Fermion ResearchFermion Research

Speech recognition/facts checked 13 November 2026

Self-hosted vs API speech-to-text: how to decide

An API is the fastest way to start and the easiest to scale down. Running an open model yourself wins on cost at steady volume and on keeping audio inside your environment. Most teams should decide on volume and data rules first.

Self-host a speech-to-text server

The short answer

Self-host speech-to-text when you have steady volume or audio that must not leave your environment; use a hosted API when volume is small or spiky, or when you need features open models lack, such as speaker labels, 90+ languages or managed streaming at scale. At steady volume, a measured self-hosted open model cost $0.00204 per audio hour, against $0.012 to $0.96 for hosted list prices.

Side by side

Self-hosted figures are Fermion Research measurements; API figures are list prices and features on the providers' own pages, read 2026-11-11 to 2026-11-13.

FactorSelf-hosted open modelHosted API
Cost per audio hour$0.00204 measured (Phonon-2, L4 GPU, 60 % utilisation)$0.012 (hosted Whisper turbo) to $0.96 (Google Chirp 3), list prices
Fixed costservers whether busy or idle, plus engineering timenone; pay per use
Latencyno network round trip when on the same machine or networkupload time plus the provider's queue
Where audio goesstays in your environmentsent to the provider; terms vary
English accuracy (Open ASR Leaderboard, seven sets)best open row 4.29 % word errorbest closed row 3.68 % word error
Languages25 (Phonon-2.5) to 99 (Whisper)up to 90+ (ElevenLabs Scribe v2)
Featurestranscription, word timestamps, hotwords on some modelsoften adds speaker labels, redaction, summaries
Scalingyou add machinesthe provider scales

When a hosted API is the better choice

  • Low or unpredictable volume: a few hundred hours a month costs little at any list price, and an idle server would cost more.
  • You need speaker labels, redaction or summaries out of the box.
  • You need languages outside the open models you can run, or the lowest possible English error rate: on the Open ASR Leaderboard snapshot of 2026-09-25, the lowest seven-set averages (Zoom Scribe v2 Pro 3.68 %, Modulate 3.92 %, Azure Speech 3.93 %) are closed APIs, against 4.29 % for the best open row.
  • No one on the team wants to run a service.

When self-hosting is the better choice

  • Steady volume: at our measured $0.00204 per audio hour, 100,000 audio hours cost about $204 of GPU and gateway time.
  • Audio that cannot leave your environment for legal or contractual reasons. The model runs in your cloud, data centre or on the device.
  • Hardware you already own: Phonon-2 runs at 143 times realtime on eight AMD Zen 5 cores (Phonon-2's published figure), so an ordinary CPU server can transcribe more than a hundred hours of audio per hour.
  • Control: a fixed model version, no deprecations, and the ability to re-run any benchmark yourself.
pip install fermion-research
fermion serve phonon-2.5 --host 0.0.0.0 --port 8000 --api-key "$FERMION_API_KEY"

A simple way to decide

Estimate monthly audio hours. Multiply by the API price you would pay, then compare with the monthly cost of the smallest server that handles your peak load plus a few hours a month of someone's time to run it. If the API bill is the smaller number, use the API and revisit when volume grows. If audio cannot leave your environment, self-host regardless of the arithmetic.

Questions

Is it cheaper to self-host speech-to-text?
At steady volume, yes: we measured $0.00204 per audio hour for Phonon-2 on an L4 GPU at 60 % utilisation, while hosted list prices run from $0.012 to $0.96 per hour. At low volume the fixed cost of a server can make an API cheaper.
Is self-hosted speech recognition as accurate as an API?
Close on English: the best open row on the Open ASR Leaderboard seven-set average is 4.29 %, against 3.68 % for the lowest closed API in the 2026-09-25 snapshot. Run both on your own audio, since domain matters more than these gaps.
Do I need a GPU to self-host speech-to-text?
Not for small models. The Phonon models run on plain CPUs on Linux, Windows and macOS; Phonon-2 runs at 143 times realtime on eight AMD Zen 5 cores (its published figure).
Which is more private, self-hosted or API?
Self-hosted, because audio stays in your environment. With an API, the provider's terms decide retention and use.
Can I switch between a self-hosted model and an API later?
Yes, if both use the same request shape. fermion serve exposes POST /v1/audio/transcriptions, the route shape OpenAI's transcription API uses, so a client can move by changing its base URL.

Related

Sources and method

  • 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.
  • Hosted list prices marked OpenRouter: OpenRouter speech-to-text model pages, read 2026-11-11.
  • Vendor prices and features read 2026-11-13: deepgram.com/pricing, assemblyai.com/pricing, elevenlabs.io/pricing/api, developers.openai.com/api/docs/pricing.
  • Rival English averages: Open ASR Leaderboard, seven English sets (LibriSpeech clean and other, AMI, Earnings-22, GigaSpeech, SPGISpeech, VoxPopuli), snapshot 2026-09-25.
  • Phonon-2 CPU speed: Phonon-2 model page, published figure. No Phonon-2.5 speed figures are published.

Canonical: https://www.fermionresearch.com/speech/self-hosted-vs-api-speech-to-text/