Imported from jeremylongshore/tons-of-skills-marketplace (skills/.curated/assemblyai-cost-tuning/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill assemblyai-cost-tuning. Copyright stays with the author (MIT).

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)

Streaming Speech-to-Text

Audio Intelligence Add-Ons

LeMUR

Instructions

Step 1: Cost Estimation Calculator

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

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

// 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

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

Step 6: Budget Alerts

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

Examples

Before enabling a feature, verify current rates and limits in AssemblyAI's authoritative billing documentation, estimate a bounded pilot from approved audio duration, and set a budget alert below the hard limit. Tag usage by opaque tenant/workload key and review a dry-run cost report before enabling bulk transcription or LeMUR jobs.

Error Handling

Resources

• AssemblyAI Pricing
• AssemblyAI Billing Dashboard
• Speech Model Comparison

Next Steps

For architecture patterns, see assemblyai-reference-architecture.