Fermion ResearchFermion Research

Speech recognition/facts checked 13 November 2026

How much does a speech-to-text API cost per hour?

Hosted transcription prices differ by almost a hundredfold for the same hour of audio. This page lists the published prices with the date they were read, and the measured cost of running an open model yourself.

Self-host a speech-to-text server

The short answer

Speech-to-text APIs cost from about $0.012 to $0.96 per audio hour at list price: hosted open models such as Whisper large-v3-turbo and Qwen3-ASR 0.6B start at $0.012 on DeepInfra, most proprietary APIs list between $0.10 and $0.36 (Microsoft MAI-Transcribe 2 $0.10, ElevenLabs Scribe v2 $0.20, AssemblyAI Universal-3.5 Pro $0.21, Deepgram Nova-3 $0.258, OpenAI whisper-1 $0.36), and Google Chirp 3 lists $0.96. Self-hosting an open model at steady volume measured $0.00204 per audio hour in our test.

Prices change often. Check the provider's own page before you commit.

List prices per audio hour

Pay-as-you-go, pre-recorded audio, converted to per hour where the provider prices per minute. Rows marked 'vendor' were read on the provider's pricing page on 2026-11-13; rows marked 'OpenRouter' are OpenRouter's model pages read on 2026-11-11.

Model (provider)Per audio hourLanguagesSource
Whisper large-v3-turbo (DeepInfra)$0.01299OpenRouter
Qwen3-ASR 0.6B (DeepInfra)$0.012see model cardOpenRouter
Qwen3-ASR 1.7B (DeepInfra)$0.02730 and 22 Chinese dialectsOpenRouter
Whisper large-v3 (DeepInfra)$0.02799OpenRouter
Whisper large-v3-turbo (Groq)$0.0499vendor
Whisper large-v3-turbo (Prime Intellect)$0.0499OpenRouter
Parakeet TDT 0.6B v3 (Prime Intellect)$0.0625 EuropeanOpenRouter
Voxtral Mini 3B (DeepInfra)$0.06 ($0.001 per minute)8 named on its cardvendor
Whisper large-v3 (Together)$0.0999OpenRouter
Parakeet TDT 0.6B v3 (Together)$0.0925 EuropeanOpenRouter
Microsoft MAI-Transcribe 2 (Azure)$0.1060OpenRouter
xAI Grok STT$0.10see providerOpenRouter
Whisper large-v3 (Groq)$0.11199vendor
Meta muse-voice-transcribe-1.0$0.18see providerOpenRouter
Voxtral Mini Transcribe 2602 (Mistral)$0.18see providerOpenRouter
ElevenLabs Scribe v2$0.20 (OpenRouter lists $0.22)90+vendor
AssemblyAI Universal-3.5 Pro$0.21 (OpenRouter lists $0.30)18vendor
Voxtral Mini Transcribe 3.0 (Mistral)$0.21see providerOpenRouter
Voxtral Small Transcribe 3.0 (Mistral)$0.24see providerOpenRouter
Deepgram Nova-3 monolingual$0.258 ($0.0043 per minute)see providervendor
OpenAI gpt-transcribe$0.27 ($0.0045 per minute)see providervendor
Deepgram Nova-3 multilingual$0.312 ($0.0052 per minute)see providervendor
OpenAI whisper-1$0.36 ($0.006 per minute)99vendor
Microsoft MAI-Transcribe 1.5 (Azure)$0.3643OpenRouter
Fish Audio$0.36see providerOpenRouter
Google Chirp 3$0.96see providerOpenRouter

What self-hosting costs

We measured Phonon-2 on an NVIDIA L4 GPU on our staging server: about 960 audio hours transcribed per GPU hour on a realistic mix of files, and a fully loaded cost of $0.00204 per audio hour at 60 % utilisation, counting the GPU and the gateway in front of it.

  • 1,000 audio hours: about $2.04.
  • 100,000 audio hours: about $204.

That figure assumes enough steady traffic to keep the GPU 60 % busy. A server costs the same whether it is busy or idle, so at low or bursty volume the cost per audio hour is higher, and engineering time to run the service is not included. Our self-hosted vs API guide works through the trade-off.

What changes the bill

  • Add-ons: custom vocabulary costs extra on some APIs (Deepgram keyterm prompting $0.0013 per minute; AssemblyAI keyterms $0.05 per hour).
  • Streaming: live transcription is usually priced higher than pre-recorded audio. Deepgram Nova-3 streaming is $0.0077 per minute against $0.0043 pre-recorded, and AssemblyAI bills streaming on session duration rather than audio sent.
  • Multilingual: Deepgram charges more for multilingual Nova-3 than for monolingual.
  • Billing granularity: per second, per 15 seconds or per minute changes the cost of short clips. Deepgram states it bills per second with no block rounding.

Questions

What is the cheapest speech-to-text API?
Among list prices read in November 2026, hosted Whisper large-v3-turbo and Qwen3-ASR 0.6B on DeepInfra cost $0.012 per audio hour (OpenRouter listing). Self-hosting an open model at steady volume measured $0.00204 per audio hour in our test.
How much does OpenAI charge for transcription?
OpenAI's pricing page lists whisper-1 at $0.006 per minute ($0.36 per hour) and gpt-transcribe at $0.0045 per minute ($0.27 per hour), read 2026-11-13.
How much does Deepgram cost per hour?
Deepgram lists Nova-3 pre-recorded at $0.0043 per minute monolingual ($0.258 per hour) and $0.0052 per minute multilingual ($0.312 per hour), read 2026-11-13.
How much does AssemblyAI cost per hour?
AssemblyAI's pricing page lists Universal-3.5 Pro at $0.21 per hour and Universal-2 at $0.15 per hour for pre-recorded audio, read 2026-11-13; keyterms prompting adds $0.05 per hour.
Is self-hosting speech-to-text cheaper than an API?
At steady volume, usually: we measured $0.00204 per audio hour for Phonon-2 on an L4 GPU at 60 % utilisation. At low volume, an idle server and the engineering time can cost more than paying per hour.

Related

Sources and method

  • Hosted list prices marked OpenRouter: OpenRouter speech-to-text model pages, read 2026-11-11.
  • Vendor pages read 2026-11-13: developers.openai.com/api/docs/pricing; deepgram.com/pricing; assemblyai.com/pricing; elevenlabs.io/pricing/api; console.groq.com/docs/speech-to-text; deepinfra.com/pricing.
  • Self-hosting cost: Fermion Research measurement of Phonon-2 on an NVIDIA L4 staging server, 2026-10-08, realistic audio mix, GPU plus gateway, 60 % utilisation.

Canonical: https://www.fermionresearch.com/speech/speech-to-text-api-cost/