Text-based editing is only as good as the words underneath it. Scenaristo runs a small family of best-in-class automatic speech recognition models on your own machine, each picked for a low word error rate, fast turnaround, and the thing every cut depends on: accurate word-level timing.
Every model in the family clears the same three bars before it ships in the app.
Fewer mistranscriptions means fewer corrections, and a transcript you can trust enough to edit by deleting text.
WER-firstTranscription runs in the background and finishes quickly, so a long recording is editable in minutes, not coffee breaks.
Real-time classEach word carries its own timestamp. That is what lets a deleted phrase ripple the footage closed on a clean boundary.
Per-word stampsFastest at the top, broadest coverage at the bottom. Each model is strongest somewhere different, and Scenaristo detects the language of your recording and offers the one that fits. Every model downloads on demand, pinned by checksum, stored locally, never fetched twice.
nvidia/parakeet-tdt-0.6b-v3The default. A streaming-friendly transducer that turns an hour of audio around in a fraction of the time, with word-level timestamps baked in.
Token-and-Duration Transducer, built for real-time throughput.
nvidia/canary-1b-v2When accuracy matters more than raw speed. A heavier model that trades a little throughput for cleaner transcripts on tough audio.
A larger encoder for the lowest word error rate of the set.
CohereLabs/cohere-transcribe-03-2025A well-rounded option focused on the most common languages, with reliable timing and punctuation out of the box.
Strong on clean speech across major business languages.
openai/whisper-large-v3-turboThe widest language coverage in the family, from the model that long set the industry standard. Turbo delivers it quickly.
A distilled decoder, markedly faster than large-v3 for a small accuracy cost.
openai/whisper-large-v3The accuracy rung of the Whisper ladder. Slower than Turbo, but the strongest option on noisy audio and the long tail of languages.
The full, un-distilled model, reach for it on the hardest audio.
Model identifiers reference their public Hugging Face repositories. Parakeet, Canary, Cohere Transcribe, and Whisper are trademarks of NVIDIA, Cohere, and OpenAI respectively. Availability and exact versions may change as the family is updated.
Most recordings never need more than the default. Parakeet is fast, accurate, and covers the languages most spoken-word video is made in, so for the typical interview or talking-head clip, it just works.
When a file is noisy, multilingual, or in a language outside the faster models' range, you can switch models on download and re-run transcription. The edit, the cuts, and the timing all stay intact.
On-device · no account · no upload