Every transcription API ranked for your volume
Same audio hours, cheapest first.
| Provider / tier | $/minute | Your $/mo |
|---|
Add-ons cost extra — the numbers above are base transcription only
Speaker diarization, keyterm/keyword boosting and specialty modes are billed as separate surcharges on top of the base rate on most providers. Verified per-provider, August 2026:
| Provider | Add-on | Surcharge |
|---|---|---|
| AssemblyAI | Speaker diarization (async) | +$0.02/hr (≈+$0.00033/min) |
| AssemblyAI | Keyterms prompting | +$0.05/hr (≈+$0.00083/min) |
| AssemblyAI | Medical Mode | +$0.15/hr (≈+$0.0025/min) |
| AssemblyAI | Streaming diarization | +$0.12/hr (≈+$0.002/min) |
| Deepgram | Speaker diarization | +$0.0020/min flat |
On a 500-hour/month batch job, AssemblyAI diarization alone adds about $10/mo — small next to the base bill, but keyterm boosting plus diarization plus a specialty mode can stack to a meaningful percentage on top. Always price the exact feature set you'll actually use, not just the headline per-minute rate.
Per-minute pricing, five-fold spread
Speech-to-text APIs bill by the minute of audio, not by file size or word count, which makes volume planning simple: estimate your hours, multiply by sixty, multiply by the per-minute rate. The catch is how wide the spread is. Deepgram and AssemblyAI's batch tiers transcribe a minute for well under half a cent, OpenAI's Whisper API sits around six tenths of a cent, and the hyperscaler services — Google, AWS Transcribe, Azure — typically charge two to four times more per minute. Over thousands of hours a month, that gap turns into thousands of dollars a year for identical output.
Batch vs real-time is the other big dial
The second cost lever is latency. Real-time streaming transcription — live captions, voice agents, call monitoring — holds an open connection and returns words as they're spoken, and providers charge a premium for it. Batch transcription of recorded files runs in the background and is markedly cheaper. If your product can wait a few minutes — meeting summaries, podcast transcripts, archived call analytics — use batch and pocket the difference. Only pay for real-time where the user is genuinely waiting on the words.
When to self-host Whisper
Open-source Whisper removes the per-minute fee but adds GPU, ops and idle cost. Below a few hundred hours a month the hosted API wins on both price and simplicity; above that, a well-utilised GPU running near the clock can beat it. Find your crossover on the self-hosted vs API calculator. If transcription feeds an LLM — summaries, Q&A over calls — add that token cost on the LLM token cost calculator, and total the whole pipeline on the AI app cost estimator.
Related tools & guides
AI app cost estimator · Self-hosted vs API calculator · AI voice agent cost calculator · Cheapest LLM API · All AI APIs