Best speech-to-text APIs for calls
Developer APIs for transcribing phone audio in real time or after the fact, compared on the per-minute rate each vendor actually publishes.
Deepgram wins on price for real-time transcription of English calls: its Nova-3 model is the cheapest per-minute rate checked here, with further volume discounts on its Growth tier. AssemblyAI wins instead if you want streaming and pre-recorded transcription priced and documented on one page with the widest add-on menu, such as speaker diarization and PII redaction, billed separately.
The short answer
- Deepgram's Nova-3 model transcribes at a promotional rate of $0.0048/minute (regular $0.0077/minute) for English, with $200 in free credit on signup, per deepgram.com/pricing.
- AssemblyAI prices streaming and pre-recorded audio separately: its Universal-Streaming model runs $0.15/hour ($0.0025/minute) while its pre-recorded Universal-2 model is also $0.15/hour, with $50 free credit and no card required, per assemblyai.com/pricing.
- Every published per-minute rate here covers transcription only. None of it includes the language model that reasons over the transcript, the text-to-speech that answers back, or the telephony carrying the call.
Speech-to-text APIs compared
| Provider | Real-time price | Free credit | Add-ons | Best for | Not for |
|---|---|---|---|---|---|
| Deepgram | $0.0048-$0.0078/min (Nova-3, Flux, promotional) | $200 | Growth tier prepaid volume discount up to 20% | Lowest cost English transcription at real volume | A team that wants one flat, unpromoted rate with no tiering |
| AssemblyAI | $0.15/hr streaming (Universal-Streaming), $0.15-$0.21/hr pre-recorded | $50, no card required | Diarization, medical mode, keyterm prompting, all billed separately | A team that wants transcription and add-ons documented on one clear page | A team that only needs bare transcription and doesn't want add-on line items |
| Google Cloud Speech-to-Text | Pricing on request via the interactive calculator | Google Cloud free-tier credit (general, not STT-specific) | Chirp/enhanced models, speaker diarization, custom models | A team already billing everything through Google Cloud | A team that wants a rate it can read without opening a calculator |
Figures checked on deepgram.com/pricing and assemblyai.com/pricing, September 2026. Google Cloud's speech-to-text pricing page renders its rate table through an interactive calculator rather than a static price list, so no flat per-minute figure is quoted here; the vendor's own pricing page is cloud.google.com/speech-to-text/pricing.
How the three differ in practice
Deepgram
Built around a family of named models (Nova-3, Flux) rather than one flat rate, with monolingual English pricing lower than multilingual detection. The current promotional rates are meaningfully below the vendor's own listed regular rates, so a rate quoted today is not guaranteed to hold at renewal. Not for: a team that wants price certainty without checking back after a promotional period ends.
AssemblyAI
Bills real-time streaming by WebSocket session duration (open-to-close time), not by the audio actually sent, which means a call left connected but silent still accrues cost. Its add-on menu, from PII redaction to medical-mode vocabulary, is itemized and stacks additively on top of the base rate rather than bundled into a single higher tier. Not for: a team that wants one number covering everything with no line items to track.
Google Cloud Speech-to-Text
Priced through Google Cloud's standard consumption model, billed in second increments of audio processed, with standard and enhanced or Chirp-family model tiers. The pricing page itself does not render as a static table; the rate depends on model, region and audio type selected in an interactive calculator, which makes it harder to quote a single number with confidence. Not for: comparing costs quickly without stepping through the calculator first.
When none of these is the answer
If you are not building a custom voice pipeline at all, a bare transcription API is the wrong layer to buy. A finished voice agent platform like Vapi or Retell AI already bundles a speech-to-text provider into its per-minute rate, so evaluating STT vendors separately only matters once you are assembling the pipeline yourself rather than buying it assembled.
SalesCrew is not included in the ranking above because it does not offer a speech-to-text API; it runs outbound calls through Vapi as one channel and logs the outcome against a contact record. Inbound call transcription and voicemail-to-text are on the roadmap, not shipped, so SalesCrew is not a candidate in a speech-to-text comparison today.
Questions
- What does speech-to-text for a phone call actually cost?
- The published per-minute rate is only the transcription cost. A live voice agent also pays for telephony to carry the call, a language model to reason over the transcript, and text-to-speech to answer back, so the STT line item is usually the smallest of the four, not the total.
- Is streaming transcription more expensive than transcribing a recording afterward?
- Yes, usually. AssemblyAI prices real-time streaming separately from pre-recorded audio, and its streaming tiers run higher per hour than its async models because the service has to hold a connection open and return words with low latency instead of processing a finished file.
- Do I need a paid plan to test one of these?
- No. Deepgram gives new accounts $200 in free credit and AssemblyAI gives $50, both usable against real usage with no credit card required at signup, per each vendor's own pricing page.
- Which one is cheapest for a high-volume call center?
- On the rates checked here, Deepgram's Nova-3 models are the lowest per-minute cost for English transcription, with further discounts on its Growth tier for prepaid annual volume. AssemblyAI's async model is competitive for calls transcribed after the fact rather than in real time.