Dictate live in 70+ languages, load an audio or video file and sync it with a searchable interactive transcript, then export clean text or subtitles. Everything runs inside your browser tab with the native Web Speech and Web Audio APIs — no account, no AI service, nothing uploaded.
Live microphone, an existing media file, or captions you already have
MP3 · WAV · M4A · OGG · FLAC · MP4 · WEBM · MOV — decoded on your device, never uploaded
Search, edit, label speakers — click any line to jump the player to that moment
Counts, speaking pace and the words you lean on most
Every file is generated locally in this tab
Autosave keeps your work in this browser only (localStorage) so a refresh never loses the transcript. Hit Clear to wipe it completely.
Faster than reaching for the mouse
Your file never leaves your device: it is decoded with the Web Audio API and played from a local blob URL. Live dictation uses the browser speech recognition engine built into Chrome, Edge and Safari, which may send microphone audio to the browser vendor for recognition — that is a browser feature, not something this page controls, and no transcript, file or recording is ever sent to apps2help.com.
Over seventy locales are listed, from English, Spanish, French, German and Portuguese through Hindi, Bengali, Tamil, Arabic, Mandarin, Japanese and Korean. Punctuation commands, filler-word detection and the stop-word lists used for keyword counts ship for the major language families, and the rest fall back to the English rules.
No browser exposes an offline speech engine to web pages, and the Web Speech API only accepts a live microphone stream. The honest workaround here is to play your file out loud and let the engine listen in, which works well in a quiet room with speakers. For files you can also detect speech regions, mark lines at the playhead, or import captions you already have.
Yes. Turn on edit mode and every line becomes editable in place, with search and replace for repeated errors, one-click speaker labels, merge-into-paragraphs, star markers for the parts worth keeping, and a single-level undo for deleted lines.
Plain text, timestamped text, SRT and WebVTT subtitles, structured JSON, Markdown, or a print-to-PDF view. If you recorded live audio, you can also save that recording as a WebM file.
Yes, like every tool on this site. There is no account, no watermark, no usage limit and no server-side processing.
The Web Speech API exposes your operating system's or browser's speech engine to a web page. Audio from the microphone is turned into a stream of hypotheses, each with a confidence score. The engine emits interim results almost immediately and then revises them into final results once it has enough surrounding context to settle on an interpretation.
This is why live transcription visibly rewrites itself as you talk. The engine hears "recognise speech", initially guesses "wreck a nice beach", and then the language model, seeing the rest of the sentence, corrects it. Watching those revisions happen is disconcerting the first time, but it is the system working as designed rather than misbehaving.
| Factor | Helps | Hurts |
|---|---|---|
| Microphone distance | 15–30 cm, or a headset | Across the room, laptop built-in |
| Background noise | Quiet room, soft furnishings | Fans, traffic, café chatter, music |
| Speaking pace | Steady, natural, complete sentences | Rushed, trailing off, long pauses mid-clause |
| Vocabulary | Everyday words | Jargon, product names, acronyms, proper nouns |
| Speakers | One at a time | Overlapping voices |
| Language setting | Matches the accent variant | Wrong regional variant selected |
Selecting the right regional variant matters more than most people expect. English recognition models are trained separately for different regions, and choosing the variant closest to the speaker's accent typically produces a noticeable jump in accuracy with no other change.
Browser engines vary in how they handle punctuation. Some insert commas and full stops automatically from prosody and grammar; others insert almost none, leaving you a continuous stream of words. Where automatic punctuation is weak, speaking the marks explicitly works — saying "comma", "full stop", "new paragraph" — though you must then remember to do it consistently or the output becomes a mixture of both conventions, which is more work to clean up than either alone.
Numbers are similarly inconsistent. "Twenty twenty five" may become the digits or the words depending on context. Currency, dates and phone numbers are worth checking every time.
Overlapping speech is the hardest case. Browser engines do not perform speaker diarisation, so a two-person conversation arrives as one undifferentiated stream with no indication of who said what. Meeting recordings where people talk over each other produce output that is genuinely difficult to reconstruct.
Recognition also degrades sharply with strong accents outside the trained variants, with speech affected by illness or fatigue, and with children's voices, which sit outside most training distributions. Heavy technical jargon and organisation-specific terminology will be wrong nearly every time because those words simply do not appear in the language model.
Finally, recognition needs live microphone audio. It is not a file transcriber: you cannot feed it an existing recording and get text out. Playing a recording into the microphone technically works but adds a full generation of acoustic degradation and produces markedly worse results than the original audio would.
Dictating first drafts of emails and documents, capturing your own thoughts while your hands are busy, taking notes in a one-to-one conversation, drafting meeting minutes as a solo note-taker, accessibility support for anyone who finds typing painful or slow, and producing a rough transcript of a lecture or interview that you will edit afterwards. In each of those the output is a working draft that a human refines, which is the realistic sweet spot for the technology.
This is worth being precise about rather than making a blanket claim. The page itself stores nothing and transmits nothing: your transcript lives in the browser tab and is yours to copy or download. However, the recognition itself is performed by your browser's speech engine, and in some browsers that engine sends audio to the vendor's servers for processing. That behaviour belongs to the browser, not to this page, and it varies by browser and platform.
If that matters for your material, check your browser's own speech and privacy settings, or use a browser and platform that performs recognition on-device. For genuinely confidential conversations, the safest position is not to rely on any live recognition service without first confirming how your specific browser handles the audio.
The engine emits provisional guesses immediately and revises them once it has heard enough of the sentence to disambiguate. Those revisions are the language model correcting itself, and the final result is more accurate than the interim one.
No. This works from live microphone input. Playing a recording into your microphone will produce something, but the extra generation of acoustic loss makes the result considerably worse.
Some browser engines insert it automatically from grammar and intonation. Where they do not, speak the marks aloud — saying comma, full stop or new paragraph — and be consistent about it throughout the session.
Those words are rare or absent in the language model, so the engine substitutes the nearest common word. Check how each one was rendered, then fix all occurrences at the end with find-and-replace.
No. Browser speech recognition does not perform speaker separation, so a conversation arrives as a single stream with no labels. Overlapping speech is particularly difficult.
A wired headset gives the biggest improvement for the least effort. A built-in laptop microphone picks up keyboard noise and room reflections, both of which reduce accuracy noticeably.
The page stores nothing and sends nothing; the transcript stays in your browser tab. Note separately that the recognition engine belongs to your browser, and some browsers process speech on their own servers, so check your browser's speech settings if the content is sensitive.
Support comes from the browser's own speech engine, so availability and quality differ between Chrome, Edge, Safari and Firefox, and between desktop and mobile versions of each.