What is the cheapest AI speech-to-text API?

The cheapest AI speech-to-text API is Deepgram Nova-2 at $0.0043/minute for pre-recorded audio. OpenAI Whisper is $0.006/minute. AssemblyAI async costs $0.0085/minute. Google Cloud Speech-to-Text starts at $0.016/minute for standard models. For most use cases, Deepgram Nova-2 offers the best combination of accuracy and cost.

How much does AI transcription cost per hour?

AI transcription costs per hour: Deepgram Nova-2 $0.26/hour, OpenAI Whisper $0.36/hour, AssemblyAI async $0.51/hour, Google Cloud STT $0.96/hour, Amazon Transcribe $1.44/hour. For 100 hours/month, costs range from $26 (Deepgram) to $144 (Amazon Transcribe). Deepgram is 5.5× cheaper than Amazon Transcribe.

Is AI speech-to-text accurate enough for production use?

Yes — modern STT APIs achieve 95-99% accuracy for clear English audio. Deepgram Nova-2 and OpenAI Whisper are the most accurate. Accuracy drops with: heavy accents, background noise, technical jargon, overlapping speakers. For production use, choose a provider with speaker diarization (speaker identification) and custom vocabulary support. Deepgram and AssemblyAI offer the best accuracy-to-cost ratio.

Cheapest AI API for Speech-to-Text

Find the cheapest AI API for speech-to-text transcription. We compared 7 providers — from $0.004 per minute.

Calculate Your Transcription Cost

Enter your monthly audio volume to see the cheapest providers for your workload.

Use case:

Audio minutes per day

Processing Mode

Days per month

Speech-to-Text API Cost Ranking

Every provider ranked by cost for a typical workload: 300 minutes/day, async batch processing.

Top Picks by Volume

Small Project (under $15/month)

Deepgram Nova-2$38.70/mo

OpenAI Whisper$54.00/mo

AssemblyAI Async$76.50/mo

Business ($50-200/month)

Deepgram Nova-2$258/mo (1K min/day)

OpenAI Whisper$360/mo

AssemblyAI Async$510/mo

Enterprise Volume ($500+/month)

Deepgram Nova-2$1,296/mo (5K min/day)

OpenAI Whisper$1,800/mo

Google Cloud STT$2,880/mo

Strategy: Async + Realtime Hybrid

Use a hybrid approach — async processing for recorded content, real-time for live applications.

Smart STT Pipeline (500 min/day)

70% recordings → Deepgram async ($0.0043/min)$45.15/mo

30% live calls → Deepgram realtime ($0.0077/min)$34.65/mo

Total with hybrid$79.80/mo (vs $135 using realtime for everything)

The hybrid approach saves 41% compared to using real-time processing for all audio. Most recorded content can be processed asynchronously at lower cost.

Find the cheapest model for your transcription workload

Enter your usage and see all providers ranked by cost. Free, no signup.

Open Savings Calculator →

Key Factors When Choosing an STT API

Per-minute pricing: STT APIs charge per minute of audio processed. Async (batch) is 40-60% cheaper than real-time streaming. If you don't need instant results, always use async.
Accuracy matters more than price: A 1% accuracy improvement can save hours of manual correction. Deepgram Nova-2 and OpenAI Whisper lead in accuracy benchmarks. The cheapest option isn't always the best value if accuracy is poor.
Speaker diarization: For meetings and multi-speaker content, speaker diarization (who said what) is essential. Deepgram, AssemblyAI, and Google all offer this. It may cost extra ($0.01-0.02/min additional).
Language support: OpenAI Whisper supports 100+ languages. Deepgram focuses on English and major European/Asian languages. Google Cloud STT has the broadest language coverage. Choose based on your target languages.
Custom vocabulary: For technical, medical, or industry-specific content, custom vocabulary support improves accuracy. AssemblyAI and Google have the best custom vocabulary options.
File size limits: Some APIs limit audio file length (e.g., 25MB for OpenAI Whisper API). For long recordings, use chunked uploads or choose a provider with no file size limits (Deepgram, Google).

Related Tools

Savings Calculator — See how much you can save by switching models
Cost Explorer — See all 42 models ranked by your usage
Cheapest AI API Finder — Find the absolute cheapest model