Cloud Models
OSTT works with eight cloud transcription providers, each with your own API key: OpenAI, Deepgram, Groq, DeepInfra, AssemblyAI, Berget, ElevenLabs, and Mistral. Audio is sent to the provider you choose, and usage is billed by that provider.
Local models are the default recommendation: they run offline with full GPU support and keep audio on your machine. The cloud is the better fit when you need provider features such as speaker labels, a hosted model with no local equivalent, or transcription on a machine without the hardware for a large local model. See Choosing a Model for the comparison.
Set Up
Save an API key, then pick a model:
ostt auth login
ostt modelModels from providers you have a key for are listed in the Cloud models section of the picker. Keys are stored apart from the config, in ~/.local/share/ostt/credentials with 0600 permissions.
When the Cloud Fits
| Need | Good starting point | Why |
|---|---|---|
| No model download | OpenAI openai/gpt-4o-transcribe or Deepgram deepgram/nova-3 | Hosted APIs need only an API key. |
| Strong general quality | OpenAI openai/gpt-transcribe, Deepgram deepgram/nova-3, or AssemblyAI assemblyai/universal-3-5-pro | Current general-purpose models with useful params for language, formatting, prompts, or keyterms. |
| Speaker labels and long meetings | Providers with diarization and formatting params | Long files often benefit from provider-specific params such as diarization, formatting, language hints, and prompts. |
| Swedish or EU hosting | Berget berget/KBLab/kb-whisper-large; berget/klang/pianissimo | Swedish-focused models hosted in Europe. Pianissimo transcribes after recording stops and does not support keyword boosting. |
| Norwegian | Berget berget/NbAiLab/nb-whisper-large | NB-Whisper is documented for Norwegian, Bokmål, Nynorsk, and English. |
Providers
| Provider | Good for | Watch out for |
|---|---|---|
| OpenAI | GPT Transcribe, GPT-4o transcription, GPT-4o Mini, hosted Whisper, prompt hints, diarization model | OSTT returns plain text even when JSON metadata is requested. |
| Deepgram | Nova models, low-latency cloud transcription, formatting, diarization, language detection, keyterms | Advanced params are provider-specific. Pick nova-3 or nova-2 intentionally. |
| Groq | Very fast hosted Whisper variants and OpenAI-compatible request shape | Groq model choices differ in accuracy, cost, and translation support. |
| DeepInfra | Hosted open speech-recognition models, including Whisper, Qwen3 ASR, Nemotron, and Voxtral options | Model availability and pricing can change. Check the provider docs. |
| AssemblyAI | Universal 3.5 Pro, Universal 2, promptable transcription, speaker labels, language detection, keyterms | Async-provider behavior and params differ from OpenAI-style endpoints. |
| Berget | Swedish Pianissimo, Swedish and Norwegian optimized Whisper models, European hosting | Choose model-specific params: Pianissimo does not support Whisper prompts or hotwords. |
| ElevenLabs | Scribe transcription and multilingual speech-to-text workflows | Advanced diarization and role params have provider-specific constraints. |
| Mistral | Voxtral transcription, context bias, diarization, timestamp granularity | OSTT uses the synchronous transcription endpoint, not streaming. |
Each provider page lists its models and params. Providers and Models explains how params are set in the config and per run.
Compare With a Local Model
Recordings made with a cloud model active are MP3 by default, which local models cannot read. To compare a cloud model with a local one on the same audio, record with the local model active and retry with the cloud model:
ostt -o local.txt
ostt retry -m deepgram/nova-3 -o deepgram.txtSee Audio Format.