Transcribe audio and video
Accurate transcripts made on your own computer. Nothing is uploaded.
Drop audio or video files here
MP3, WAV, M4A, MP4, MOV, WEBM and more. Add as many as you like: they run one after another.
Not local The link is looked up through the LocalScribe server, which passes the file to your browser without storing it. Transcription still happens on this device. How links work
More options
Queue
Record
Record your microphone, the sound of a browser tab or meeting, or both. The recording stays on this computer.
Compare recordings
Practice tool for interpreters: transcribe an original speech and your interpretation, then see them side by side with timing and simple statistics.
1. Original speech
2. Your interpretation
Both files use the quality chosen on the New transcription page. Tip: the interpretation should be recorded from the same starting point as the original, so the times line up.
Previous comparisons
Comparison
Glossary
Your terms help LocalScribe catch spelling mistakes in names, drugs, products and other key words. Everything is stored in this browser.
Terms to check
One term per line, written exactly as it should appear. LocalScribe flags words in a transcript that look like a misspelling of these terms.
Always replace
One rule per line: wrong => right. Rules are whole words and not case sensitive.
Saved automatically.
Settings
Appearance
Subtitles
For detailed subtitle work, open the SRT in Subtitle Refiner (separate app).
Cloud and AI features Not local
Keys are stored only in this browser on this computer. Anyone who uses this browser profile can use them, so do not save keys on a shared computer.
Storage
Calculating…
Deleting the models frees space but keeps your transcripts, glossary and settings. Models download again when needed.
About
LocalScribe v3.2 · noradiaz.com/apps
What's new in this version
Introduction
Help and frequently asked questions
Getting started
What does LocalScribe do?
It turns speech in audio and video files into editable text with timestamps. The speech model runs inside your browser, on your own computer, so your recordings are never uploaded. When you finish, you can export the text as a Word document, plain text, subtitles (SRT or VTT) or a spreadsheet (CSV).
How do I transcribe my first file?
- Choose the quality and the spoken language.
- Drop the file on the large box, or click Choose files.
- The first time you use a quality level, LocalScribe downloads the speech model. This happens only once.
- Follow the progress in the queue. Click View live to read lines as they appear, or Open when it is done.
Which quality should I choose?
- Fast (about 210 MB download): quick, good for clear audio with one speaker.
- Balanced (about 410 MB): the best everyday choice. Handles accents and some background noise well.
- Best (about 1.6 GB): the most accurate, especially for technical vocabulary and difficult audio. Needs a graphics card.
- Best, full precision: the same model as Best without compression. A little slower, with fewer added or missing words. If you already use Best, only about 340 MB more is downloaded. Needs a graphics card that supports half precision (most recent ones do).
Downloads are approximate and happen once per quality level. Files without a graphics card use smaller versions of the Fast and Balanced models.
Which spoken language should I choose?
Choosing the language yourself is the most reliable option. With Detect automatically, LocalScribe listens to the first minute and a half, decides the language and keeps it for the whole file. If a recording switches between languages (for example an interpreted interview in English and Spanish), choose Several languages: the language is checked again every 15 seconds, each change is labeled in the text (EN, ES…) and starts a new paragraph.
Limits: when people switch languages within the same sentence or every few seconds, part of a 15-second piece can come out in the wrong language. Choosing a single language works best for single-language recordings. With Original and English translation, parts already in English are not translated again.
Tip: if a transcript comes out in English when the person spoke another language, the language was not recognized. Choose it in the list and transcribe again.
Can I get the original and an English translation at the same time?
Yes. In Output, choose Original and English translation. The model runs twice, so it takes about twice as long. The editor then shows Original, Translation or Both side by side, and exports can include both. Translation into other languages is available with Translate in AI tools (a cloud feature).
The Best quality model was not trained to translate, so the English translation always uses the Balanced model.
Can I watch the video while I edit? Does it slow down the transcription?
Yes. Video files play in the panel next to the text, and the text follows playback. Showing the video has no effect on accuracy. You can open a file while it is still being transcribed (View live); playing it during that time uses a little of the graphics card, so the transcription may be slightly slower until you pause.
Does it work without internet?
Yes, after the first visit. Install LocalScribe as an app (the Install app button in the top bar, or the install icon in the address bar of Chrome or Edge) and use each quality level once while online. After that, LocalScribe opens and transcribes with no connection. Cloud features always need internet.
Is my audio really private?
Yes. With On this computer selected, the audio is processed in this browser tab and never leaves your device. The internet is used only to load the app and to download the speech models. Your transcripts are saved in this browser's storage on this computer. Only the optional cloud features send anything elsewhere, and each one is clearly marked with a Not local label.
Your computer and the graphics card
What does “Graphics card ready” mean?
Speech models do millions of small calculations for every second of audio. A graphics card (GPU) is designed to do many of these calculations at the same time, so it transcribes many times faster than the main processor (CPU).
“Graphics card ready” means your browser can use your computer's graphics card through a browser feature called WebGPU. LocalScribe will use it automatically. Click the indicator in the top bar to see which graphics card was found and which quality levels it can run.
What does “Processor mode” mean?
Your browser could not access the graphics card, so LocalScribe uses the main processor. It still works and the results are just as accurate, but it is slower, and the Best quality is not available. To fix it, use an up-to-date Google Chrome or Microsoft Edge, and make sure the computer's graphics drivers are up to date.
How long will a transcription take?
It depends on the computer and the quality level. With a recent graphics card, Fast and Balanced are usually several times faster than the length of the recording. In processor mode, expect it to take about as long as the recording, or longer. The queue shows an estimate after the first minute of audio.
Accuracy estimate
What is the “estimated accuracy” percentage?
For every word it writes, the speech model also reports how certain it is. LocalScribe averages those certainty scores for the whole transcript and shows the result, for example “≈ 94% estimated accuracy”.
It is an estimate, not a measurement: nobody has compared the transcript with what was actually said. In practice it tracks real accuracy reasonably well for clear recordings. It can be too optimistic with heavy background music, overlapping speakers or very poor audio, because the model can be confidently wrong.
- 95% and above: usually only small fixes.
- 85 to 95%: review the highlighted words.
- Below 85%: listen to the whole recording while you review, or try a higher quality level.
What are the highlighted words?
Words with an orange background are the ones the model was least certain about. Use Next unsure word in the editor to jump from one to the next: playback moves to that word so you can listen and fix it. When you edit a word, it counts as checked. You can turn highlighting off in Settings.
The editor
Keyboard shortcuts and editing
| Esc | Play or pause |
| Ctrl + ← / → | Back or forward 3 seconds |
| Enter | Split the line where the cursor is |
| Backspace at the start of a line | Join it with the line above |
| Click a time | Play from there |
| Click a word | Move playback to that word and start editing |
| Right-click a sentence | Line options in Running text |
Changes are saved automatically. Undo reverts the last bulk change (replace all, cleanup, glossary fixes, speaker detection, splits and joins).
What can I do with a right-click?
Right-click any sentence (or a word inside it) for these options:
- A word: Replace everywhere fixes the same word in the whole transcript and can add a replacement rule to your glossary. Add to glossary saves the word as a term.
- Listen: play from that word, loop the sentence until you click Stop, or play it at 0.75× speed.
- Transcribe this sentence again: runs only that passage through another quality level or language, then lets you keep the current version or use the new one. Useful for a passage the model got wrong.
- Edit: split the sentence where you clicked, start a new paragraph, change the speaker, insert [inaudible] or [crosstalk] with the time (select words first to replace them), join lines or delete one.
- Mark and share: add a note, highlight in four colors, copy the sentence (alone or with time and speaker) or translate it (a cloud feature). Notes and highlights in the toolbar lists them all.
Notes are exported as comments in Word, as a Note column in CSV and at the end of TXT files. Highlights keep their color in Word.
Running text or lines?
Running text shows the transcript as paragraphs, like a document, with the time of each sentence in grey, for example (0:07). Click a time to play from there. Turn timestamps off with the Timestamps box. Right-click a sentence for more options (join, delete, change speaker).
Lines shows one line per sentence with its time and speaker label, which is handy for subtitles and careful review.
Each transcript also shows when it was transcribed and how long it took.
Verbatim or clean?
Speech models tend to tidy up what people say: they often drop repetitions, false starts (“la app puedes abrirla… puedes abrir la app”) and hesitations. In Verbatim style, the default, LocalScribe shows the model an example of spontaneous speech with hesitations and repetitions, which makes it keep many more of them. This helps a lot, but it is not a guarantee: some repetitions can still be left out, so listen to key passages when you need a strict verbatim record.
Clean style lets the model tidy the text. You can also transcribe verbatim and use Clean up afterwards to remove fillers and repeated words, with Undo if you change your mind.
What does “Clean up” do?
It removes filler sounds (um, uh, eh, mmm and similar) and accidental repeated words (“the the”). Use it when you need a clean, readable transcript rather than a verbatim one. You can undo it.
How does the glossary help?
Add your terms on the Glossary page. In the editor, Check terms lists words that look like misspellings of your terms (for example “pembrolizumap” for “pembrolizumab”), with the place where each appears. Fix them one by one or all at once. Replacement rules (wrong => right) can also be applied automatically to every new transcript.
Speakers
How does automatic speaker detection work?
A second, small model listens for changes of voice and labels each part Speaker 1, Speaker 2 and so on. Rename the speakers in the bar above the transcript and the names update everywhere. It works best with two or three people who take clear turns. When people talk over each other, or when one recording is very long, check the labels: click any label to change it, with the option to change the lines that follow too. If one person was split into two labels, use the ⋯ button next to a speaker name to merge them, or It was only one speaker. When you know how many people speak, choose Only one speaker or Two speakers under More options before transcribing.
If you did not turn it on before transcribing, open the transcript, attach the original file if asked, and click Detect speakers.
Exports and backups
Which formats can I export?
- Word (DOCX): a formatted document in running text with timestamps like (0:07), one line per timestamp, or no timestamps, plus speaker names and the date it was transcribed. With a translation, a two-column bilingual table.
- TXT: plain text with the same timestamp choices.
- SRT and VTT: subtitles, split according to the rules in Settings.
- CSV: a table for Excel or a CAT tool, with start, end, speaker and text.
Can I search inside my transcripts?
Yes. The search box above Saved transcripts looks in the titles and in the full text of every transcript. Accents do not matter, so “medico” also finds “médico”. Each result shows the passage where the words appear and how many times. Click it to open the transcript at the first match, then use Next to go to the following ones.
Where are my transcripts saved? How do I move them to another computer?
They are saved in this browser, on this computer, together with a copy of the audio or video file (you can turn that off in Settings). Clearing the browser's site data deletes them. Use Back up at the bottom of the sidebar regularly: it downloads one file with all your transcripts. On another computer or browser, use Restore with that file. Audio and video files are not included in backups, so keep your originals.
Cloud features
Which features are not local?
Everything marked Not local:
- AI notes (summary, action items, chapters, minutes): sends the transcript text.
- Translate: sends the transcript text.
- Ask: sends the transcript text and your question.
- AI review of meaning in Compare recordings: sends both transcripts.
- Cloud transcription with OpenAI: sends the audio itself.
They need your own key from Anthropic, OpenAI or Google, added in Settings. Usage is billed to your account by that company. Before the first use, LocalScribe asks you to confirm.
When should I use cloud transcription?
Only when your computer is too slow for the local models and the content is not confidential. Your audio is sent to OpenAI. For confidential material (medical, legal, business), stay with On this computer.
Transcribing from a link
Which links work?
Paste a link in the box under the drop area, or drop a link from your browser onto it. These work:
- Dropbox, Google Drive, OneDrive, SharePoint and Box share links, as long as the file is shared as “Anyone with the link can view”.
- Podcasts: an episode page or the show's RSS feed (the newest episode is used). Many podcast sites work; if one does not, use the episode's download link.
- Direct links to audio or video files (ending in .mp3, .m4a, .mp4 and so on) and web pages with an audio or video player.
Files up to 2 GB.
Why not YouTube or TikTok?
YouTube, TikTok, Instagram, Vimeo, Spotify and similar platforms do not allow their videos to be downloaded by other tools, so LocalScribe does not fetch them. If the video is yours, download it from the platform (in YouTube Studio, open the video and choose Download) and drop the file here.
Is a link private?
Less private than a file on your computer, which is why it is marked Not local. A web page cannot download files from most other sites directly, so the link is sent to a small relay on the LocalScribe server, which looks up the file and passes it on to your browser. The file is not stored or kept there. When a site allows direct downloads, the file comes straight from that site instead. Either way, the transcription happens on your computer, as with any other file. For confidential recordings, download the file and drop it here.
Samples and credits
Where do the samples come from?
The samples are short recordings that are free to use. They are downloaded from their source when you choose one and are transcribed on your computer like any other file.
- English: John F. Kennedy, inaugural address, 1961. Public domain (U.S. government work).
- Spanish: Spoken Wikipedia, introduction to Isla de Margarita, by Wilfredor. CC0.
- French: speech by the mayor of Belfort, recorded by Thomas Bresson. CC BY 4.0.
- Italian: ESA astronaut Paolo Nespoli. ESA, CC BY 3.0 IGO.
- English and Spanish: Ronald Reagan, American Red Cross message recorded in English and Spanish, joined with a short pause. Public domain (U.S. government work).
Troubleshooting
The model download stopped or failed
Check your internet connection and try again: finished parts are kept, so it continues where it stopped. Large models need free disk space (about 2 GB for Best). Private or incognito windows may not allow large downloads.
“This file could not be read”
Some video formats or rare audio codecs cannot be opened by the browser. Convert the file to MP3 or MP4 (for example with a free converter or by exporting it again from your editing program) and try again.
The text repeats the same sentence or invents words during silence or music
Speech models sometimes do this with long silences, music or applause. LocalScribe removes the most common cases automatically. Delete any leftovers with the line menu (⋯), or try a higher quality level.
“The graphics card stopped responding”
The speech model ran out of graphics memory, or the graphics card was busy with something else. LocalScribe restarts the speech engine and tries again automatically. If it happens again, close other tabs or apps that use the graphics card (games, video editors, another LocalScribe tab, video calls) and click Try again, or click Try again with Balanced: Balanced needs much less graphics memory than Best.
The computer gets slow or the tab closes on very long files
Very long recordings (several hours) need a lot of memory. Split them into parts of one or two hours, close other heavy tabs, or use the Fast quality.
Something else
Reload the page: your saved transcripts stay. If the problem continues, contact me through noradiaz.com/apps.