Choosing a Transcription Model or Provider ​

OSTT is provider-neutral. You can use hosted cloud transcription, built-in local Whisper-compatible models, or your own external command or HTTP engine. The right choice depends on privacy, language, latency, cost, hardware, and how much setup you want to own.

If you are unsure, start with one good default and let OSTT make comparison practical: record once, then use ostt retry -m PROVIDER/MODEL to transcribe the same audio with another model.

Quick Recommendations ​

These are starting points, not benchmark rankings. Hardware, microphone quality, language, accent, noise, pricing, and provider policies change over time.

Compare Models With Retry ​

Most dictation tools hide model choice behind a global setting. OSTT saves recordings locally, so model choice becomes testable instead of theoretical.

# Record once with your current default
ostt -o first.txt

# Retry the same audio with different providers
ostt retry -m deepgram/nova-3 -o deepgram.txt
ostt retry -m openai/gpt-4o-transcribe -o openai.txt
ostt retry -m berget/KBLab/kb-whisper-large -o berget.txt
ostt retry -m whisper/turbo -o local.txt

Use this for real audio from your microphone, your accent, your room, and your vocabulary. That is more useful than a generic benchmark table.

Cloud Providers ​

Cloud providers are usually easiest when you want strong transcription without downloading local models. They also move local CPU/GPU load off your machine.

Run ostt auth login before selecting a cloud provider:

ostt auth login
ostt model

Built-In Local Whisper ​

Use local Whisper-compatible models when privacy, offline use, or predictable cost matters. OSTT's built-in local path uses whisper-rs with GGUF or ggml-*.bin model files.

Open the model picker:

ostt model

Choose Local provider to download curated models, activate a downloaded model, inspect metadata, delete model files, or add a custom Hugging Face/direct model URL.

For repeated local dictation, start Daemon Mode:

ostt daemon start
ostt launch --paste

Daemon mode keeps the active local model loaded so each transcription avoids model load time.

External Local Engines ​

Built-in local support intentionally stays focused on Whisper-compatible models. If you want faster-whisper, Parakeet, Cohere Transcribe, Speaches, LocalAI, onnx-asr, or a custom research model, run that engine yourself and let OSTT call it.

Use command/<profile> when you have a CLI or wrapper script:

[transcription]
provider = "command"
model = "parakeet"

[command.parakeet]
display_name = "Parakeet"
command = "/home/you/asr/parakeet-transcribe.sh {audio_path}"
output_format = "pcm_s16le -ar 16000"
timeout_secs = 300

Use http/<profile> when your engine exposes an OpenAI-compatible /v1/audio/transcriptions endpoint:

[transcription]
provider = "http"
model = "speaches"

[http.speaches]
display_name = "Speaches"
endpoint = "http://127.0.0.1:8000/v1/audio/transcriptions"
output_format = "pcm_s16le -ar 16000"
timeout_secs = 300

[http.speaches.params]
model = "Systran/faster-whisper-large-v3"
response_format = "json"

Then select or use it like any other model:

ostt model select http/speaches
ostt -m http/speaches --paste
ostt retry -m command/parakeet

See External Engines for the full contract.

Developer Dictation Needs More Than a Model ​

For code, product names, acronyms, APIs, and unusual proper nouns, combine model choice with OSTT's transcript cleanup tools.

Add recognition hints before transcription:

ostt keyword add Kubernetes
ostt keyword add VitePress

Fix final casing and common misrecognitions after transcription:

[text.replace]
"ostt" = "OSTT"
"api" = "API"
"github" = "GitHub"
"open ai" = "OpenAI"

Then use processing actions for transformations that need AI or shell commands:

ostt launch --paste -p clean
ostt transcribe meeting.mp3 -p summary -o summary.md

Tradeoffs ​

Related Pages ​

• Providers and Models
• Local Models
• External Engines
• GPU Acceleration
• Daemon Mode