Speech-to-Text Integration

​

Available Providers

​

Deepgram (Recommended - Best Performance)

• Lowest latency (200ms)
• Highest accuracy
• Excellent cost/performance ratio
• Best for real-time conversations

​

AssemblyAI (Best for Accuracy)

• Exceptional accuracy
• Great documentation
• Reliable service
• Good cost structure

​

Cartesia (Enterprise Grade)

• Enterprise-level performance
• Custom models available
• Premium support
• Dedicated infrastructure

• Latency: < 250ms
• Accuracy: 96%+
• Cost: Custom pricing
• Languages: 40+
• Special Features: Domain adaptation, custom models

​

Provider Comparison

​

Setup

​

Deepgram Setup (Recommended)

1. Visit deepgram.com
2. Sign up for free account
3. Verify email

1. Go to Console
2. Click “API Keys” in left menu
3. Click “Create a new API Key”
4. Select “Scoped API Key”
5. Assign “speech-to-text” scope
6. Copy key

1. Go to Settings → Developer Settings
2. Click “API Keys”
3. Select “Deepgram” from provider list
4. Paste API key
5. Click “Test Connection”
6. Save

1. Create/Edit Agent
2. Under “Speech-to-Text”, select:
   • nova-2 (default, best accuracy)
   • nova (faster)
   • enhanced (most accurate)
3. Select language and options
4. Save

​

AssemblyAI Setup

1. Visit assemblyai.com
2. Sign up for account
3. Verify email

1. Go to Dashboard
2. Click “API Keys” in sidebar
3. Copy API key under “Your API Token”

1. Go to Settings → Developer Settings
2. Click “API Keys”
3. Select “AssemblyAI” from provider list
4. Paste API key
5. Click “Test Connection”
6. Save

1. Create/Edit Agent
2. Under “Speech-to-Text”, select:
   • “Standard” (default)
   • “Enhanced” (more accurate)
3. Configure language
4. Enable diarization if needed
5. Save

​

Cartesia Setup

1. Visit cartesia.ai
2. Request enterprise access
3. Complete onboarding
4. Receive API credentials

1. Go to Settings → Developer Settings
2. Click “API Keys”
3. Select “Cartesia” from provider list
4. Paste API key
5. Click “Test Connection”
6. Save

1. Create/Edit Agent
2. Under “Speech-to-Text”, select Cartesia
3. Configure custom model (if provided)
4. Set language and options
5. Save

​

Cost Comparison

​

Monthly Costs (1000 calls, 2 min avg)

• 1000 calls × 2 minutes = 2000 minutes
• 2000 × 0.0043 USD = 8.60 USD/month
• Cost per call: 0.0086 USD

• 1000 calls × 2 minutes = 2000 minutes
• 2000 × 0.0075 USD = 15 USD/month
• Cost per call: 0.015 USD

• Custom pricing (typically 5000-10000 USD/month for enterprise)

​

Language Support

​

Deepgram Languages

English (US, UK, Australia, India, New Zealand)
Spanish, French, German, Italian, Dutch, Russian
Portuguese, Mandarin, Cantonese, Korean, Japanese
Arabic, Hindi, and 20+ more

1. Edit Agent
2. Under STT, select language
3. Use language codes:
   - en (English)
   - es (Spanish)
   - fr (French)
   - de (German)
   - etc.
4. Save

​

AssemblyAI Languages

1. Set language to "Auto"
2. System detects language automatically
3. Great for multilingual operations

​

Configuration Options

​

Model Selection

​

Deepgram Models

Best For: Most use cases
Accuracy: Highest
Speed: Fast
Cost: Standard
Recommended: YES ✓

Best For: Budget-conscious
Accuracy: Excellent
Speed: Very Fast
Cost: Standard

Best For: Noisy environments
Accuracy: Very High
Speed: Standard
Cost: Standard

​

Key Configuration Options

Enabled: Add punctuation to transcripts
Disabled: Raw text without punctuation
Default: Enabled
Recommendation: Enable for better readability

Enabled: Replace profanity with [profanity]
Disabled: Include all words
Default: Disabled
Recommendation: Enable for customer interactions

Enabled: Convert "one" to "1"
Disabled: Keep as written text
Default: Enabled
Recommendation: Enable for phone numbers, quantities

Enabled: Identify different speakers
Disabled: Single speaker transcript
Cost: Additional (check pricing)
Recommendation: Enable for multi-party calls

​

Advanced Features

​

Real-Time Transcription

Advantage: Immediate feedback
Use Case: Live coaching, monitoring
Cost: Higher (streaming mode)
Latency: 200-500ms

​

Accuracy Enhancements

1. Provide domain vocabulary
2. Add expected phrases
3. Use correct language setting
4. Choose higher-accuracy model

Your company: "Acme Corp"
Products: "WidgetPro", "GadgetMax"
Add to STT configuration
Result: Better recognition of custom terms

​

Sentiment Analysis

Sentiment Detection:
- Positive, Negative, Neutral
- Confidence score
- Available in: AssemblyAI

​

Troubleshooting

​

Best Practices

​

1. Start with Deepgram Nova 2

Best balance of:
- Accuracy (highest)
- Speed (fast)
- Cost (competitive)
- Languages (40+)

​

2. Monitor Accuracy

Weekly Review:
- Sample 10 calls
- Check transcript accuracy
- Note problem phrases
- Adjust configuration if needed

​

3. Use Language Setting

Always specify language explicitly
Auto-detection works but is slower
Specified language: 5-10ms faster

​

4. Enable Helpful Features

✓ Enable: Punctuation (readability)
✓ Enable: Number conversion (clarity)
✓ Consider: Profanity filter (compliance)
✓ Consider: Diarization (if multi-party)

​

Performance Tips

​

Optimize Response Time

Action               | Impact    | Recommendation
---------------------|-----------|----------------
Lower-latency model  | Critical  | Deepgram Nova 2
Specify language     | ~10ms     | Always specify
Enable streaming     | 200-300ms | Default in CallIntel
Reduce audio quality | -5ms      | Not recommended

​

Monitor Quality

- Transcription accuracy %
- Processing latency (ms)
- Cost per minute
- Language breakdown
- Error rate

​

See Also

​

Support