How to use
- Choose File to open audio or video (MP3, WAV, M4A, OGG, FLAC, MP4 and more), or Microphone to record right on the page. For a video, only the sound is used.
- Pick the language. If you know it, choosing Korean or English directly is more reliable. Auto-detect decides once, from the first 30-second part that has sound, and sticks with that language to the end.
- Press Transcribe. The first time, you’ll see the model download (about 77 MB). After that, progress shows for each 30-second part along with a token count. Text fills in below as each part finishes, and you can cancel at any time.
- You can edit the text. Copy or Save it, save SRT or VTT subtitles, or use Open in subtitle editor to keep polishing.
Output formats
| Format | Content | Saved as |
|---|---|---|
| Text | One line per segment, no times | .txt |
| Timed text | Each segment’s start time up front, like [01:15] sentence (1:01:01 past the one-hour mark) |
.txt |
| SRT | Number, 00:00:03,200 --> 00:00:06,000, text |
.srt |
| VTT | WEBVTT header, times written with a dot |
.vtt |
For example, “Hello.” (0 to 2.5 s) and “Nice to meet you.” (3.2 to 6 s) come out as SRT like this:
1
00:00:00,000 --> 00:00:02,500
Hello.
2
00:00:03,200 --> 00:00:06,000
Nice to meet you.
Changing the output format only changes the text on screen. It doesn’t transcribe again. Switching formats throws away any edits you’ve made, so pick the format before you start editing.
How it works
The tool runs OpenAI’s Whisper base model inside your browser. Everything from preprocessing to decoding is built in, with no transcription library.
- Preprocessing: the audio is converted to 16 kHz mono and cut into 30-second windows (480,000 samples). A 400-sample Hann window, an exact 400-point DFT, 80 mel filters (Slaney scale), a log, a floor at the maximum minus 8 and (x+4)/4 scaling produce an 80 × 3000 feature matrix.
- Encoder and decoder: the encoder reads the features once, and the decoder produces tokens one at a time. A KV cache stores earlier work so each step only computes the new token, and the most likely token is picked each time (greedy decoding). Suppressed tokens (symbols and special tokens) are blocked, and a group that repeats for 12 tokens or more is treated as a looping hallucination and stops decoding.
- Timing: timestamp tokens count in 0.02-second steps. Subtract 50364 from the token id and multiply by 0.02 to get seconds, so 50364+150 is 3 s and 50364+1234 is 24.68 s. Each
start time token → text → end time tokengroup becomes one segment. - Text recovery: tokens are byte-level BPE. The Korean syllable “한” (UTF-8
ed 95 9c) is stored as the token stringíķľ, and one character can be split across several tokens. The tool maps token strings back to bytes, gathers them, and decodes the UTF-8 all at once so characters never come out garbled. - A 30-second part with almost no sound (RMS below 0.0005) is skipped. This keeps the model from inventing sentences out of silence.
The 30-second windows are fixed and don’t overlap. If a word falls right on a window edge, it may be dropped or split across two segments.
Speed and accuracy
These numbers come from running the real model in Node on an N100-class PC, transcribing an 11-second English speech. The encoder took about 4 to 5.5 seconds per 30-second window, however much of it was speech, and the decoder took about 1.8 to 2.3 seconds for 26 tokens (0.07 to 0.09 s per token). With around 100 tokens per window, that’s roughly 12 seconds per window, a little over 2 times real time. A one-hour file has 120 windows, so the estimate is about 24 minutes. Korean needs more tokens for the same amount of speech, so it can be slower, and browser speed depends on your device. The estimate shown on the page is based on these figures. Transcription handles up to 2 hours, and the file you open can be up to 500 MB and 1 hour.
Accuracy depends on model size. The base model does well with clearly spoken standard speech recorded in a quiet room, but background noise, accents and dialects, technical terms, names and overlapping voices cause plenty of mistakes. It doesn’t tell speakers apart, and numbers, spacing and punctuation can be off. For noisy recordings, cleaning them up first with Audio Noise Reduction sometimes helps. Always check anything important yourself.
The model and where it’s stored
On first use, the encoder (about 23 MB) and decoder (about 54 MB) are downloaded. Each file is verified by size and SHA-256 and saved in this browser’s cache, so later runs start without a download. Delete saved models on the page removes them at any time. The models come only from this site, with no outside server or CDN. The weights are OpenAI’s released Whisper model (Apache-2.0).
To burn the subtitles into a video, fine-tune the timing in the Subtitle Editor, then use Burn Subtitles into Video. To just switch subtitle formats, use the Subtitle Converter.
Your audio is processed only in your browser and never sent to a server.