voice-ai-integration

Build production-ready voice AI training platforms with Hume EVI (Empathic Voice Interface), Twilio phone integration, and modern web stack (React + tRPC + Drizzle ORM).

When to Use This Skill

• Building voice-based roleplay or training systems
• Integrating Hume AI for emotionally intelligent conversations
• Adding phone calling capabilities (inbound/outbound) to web apps
• Creating practice/training platforms with conversation tracking
• Implementing multi-modal voice interfaces (web + phone)

Core Architecture Pattern

Frontend (React + tRPC)
    ↓
Backend (Express + tRPC + Drizzle)
    ↓
Voice AI (Hume EVI) + Phone (Twilio)
    ↓
Database (MySQL/TiDB) - Session & conversation tracking

Implementation Workflow

Phase 1: Environment Setup

1. Add Hume AI Credentials

Use webdev_request_secrets to add:

• HUME_API_KEY - API key from Hume dashboard
• HUME_SECRET_KEY - Secret key for access token generation
• HUME_FEMALE_CONFIG_ID - Female voice configuration UUID
• HUME_MALE_CONFIG_ID - Male voice configuration UUID
• HUME_WEBHOOK_URL - Webhook endpoint for call events (optional)

2. Add Twilio Credentials (if phone integration needed)

• TWILIO_ACCOUNT_SID - Account SID from Twilio console
• TWILIO_AUTH_TOKEN - Auth token from Twilio console
• TWILIO_PHONE_NUMBER - Purchased phone number in E.164 format (+15551234567)

3. Update server environment config

Add to server/_core/env.ts:

export const ENV = {
  // ... existing vars
  humeApiKey: process.env.HUME_API_KEY!,
  humeSecretKey: process.env.HUME_SECRET_KEY!,
  humeFemaleConfigId: process.env.HUME_FEMALE_CONFIG_ID!,
  humeMaleConfigId: process.env.HUME_MALE_CONFIG_ID!,
  humeWebhookUrl: process.env.HUME_WEBHOOK_URL,
  twilioAccountSid: process.env.TWILIO_ACCOUNT_SID,
  twilioAuthToken: process.env.TWILIO_AUTH_TOKEN,
  twilioPhoneNumber: process.env.TWILIO_PHONE_NUMBER,
};

Phase 2: Database Schema

Core tables needed:

1. Sessions table - Practice/training sessions

   • userId, status, archetype/scenario, difficulty, mode
   • voiceGender, roleplayMode (standard/reverse)
   • callType (web/phone_inbound/phone_outbound)
   • userPhoneNumber, twilioCallSid
   • score, feedback, duration
   • userCode (for tracking individual users)

2. Conversation turns table - Message-by-message tracking

   • sessionId, speaker (user/ai), message, timestamp
   • emotions (JSON - from Hume), confidence

3. User progress table - Aggregate stats

   • userId, totalSessions, avgScore
   • Scenario-specific completion counters

4. Scripts table (if using predefined content)

   • userId, name, content, fileUrl

Example schema (Drizzle ORM):

export const practiceSessions = mysqlTable('practice_sessions', {
  id: int('id').primaryKey().autoincrement(),
  userId: int('user_id').notNull(),
  userCode: varchar('user_code', { length: 50 }),
  archetype: varchar('archetype', { length: 50 }).notNull(),
  difficulty: varchar('difficulty', { length: 50 }).notNull(),
  practiceMode: varchar('practice_mode', { length: 50 }).notNull(),
  voiceGender: varchar('voice_gender', { length: 10 }).notNull(),
  roleplayMode: varchar('roleplay_mode', { length: 20 }).default('standard'),
  callType: varchar('call_type', { length: 20 }).default('web'),
  userPhoneNumber: varchar('user_phone_number', { length: 20 }),
  twilioCallSid: varchar('twilio_call_sid', { length: 100 }),
  status: varchar('status', { length: 20 }).notNull(),
  score: int('score'),
  feedback: text('feedback'),
  duration: int('duration'),
  createdAt: timestamp('created_at').defaultNow(),
});

Phase 3: Hume Service Layer

Install Hume SDK:

pnpm add hume

Create server/hume-service.ts:

Key functions:

1. getHumeAccessToken() - Generate access token for client
2. generateSystemPrompt() - Dynamic prompt based on scenario/difficulty/mode
3. prepareHumeSession() - Return configId + systemPrompt
4. getConfigId() - Select male/female voice config

System prompt pattern:

function generateSystemPrompt(params: {
  archetype: string;
  difficulty: string;
  practiceMode: string;
  roleplayMode: 'standard' | 'reverse';
  scriptContent?: string;
}): string {
  if (params.roleplayMode === 'reverse') {
    // AI plays expert sales rep, user plays homeowner
    return `You are an expert door-to-door fiber sales representative...`;
  }
  
  // Standard: AI plays homeowner, user plays sales rep
  return `You are a ${params.archetype} homeowner. ${getPersonalityTraits(params.archetype)}...`;
}

Phase 4: Twilio Service Layer (Optional)

Install Twilio SDK:

pnpm add twilio

Create server/twilio-service.ts:

Key functions:

1. getTwilioPhoneNumber() - Return configured phone number
2. initiateOutboundCall() - Call user's phone number with Hume config

Outbound call pattern:

export async function initiateOutboundCall(params: {
  toNumber: string;
  configId: string;
  apiKey: string;
}): Promise<{ success: boolean; callSid?: string; error?: string }> {
  const client = twilio(ENV.twilioAccountSid, ENV.twilioAuthToken);
  
  const call = await client.calls.create({
    to: params.toNumber,
    from: ENV.twilioPhoneNumber,
    twiml: `<Response><Connect><Stream url="wss://api.hume.ai/v0/evi/chat?config_id=${params.configId}&api_key=${params.apiKey}" /></Connect></Response>`,
  });
  
  return { success: true, callSid: call.sid };
}

Phase 5: tRPC Router

Create server/practice-router.ts (or similar):

Essential endpoints:

1. startSession - Create session, return sessionId
2. getHumeToken - Get access token for web voice calls
3. prepareHumeSession - Get configId + systemPrompt
4. getTwilioPhoneNumber - Return phone number for call-in
5. initiatePhoneCall - Trigger outbound call
6. updateSession - Update score/feedback/duration
7. getMySessions - List user's sessions
8. getSession - Get single session details
9. getConversation - Get message-by-message transcript

Key pattern - session creation:

startSession: protectedProcedure
  .input(z.object({
    scriptId: z.number().optional(),
    userCode: z.string().optional(),
    archetype: z.enum(['type1', 'type2', 'type3']),
    difficulty: z.enum(['easy', 'medium', 'hard']),
    practiceMode: z.enum(['mode1', 'mode2']),
    voiceGender: z.enum(['male', 'female']),
  }))
  .mutation(async ({ ctx, input }) => {
    const result = await createPracticeSession({
      userId: ctx.user.id,
      ...input,
      status: 'in_progress',
    });
    return { sessionId: result.insertId };
  })

Phase 6: Frontend React Hooks

Create client/src/hooks/useHumeVoice.ts:

Manages WebSocket connection to Hume EVI:

• Connect/disconnect
• Send/receive audio
• Handle messages with emotional data
• Track connection status

Pattern:

export function useHumeVoice(sessionId: number | null) {
  const [status, setStatus] = useState<'idle' | 'connecting' | 'connected' | 'disconnected'>('idle');
  const [messages, setMessages] = useState<Message[]>([]);
  
  const connect = async () => {
    const token = await trpc.practice.getHumeToken.query();
    const config = await trpc.practice.prepareHumeSession.query({ sessionId });
    
    // WebSocket connection to Hume EVI
    const ws = new WebSocket(`wss://api.hume.ai/v0/evi/chat?access_token=${token}&config_id=${config.configId}`);
    // ... handle audio streaming
  };
  
  return { status, messages, connect, disconnect };
}

Phase 7: Voice Call UI Component

Create client/src/components/VoiceCall.tsx:

Features:

• Start/end call buttons
• Mute/unmute toggle
• Real-time transcript display
• Call duration timer
• Visual status indicators

Integration with useHumeVoice hook:

export function VoiceCall({ sessionId }: { sessionId: number }) {
  const { status, messages, connect, disconnect } = useHumeVoice(sessionId);
  
  return (
    <Card>
      <CardHeader>
        <CardTitle>Practice Session</CardTitle>
        <Badge>{status}</Badge>
      </CardHeader>
      <CardContent>
        <div className="space-y-4">
          {messages.map((msg, i) => (
            <div key={i} className={msg.speaker === 'user' ? 'text-right' : 'text-left'}>
              <p>{msg.content}</p>
            </div>
          ))}
        </div>
        <Button onClick={status === 'connected' ? disconnect : connect}>
          {status === 'connected' ? 'End Call' : 'Start Call'}
        </Button>
      </CardContent>
    </Card>
  );
}

Phase 8: Practice Configuration UI

Create client/src/pages/Practice.tsx:

Configuration sections:

1. Script selection (optional)
2. User code input (for tracking)
3. Scenario/archetype selection
4. Difficulty level
5. Practice mode (full/partial)
6. Roleplay mode (standard/reverse)
7. Call method (web/phone-in/phone-out)
8. Voice gender

Phone integration UI:

// Call method selection
<RadioGroup value={callMethod} onValueChange={setCallMethod}>
  <RadioGroupItem value="web">Web Browser</RadioGroupItem>
  <RadioGroupItem value="phone_inbound">
    Call In: {twilioPhone?.phoneNumber}
  </RadioGroupItem>
  <RadioGroupItem value="phone_outbound">
    Receive Call
  </RadioGroupItem>
</RadioGroup>

// Phone number input (for outbound)
{callMethod === 'phone_outbound' && (
  <Input
    placeholder="+15551234567"
    value={userPhoneNumber}
    onChange={(e) => setUserPhoneNumber(e.target.value)}
  />
)}

Phase 9: History & Analytics

Create client/src/pages/History.tsx:

Display:

• Session list with filters
• Session details dialog
• Full conversation transcript
• Performance metrics
• Progress by scenario/difficulty

Create client/src/pages/Analytics.tsx:

Aggregate stats:

• Total sessions
• Average score
• Completion rates by scenario
• Difficulty progression
• Time-series charts

Testing Strategy

1. Credential validation tests:

// server/hume.credentials.test.ts
test('Hume credentials are valid', async () => {
  const token = await getHumeAccessToken();
  expect(token).toBeTruthy();
});

2. Integration tests:

// server/phone-integration.test.ts
test('System prompt includes roleplay mode', () => {
  const prompt = generateSystemPrompt({ roleplayMode: 'reverse', ... });
  expect(prompt).toContain('expert sales representative');
});

3. End-to-end flow:

• Start session → Connect to Hume → Have conversation → End session → View history

Common Patterns

Reverse Roleplay Mode

Allow users to practice by playing the "customer" role while AI demonstrates expert responses:

if (roleplayMode === 'reverse') {
  systemPrompt = `You are an expert ${profession}. The user will play the role of ${customerType} and present objections or scenarios. Demonstrate how to handle them professionally...`;
} else {
  systemPrompt = `You are a ${customerType}. The user is a ${profession} trying to ${goal}...`;
}

User Code Tracking

Enable managers to track individual users without requiring login:

// Auto-fill from profile or manual entry
const repCode = user?.repCode || manualRepCode || undefined;

// Store with session
await createPracticeSession({
  userId: ctx.user.id,
  userCode: repCode,
  ...otherParams
});

// Filter in manager dashboard
const sessions = await getUserSessions({ userCode: 'REP-1234' });

Dynamic System Prompts

Adjust AI behavior based on difficulty:

const difficultyModifiers = {
  rookie: 'Be friendly and receptive. Show clear buying signals.',
  intermediate: 'Be neutral. Require some persuasion.',
  expert: 'Be skeptical. Present common objections.',
  nightmare: 'Be hostile and dismissive. Maximum difficulty.',
};

systemPrompt += ` ${difficultyModifiers[difficulty]}`;

Deployment Checklist

• [ ] All credentials added via webdev_request_secrets
• [ ] Database migrations applied (pnpm db:push)
• [ ] Hume credentials validated (run credential tests)
• [ ] Twilio phone number configured (if using phone)
• [ ] Webhook endpoints secured (if using webhooks)
• [ ] Test web voice calls end-to-end
• [ ] Test phone calls (inbound + outbound)
• [ ] Verify conversation tracking in database
• [ ] Check analytics/history pages display correctly

Troubleshooting

Hume 401 errors:

• Regenerate API keys from Hume dashboard
• Verify HUME_API_KEY and HUME_SECRET_KEY are correct
• Check access token generation function

Twilio call failures:

• Verify phone number is in E.164 format (+15551234567)
• Check TWILIO_ACCOUNT_SID and TWILIO_AUTH_TOKEN
• Ensure Twilio account has sufficient balance
• Test with Twilio console first

WebSocket connection issues:

• Check CORS settings for Hume WebSocket
• Verify access token is valid and not expired
• Ensure browser supports WebRTC/MediaRecorder

No audio in calls:

• Check microphone permissions in browser
• Verify audio devices are working
• Test with simple WebRTC example first

References

• Hume EVI Documentation: https://dev.hume.ai/docs/empathic-voice-interface-evi/overview
• Twilio Voice API: https://www.twilio.com/docs/voice
• tRPC Documentation: https://trpc.io/docs
• Drizzle ORM: https://orm.drizzle.team/docs/overview