Fewer hallucinations
Whisper-1 can generate text that was never spoken on silent or low-quality audio. gpt-4o-transcribe is less prone to this, which matters most on long meeting recordings.
gpt-4o-transcribe is OpenAI's GPT-4o-based transcription model, built on GPT-4o rather than Whisper, with a 4.1% word error rate and fewer hallucinations on long recordings. OSTT lets you use it from a Linux or macOS terminal, bound to a hotkey and wired into your shell.
GPT-4o Transcribe
gpt-4o-transcribe delivers a 4.1% WER, compared with whisper-1's 5.3%, with the largest gains on noisy audio, accented speech, and long recordings. It is less prone to Whisper's habit of inventing text during silence. At $0.006/min, the same as whisper-1, there is no cost reason to stay on the legacy model for standard transcription work.
# ~/.config/ostt/ostt.toml
[transcription]
provider = "openai"
model = "gpt-4o-transcribe"
[openai.gpt-4o-transcribe.params]
language = "en"
prompt = "Technical dictation with project names."
include = ["logprobs"]
# Pick interactively
ostt model
# Record, transcribe with gpt-4o-transcribe, copy result
ostt launch -cWhisper-1 can generate text that was never spoken on silent or low-quality audio. gpt-4o-transcribe is less prone to this, which matters most on long meeting recordings.
gpt-4o-transcribe auto-inserts punctuation, sentence casing, and handles code-switching between languages, so you often need no separate LLM cleanup step.
Pass a prompt to gpt-4o-transcribe to guide style, domain vocabulary, and formatting. Whisper's prompt support is limited to 224 tokens and ignores instructions; gpt-4o-transcribe follows them.
OpenAI can return token log probabilities for gpt-4o-transcribe and gpt-4o-mini-transcribe with include=logprobs. OSTT validates the option and still emits clean transcript text for shell workflows.
Switch to openai/gpt-4o-transcribe-diarize when you need OpenAI's diarized JSON response. OSTT returns the combined transcript text while preserving the model-specific request options.
Send transcription to the clipboard with -c, write to a file with -o, paste into the focused app with --paste, or print to stdout. OSTT also runs Parakeet and Whisper models locally when you want to stay offline.
Run ostt model to switch between gpt-transcribe, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize, and whisper-1. One command, no config file editing required.
Workflow
ostt in the terminal.Pipeline
OSTT routes gpt-4o-transcribe output to wherever your workflow needs it. Print to stdout, copy to clipboard, write to a file, or pipe through any CLI. Use --param language=en, --param prompt=..., or --param include=logprobs for per-run OpenAI params.
# Transcribe a recording, write to file
ostt transcribe meeting.mp3 -o notes.md
# Record, process with "summary" action, copy
ostt -p summary -c
# Transcribe and pipe to custom script
ostt | ./process.sh