learn.howToCalculate

learn.whatIsHeading

The Speech-to-Text API Cost calculator estimates the expense of transcribing audio using services like OpenAI Whisper API, Deepgram, AssemblyAI, Google Cloud Speech-to-Text, and AWS Transcribe. It covers both batch transcription and real-time streaming pricing.

Formel

Transcription Cost = Audio Duration (minutes) × Price per Minute

D: Audio Duration (minutes) — Total audio duration to transcribe
P: Price per Minute ($/minute) — Service rate per minute of audio
F: Features (list) — Additional features (diarization, timestamps, language detection)
M: Mode (batch/streaming) — Batch transcription vs. real-time streaming

Trin-for-trin guide

1Enter the total audio duration to transcribe (minutes or hours)
2Select the speech-to-text service and quality tier
3Specify features needed: speaker diarization, timestamps, language detection
4View cost comparison across providers with feature parity notes

Løste eksempler

Input

1,000 hours of podcast audio, OpenAI Whisper API

Resultat

Cost: 60,000 minutes × $0.006/min = $360. Deepgram Nova-2: 60,000 × $0.0043/min = $258. AssemblyAI: 60,000 × $0.0065/min = $390.

Input

Real-time transcription for a call center, 50,000 minutes/month

Resultat

Deepgram streaming: 50,000 × $0.0059/min = $295/month. Google Speech-to-Text: 50,000 × $0.009/min = $450/month.

Almindelige fejl at undgå

✕Not checking whether pricing includes speaker diarization and timestamps or if those are add-on features
✕Comparing batch pricing to real-time streaming pricing, which is typically 20-40% more expensive
✕Forgetting about audio preprocessing costs (format conversion, silence trimming) that can reduce billable minutes

Ofte stillede spørgsmål

Which speech-to-text service is cheapest?

For batch transcription, Deepgram Nova-2 is typically cheapest at $0.0043/minute ($0.26/hour). OpenAI Whisper API is $0.006/minute ($0.36/hour). Self-hosted Whisper on a GPU is cheapest at scale: an A10G at $0.60/hr processes ~180 min/hr audio, costing ~$0.003/minute.

How accurate is AI speech-to-text compared to human transcription?

Top AI models (Whisper large-v3, Deepgram Nova-2) achieve 5-10% word error rate on clean audio, approaching human transcriptionist accuracy (~3-5% WER). Accuracy drops significantly with background noise, accents, technical jargon, and multiple overlapping speakers. For legal or medical use, human review of AI transcription is still recommended.

Klar til at beregne? Prøv den gratis Speech-to-Text API Cost-beregner

Prøv det selv →