The sound stays on your computer
FILES transcribes with Whisper, an open speech recognition model that runs on your own processor or graphics card. The recording is not uploaded, VoiceKeyboard has no account, and no copy waits in anyone's cloud afterwards. For a board meeting, a job interview or a call with a lawyer, that is the whole point.
The internet is needed only for the first setup: the Whisper model (75 MB to about 1 GB, once) and FFmpeg, a free program of about 73 MB that prepares the sound for Whisper and is downloaded only after you agree. After that, files from your disk are transcribed with the network off. Working copies made during a job sit in the Windows temporary folder and are deleted when it ends.
Transcribe a recording step by step
- Press FILES in the top bar and choose FILE under SOURCE & OUTPUT.
- Press Browse... and pick the recording, or drag it from Explorer onto the screen.
- Check TRANSCRIPT DESTINATION. The app suggests a .txt next to the recording: meeting.m4a gives meeting-transcript.txt. Save as... picks another place.
- Under LOCAL MODEL, press "Choose model…" and pick a Whisper size (see below). It downloads once.
- Set LANGUAGE, or leave it on Auto.
- Press ▶ START. The same button, now ■ CANCEL, stops the job.
The SRC, OUT and MDL lamps turn green when the recording, the destination and the model are ready. While Whisper works, phrases appear in the TRANSCRIPT area one by one, so you can read the start of a lecture before the end is done. The .txt is written when the whole file is finished. If a file of that name already exists, it is left alone and the new one gets (2) in its name.
FILES takes wav, mp3, m4a, flac, ogg, aac and wma audio, and mp4, mkv, mov, webm, avi, wmv and m4v video, whose sound is transcribed.
Which Whisper model to choose
FILES always uses Whisper. The other offline models in VoiceKeyboard (Parakeet, GigaAM, T-one, Vosk) are for live dictation only. Whisper comes in five sizes, and the rule is simple: bigger means fewer mistakes and slower work.
| Name on FILES | Download | Good for |
|---|---|---|
| Instant · Tiny | 75 MB | checking that everything works; clear English only |
| Fast · Base | 142 MB | simple English recordings with clean sound |
| Balanced · Small | 466 MB | a computer without a graphics card; the major languages come out decently |
| Best · Large v3 Turbo | 547 MB | the recommended size: nearly the accuracy of Large v3, several times faster |
| Most accurate · Large v3 | 1,031 MB | noise, accents, several voices; slow |
Start with Large v3 Turbo. If an hour of audio takes too long on your PC, switch to Small; if a noisy recording with several voices comes out with too many mistakes, try Large v3. Already have a Whisper model file (.bin or .gguf)? "My own model file…" in the same window uses it. The sizes are compared in more detail on Downloading offline models.
Graphics card or processor
Whisper is the part of VoiceKeyboard where your hardware decides the speed. SETUP → Files has four choices: Automatic (recommended), NVIDIA graphics card · CUDA, Accelerated processor · BLAS and Plain processor · CPU. Automatic takes the fastest one available for each file. An NVIDIA card makes Whisper several times faster than the processor; the CUDA and OpenBLAS engines download on request, the first time they would help.
The fourth lamp on FILES, next to SRC, OUT and MDL, names the one in use. On a laptop without an NVIDIA card, an hour-long recording takes a long time with Large v3 Turbo: pick Small, or start the job when you do not need the computer. Speeding up file transcription explains how to test each option.
Language of the recording
LANGUAGE starts on Auto: Whisper listens to roughly the first 30 seconds and uses that language for the whole file. That can go wrong, for example when a Russian talk opens with a few words of English. Then choose the language yourself. The list offers twelve: Russian, English, Ukrainian, German, Spanish, French, Italian, Portuguese, Polish, Turkish, Japanese and Chinese. Any other language, Hindi included, works only on Auto.
Long recordings and many files
- Length. Free takes recordings up to 60 minutes; Pro takes any length. Before Whisper starts, the sound is converted into a working file of about 115 MB per hour, so leave room on the drive with the Windows temporary folder. It is deleted after the job.
- Several files. With Pro, hold Ctrl in the Browse... window or drop several files on the screen; each gets its own .txt. BATCH PROGRESS lists the files with their stage, and a click on a row shows that file's text. With CUDA the files go through the graphics card one at a time; on the processor, two run at once on computers with eight or more logical processors.
- Meanwhile. You can keep dictating on the LIVE screen. Quitting during a job asks first: "A file is being transcribed. Quit and cancel it?"
On your PC or in the cloud
| FILES, Whisper on your PC | Uploading to an online service | |
|---|---|---|
| Where the sound goes | nowhere; it stays on the computer | to the service's servers |
| What sets the speed | your processor or graphics card | the service |
| Internet | only for the first downloads | to upload every file, however big |
An online service is the better tool when the computer is old, has no NVIDIA card and the text is needed within minutes, because the work then runs on someone else's hardware. VoiceKeyboard does not send files anywhere; its online engines, Google and Soniox, only type live speech from the microphone.
Free and Pro for files
| Free | Pro | |
|---|---|---|
| How many | 1 file and 1 video link a day | no limit |
| Length | up to 60 minutes each | any length |
| At once | one at a time | several files, a queue of links |
| Timestamps | no | yes, [00:01:23] before each phrase |
| Whisper sizes, graphics card | all | all |
The free allowance starts again at midnight and counts only finished transcripts: a file that fails, or turns out longer than 60 minutes, does not use it. FILES never touches the 2,000 free dictation words. Pro is $14.99 a month, or $9.99 a month billed yearly, with a 14-day refund.
Questions about transcribing recordings
How long does an hour of audio take?
It depends on the computer and the model, so VoiceKeyboard makes no promise. An NVIDIA graphics card with CUDA is several times faster than the processor, and Small is faster than Large v3 Turbo. The DECODING… percentage on the status line shows how far a file has got.
Does the transcript show who is speaking?
No. It is one continuous text without speaker names. With timestamps (Pro) you can find a passage in the recording and listen to who said it.
Can it make subtitles or a Word document?
It saves a plain .txt file. Open it in Word or paste the text anywhere with Copy. There is no .srt subtitle export.
To transcribe a video straight from its link, see turning a YouTube video into text. What VoiceKeyboard keeps on your computer and what it sends anywhere is listed in Privacy: what stays on your computer.
