Local API (Premium)

New Local API page under "Developer" in the sidebar. It lets other apps and scripts on your computer send audio or video to ScribeForge and get the transcript back automatically.

Works with apps that support OpenAI's speech-to-text: point them at ScribeForge's address and they transcribe on your computer instead of in the cloud.

Start or stop the server with one click. If you leave it on, it starts again when you reopen ScribeForge.

The server only accepts connections from your own computer, and every request needs your private API key. You can show, copy or regenerate the key at any time.

Transcribe a file you upload, a file already on your computer, or a YouTube or other video link. Long files run in the background and report their progress.

Get results as plain text, subtitles (SRT or VTT) or detailed JSON with timestamps. You can also search your library and download any transcript.

Everything transcribed through the API is saved to your library, like any other transcript.

Requests wait in a queue and are transcribed one at a time, so your computer stays responsive.

The "Getting started" tab explains the API in plain language and walks you through setup in 3 steps. It includes a glossary and answers to common problems.

"Try it now" buttons check your connection and transcribe a test file through the API, so you can make sure everything works before connecting another app.

Ready-to-copy examples for PowerShell, cURL, Python, Node.js and automation tools such as n8n. Your address and key are already filled in.

The "Server" tab shows recent jobs and a log of incoming requests. Each list scrolls in its own panel.

Transcription Accuracy

Fixed: Whisper transcriptions sometimes stopped with a "failed to encode" or "failed to decode" error, even on the first file after opening the app. This affected the Transcribe page, Batch, Watch Folder and the Local API.

Long recordings no longer get stuck repeating one phrase. Before, one misheard sentence could repeat until the end of the file, and the rest of the recording was lost.

ScribeForge now skips silence and music before transcribing, which removes most made-up text in quiet parts. It downloads a small (1 MB) helper model the first time you transcribe, and works without it when you are offline.

If a Whisper model skips part of the speech, ScribeForge now goes back and transcribes that part again. In a 1-hour test, this recovered 48 seconds that were missing before.

Words repeated over and over inside a sentence ("why why why why…") are now reduced to one, and the rest of the sentence is kept.

Short replies like "Thank you" are no longer removed from transcripts.

Whisper transcripts are more accurate, because Whisper now always checks several possible wordings and keeps the best one.

Parakeet no longer writes some words twice.

Parakeet transcripts are now split into sentences with accurate timestamps, instead of 30-second blocks.

Parakeet now splits audio only at pauses, so words are no longer cut in half.

For Urdu, we recommend Whisper Large V3 Turbo: it gave the clearest and most complete transcript of a 1-hour Urdu class.

AI Models

Added OpenAI GPT-6 Luna, the cheapest OpenAI model. It is now the default AI model.

Added OpenAI GPT-6 Sol and GPT-5.6 Luna.

Added Google Gemini 3.1 Flash-Lite, 3.5 Flash-Lite, 3.8 Flash and 3.1 Pro (preview). Gemini 3.1 Flash-Lite is the new default for Gemini, and it can be used free of charge.

Removed Groq's Llama models. Groq now offers GPT-OSS 120B (the new default) and GPT-OSS 20B.

If your saved model is no longer available, ScribeForge switches to that provider's default model.

Cancel now stops a transcription within a second or so, instead of waiting for the whole file to finish.

Batch

You can now drag and drop audio and video files straight onto the Batch page.

Cancel All now also stops the file that is currently being downloaded or transcribed.

Clicking "View transcript" on a finished batch item now opens the transcript.

Loading a project that was saved while a batch was running now works.

Saving the same batch project again updates it instead of creating a duplicate.

Items added from a link now keep the video's title.

Pausing a batch now tells you it will pause after the current file finishes.

Batches run faster because the model stays loaded between files.

Live Transcription

Live transcription keeps up with your speech, even with larger models.

The language you choose on the Live page is now used.

The last few words you say before pressing Stop are now saved.

A "saving" message now shows while your live transcript is being saved.

More microphones now work with Live Transcription and Voice Recording.

If you leave the Live page while recording, it now shows the recording when you come back.

Live recordings now save their length, model and language in your library.

Parakeet models are no longer offered on the Live page, since they don't support live transcription yet.

Watch Folder

Watch Folder now keeps working when you switch to another page.

Watch Folder starts again automatically when you reopen ScribeForge, if you left it on.

Files dropped into the folder are now transcribed one at a time, so your computer stays responsive.

ScribeForge now waits until a file has finished copying before transcribing it.

The activity list now shows files waiting in line, files still copying, and progress for the current file.

Stopping Watch Folder lets the current file finish and skips the rest.

Changelog