assemblyai-cost-tuning

# AssemblyAI Cost Tuning

## Overview
Optimize AssemblyAI costs through model selection, feature-aware billing, and usage monitoring. AssemblyAI charges per audio hour with add-on pricing for intelligence features.

## Prerequisites
- Access to AssemblyAI billing dashboard at https://www.assemblyai.com/app
- Understanding of current usage patterns

## Actual Pricing (Pay-As-You-Go)

### Speech-to-Text (Async)
| Model | Price per Hour | Best For |
|-------|---------------|----------|
| **Best** (Universal-3) | $0.37/hr | Highest accuracy, production |
| **Nano** | $0.12/hr | High volume, cost-sensitive |

### Streaming Speech-to-Text
| Model | Price per Hour |
|-------|---------------|
| Universal Streaming | $0.47/hr |

### Audio Intelligence Add-Ons
| Feature | Additional Cost per Hour |
|---------|-------------------------|
| Speaker Diarization | $0.02/hr |
| Sentiment Analysis | $0.02/hr |
| Entity Detection | $0.08/hr |
| Auto Highlights | Included |
| Content Safety | $0.02/hr |
| IAB Categories | $0.02/hr |
| Summarization | Included (uses LeMUR) |
| PII Redaction | $0.02/hr |
| PII Audio Redaction | +processing time |

### LeMUR
| Model | Price per Input Token | Price per Output Token |
|-------|----------------------|----------------------|
| Default | ~$0.003/1K tokens | ~$0.015/1K tokens |

## Instructions

### Step 1: Cost Estimation Calculator

```typescript
interface CostEstimate {
  baseTranscriptionCost: number;
  featuresCost: number;
  totalCost: number;
  breakdown: Record<string, number>;
}

function estimateTranscriptionCost(
  audioHours: number,
  options: {
    model?: 'best' | 'nano';
    speakerLabels?: boolean;
    sentimentAnalysis?: boolean;
    entityDetection?: boolean;
    contentSafety?: boolean;
    iabCategories?: boolean;
    piiRedaction?: boolean;
  } = {}
): CostEstimate {
  const model = options.model ?? 'best';
  const baseRate = model === 'best' ? 0.37 : 0.12;
  const baseCost = audioHours * baseRate;

  const breakdown: Record<string, number> = {
    [`transcription (${model})`]: baseCost,
  };

  let featuresCost = 0;

  if (options.speakerLabels) {
    const cost = audioHours * 0.02;
    breakdown['speaker_labels'] = cost;
    featuresCost += cost;
  }
  if (options.sentimentAnalysis) {
    const cost = audioHours * 0.02;
    breakdown['sentiment_analysis'] = cost;
    featuresCost += cost;
  }
  if (options.entityDetection) {
    const cost = audioHours * 0.08;
    breakdown['entity_detection'] = cost;
    featuresCost += cost;
  }
  if (options.contentSafety) {
    const cost = audioHours * 0.02;
    breakdown['content_safety'] = cost;
    featuresCost += cost;
  }
  if (options.iabCategories) {
    const cost = audioHours * 0.02;
    breakdown['iab_categories'] = cost;
    featuresCost += cost;
  }
  if (options.piiRedaction) {
    const cost = audioHours * 0.02;
    breakdown['pii_redaction'] = cost;
    featuresCost += cost;
  }

  return {
    baseTranscriptionCost: baseCost,
    featuresCost,
    totalCost: baseCost + featuresCost,
    breakdown,
  };
}

// Example: 100 hours with Best model + diarization + sentiment
const estimate = estimateTranscriptionCost(100, {
  model: 'best',
  speakerLabels: true,
  sentimentAnalysis: true,
});
// Result: $37 (transcription) + $2 (speakers) + $2 (sentiment) = $41
```

### Step 2: Model Selection Strategy

```typescript
import { AssemblyAI } from 'assemblyai';

const client = new AssemblyAI({
  apiKey: process.env.ASSEMBLYAI_API_KEY!,
});

// Use Nano for high-volume, cost-sensitive workloads
// - 3x cheaper than Best ($0.12 vs $0.37)
// - Good enough for search indexing, keyword detection
const cheapTranscript = await client.transcripts.transcribe({
  audio: audioUrl,
  speech_model: 'nano',
});

// Use Best for critical, accuracy-sensitive workloads
// - Medical transcription, legal proceedings, compliance
// - Supports word_boost for domain terminology
const accurateTranscript = await client.transcripts.transcribe({
  audio: audioUrl,
  speech_model: 'best',
  word_boost: ['specialized', 'domain', 'terms'],
  boost_param: 'high',
});
```

### Step 3: Feature Budget — Only Enable What You Need

```typescript
// EXPENSIVE: All features enabled ($0.37 + $0.16 = $0.53/hr)
const expensive = await client.transcripts.transcribe({
  audio: audioUrl,
  speech_model: 'best',        // $0.37/hr
  speaker_labels: true,         // +$0.02/hr
  sentiment_analysis: true,     // +$0.02/hr
  entity_detection: true,       // +$0.08/hr
  content_safety: true,         // +$0.02/hr
  iab_categories: true,         // +$0.02/hr
});

// CHEAP: Only what's needed ($0.12 + $0.02 = $0.14/hr)
const cheap = await client.transcripts.transcribe({
  audio: audioUrl,
  speech_model: 'nano',         // $0.12/hr
  speaker_labels: true,         // +$0.02/hr
  // Skip features you don't use
});
```

### Step 4: Usage Tracking

```typescript
class AssemblyAIUsageTracker {
  private totalAudioHours = 0;
  private totalCost = 0;
  private transcriptionCount = 0;

  track(audioDurationSeconds: number, model: 'best' | 'nano', features: string[]) {
    const hours = audioDurationSeconds / 3600;
    this.totalAudioHours += hours;
    this.transcriptionCount++;

    const estimate = estimateTranscriptionCost(hours, {
      model,
      speakerLabels: features.includes('speaker_labels'),
      sentimentAnalysis: features.includes('sentiment_analysis'),
      entityDetection: features.includes('entity_detection'),
      contentSafety: features.includes('content_safety'),
      iabCategories: features.includes('iab_categories'),
      piiRedaction: features.includes('redact_pii'),
    });

    this.totalCost += estimate.totalCost;

    return estimate;
  }

  getSummary() {
    return {
      totalAudioHours: this.totalAudioHours.toFixed(2),
      totalCost: `$${this.totalCost.toFixed(2)}`,
      transcriptionCount: this.transcriptionCount,
      avgCostPerTranscription: `$${(this.totalCost / this.transcriptionCount).toFixed(4)}`,
    };
  }
}
```

### Step 5: Cost Reduction Strategies

| Strategy | Savings | Trade-off |
|----------|---------|-----------|
| Use Nano instead of Best | 68% cheaper | Slightly lower accuracy |
| Disable unused features | Up to $0.16/hr | Missing insights |
| Cache transcript results | Eliminate re-fetch costs | Stale data risk |
| Use LeMUR instead of per-feature AI | Often cheaper for summaries | Different output format |
| Pre-filter audio (skip silence) | Proportional savings | Requires preprocessing |
| Batch with webhooks | No savings, but better throughput | More complex architecture |

### Step 6: Budget Alerts

```typescript
const MONTHLY_BUDGET = 100; // $100
const tracker = new AssemblyAIUsageTracker();

// After each transcription
const estimate = tracker.track(transcript.audio_duration ?? 0, 'best', ['speaker_labels']);
const summary = tracker.getSummary();

if (parseFloat(summary.totalCost.replace('$', '')) > MONTHLY_BUDGET * 0.8) {
  console.warn(`Budget warning: ${summary.totalCost} of $${MONTHLY_BUDGET} used`);
  // Send alert to Slack, email, etc.
}
```

## Output
- Accurate cost estimation with feature-level breakdown
- Model selection strategy (Best vs. Nano)
- Feature budgeting to eliminate unnecessary costs
- Usage tracking with budget alerts
- Cost reduction strategies ranked by impact

## Error Handling
| Issue | Cause | Solution |
|-------|-------|----------|
| Unexpected high bill | Entity detection enabled everywhere | Audit features per endpoint |
| Nano accuracy too low | Wrong model for use case | Switch critical paths to Best |
| Budget exceeded | No monitoring | Implement usage tracker + alerts |
| Double billing | Re-transcribing same audio | Cache transcript IDs, check before submitting |

## Resources
- [AssemblyAI Pricing](https://www.assemblyai.com/pricing)
- [AssemblyAI Billing Dashboard](https://www.assemblyai.com/app)
- [Speech Model Comparison](https://www.assemblyai.com/docs/speech-to-text/speech-recognition)

## Next Steps
For architecture patterns, see `assemblyai-reference-architecture`.