, kill the orphans:

# Stop all bastionclaw containers, then restart the service
container ls -a --format json | python3 -c "
import sys, json
data = json.load(sys.stdin)
running = [c['configuration']['id'] for c in data if c['configuration']['id'].startswith('bastionclaw-') and c['status'] == 'running']
if running: print(' '.join(running))
" | xargs -r container stop 2>/dev/null
container ls -a --format json | python3 -c "
import sys, json
data = json.load(sys.stdin)
nc = [c['configuration']['id'] for c in data if c['configuration']['id'].startswith('bastionclaw-')]
if nc: print(' '.join(nc))
" | xargs -r container rm 2>/dev/null
launchctl kickstart -k gui/$(id -u)/com.bastionclaw

Root cause: The ensureContainerSystemRunning() function previously used container ls --format {{.Names}} which silently fails on Apple Container (only json and table formats are supported). Step 6 of this skill fixes this. If you haven't applied Step 6, the orphan problem will recur on every restart.

"Transcription unavailable" or "Transcription failed"

Check logs for specific errors:

tail -100 logs/bastionclaw.log | grep -i transcription

Common causes:

• API key not configured or invalid
• No API credits remaining
• Network connectivity issues
• Audio format not supported by Whisper

Voice messages not being detected

• Ensure you're sending actual voice notes (microphone button), not audio file attachments
• Check that audioMessage.ptt is true in the message object

ES Module errors (__dirname is not defined)

The fix is already included in the implementation above using:

const __filename = fileURLToPath(import.meta.url);
const __dirname = dirname(__filename);

Dependency conflicts (Zod versions)

The OpenAI SDK requires Zod v3, but BastionClaw uses Zod v4. This conflict is guaranteed — always use:

npm install --legacy-peer-deps

---

Security Notes

• The .transcription.config.json file contains your API key and should NOT be committed to version control
• It's added to .gitignore by this skill
• Audio files are sent to OpenAI for transcription - review their data usage policy
• No audio files are stored locally after transcription
• Transcripts are stored in the SQLite database like regular text messages

---

Cost Management

Monitor usage in your OpenAI dashboard: https://platform.openai.com/usage

Tips to control costs:

• Set spending limits in OpenAI account settings
• Disable transcription during development/testing with "enabled": false
• Typical usage: 100 voice notes/month (~3 minutes average) = ~$1.80

---

Removing Voice Transcription

To remove the feature:

1. Remove from package.json:

   npm uninstall openai
   

2. Delete src/transcription.ts

3. Revert changes in src/index.ts:

   • Remove the voice message handling block
   • Change callback back to synchronous if desired

4. Revert changes in src/db.ts:

   • Remove the transcribedContent parameter from storeMessage

5. Delete .transcription.config.json

6. Rebuild:

   npm run build
   launchctl kickstart -k gui/$(id -u)/com.bastionclaw
   

---

Future Enhancements

Potential additions:

• Local Whisper: Use whisper.cpp or faster-whisper for offline transcription
• Groq Integration: Free tier with Whisper, very fast
• Deepgram: Alternative cloud provider
• Language Detection: Auto-detect and transcribe non-English voice notes
• Cost Tracking: Log transcription costs per message
• Speaker Diarization: Identify different speakers in voice notes