ElevenLabs Scribe v2 scores a 2.2% word error rate on the Artificial Analysis speech-to-text benchmark, lower than OpenAI GPT Transcribe, AssemblyAI Universal 3.5 Pro, and Deepgram Nova-3. It transcribes in 90+ languages, tags audio events, and detects entities. OSTT connects it directly to your terminal, hotkey, and shell.
ElevenLabs Scribe v2
On the Artificial Analysis AA-WER v2 benchmark, which weights real-world audio including diverse accents and domain-specific language, Scribe v2 scores 2.2%, ahead of AssemblyAI Universal 3.5 Pro (3.0%), OpenAI GPT Transcribe (3.3%), and Deepgram Nova-3 (5.2%). It handles long-form audio with consistent accuracy across speaker changes, accents, and recording conditions. It tags audio events such as laughter and footsteps, detects 65 entity types, and labels up to 32 speakers. OSTT brings all of this to your keyboard.
# ~/.config/ostt/ostt.toml
[transcription]
provider = "elevenlabs"
model = "scribe_v2"
[elevenlabs.scribe_v2.params]
# language_code = "eng" # Optional: set to force language
# diarize = true
# keyterms = ["OSTT", "Scribe"]
# Pick interactively
ostt model
# Record with hotkey, transcribe with Scribe v2, copy to clipboard
ostt launch -cScribe v2 scores 2.2% on the Artificial Analysis AA-WER leaderboard, ahead of AssemblyAI Universal 3.5 Pro (3.0%), OpenAI GPT Transcribe (3.3%), and Deepgram Nova-3 (5.2%). Use scribe_v2_medical for clinical audio and medical terminology.
Scribe v2 supports 90+ languages with automatic language detection. It handles mid-file language switches without configuration. Set a language_code in ostt.toml to lock a specific language and improve accuracy when the source is known.
Turn on tag_audio_events and Scribe v2 labels non-speech events such as laughter, applause, and background noise. Useful for meeting notes, research transcription, and media workflows where context matters beyond the words spoken.
Built-in detection across 65 entity types including PII, health data, and payment details, with precise timestamps. Set entity_detection, or entity_redaction to redact them. Useful for redaction workflows and structured audio analysis.
Turn on diarize to identify and label up to 32 speakers with precise word-level timestamps. Set num_speakers when you know the count.
Provide up to 1,000 domain-specific terms, product names, or technical vocabulary. Scribe v2 applies them in context, not just as keyword matching. OSTT sends your saved ostt keyword terms as keyterms unless you set keyterms yourself.
OSTT also runs Parakeet and Whisper models locally, with Metal, CUDA, or Vulkan builds for GPU acceleration. Add --paste to send the transcript straight into the focused app.
Workflow
ostt in the terminal.Pipeline
OSTT connects Scribe v2 to your shell like any other Unix tool. Transcription output lands on stdout, so you can pipe it through jq, sed, or any CLI tool. Use -p to chain AI processing actions on the result. Add technical vocabulary to OSTT keywords once and improve accuracy across every recording.
# Transcribe an interview with Scribe v2
ostt transcribe interview.mp3 -o transcript.txt
# Record, process with AI action, copy result
ostt -p clean -c
# Transcribe a long recording, summarize, write to file
ostt transcribe lecture.mp3 -p summary -o notes.md