Speech to Text — Free, Local, No Upload

Apps / Speech to Text

Drop a file here, or pick / record. Supported: MP3, WAV, M4A, WebM, OGG, FLAC, AAC, MP4 / MOV (audio track extracted).

Your audio never leaves the browser. The first run downloads a Whisper model (~75–150 MB); subsequent runs are cached.

About this speech-to-text tool

What is this?

A free, online speech-to-text transcriber that runs OpenAI's Whisper model in your browser. Record from your microphone or drop an audio file — you get a transcript with timestamps, downloadable as plain text, SRT subtitles, or VTT.

Privacy — your audio never leaves the page

Unlike "free" tools that upload your audio to a server, this app runs Whisper locally in your browser using WebAssembly. No upload, no signup, no daily quota, no watermark. Once the model has cached (after the first run), the entire tool works offline.

How does it work?

We load a Whisper variant (default: Xenova/whisper-base, ~150 MB) as ONNX. The model runs under the WebAssembly backend of 🤗 transformers.js. Your audio is decoded to 16 kHz mono PCM via the Web Audio API and fed directly to the model — there's no intermediate server.

Sources we accept

Live microphone — real-time streaming transcription via Web Audio + an AudioWorklet PCM tap. Sentences appear as you finish them.
Audio files — MP3, WAV, M4A, AAC, OGG, FLAC, Opus, WebM
Video files — MP4 / MOV / WebM (audio track is extracted)
Drag & drop — just drop a file onto the tool surface

How live transcription works

When you click Live transcribe, the browser opens your microphone, samples 16 kHz mono PCM via anAudioWorklet, and feeds it to a voice-activity detector. As you talk, the detector watches the energy envelope of each ~8 ms frame:

Pause > 800 ms after speech → flush the chunk to Whisper, append the result to the live transcript.
No pause for 8 s → force-flush to keep the transcript live during long, breathless monologues.

Whisper inferences run sequentially in the background while you continue speaking. The complete audio is captured into a WAV so you can play it back after stopping.

Model sizes & trade-offs

Whisper comes in multiple sizes. Bigger = more accurate but slower and a larger download:

tiny (~40–75 MB) — fastest. Good for clean audio.
base (~75–150 MB) — the sweet spot. Default.
small (~470 MB) — best practical accuracy in the browser.

Switch under Advanced options. Each model is downloaded once and cached.

Multi-language & translation

Whisper supports ~100 languages — pick yours from the language dropdown, or leave it on Auto-detect. The Translate task converts non-English audio directly to English text (useful for foreign-language interviews).

Output formats

Plain text (.txt) — for notes and search
Timestamped text — each line prefixed with [mm:ss]
.srt subtitles — drop into Premiere, DaVinci, Final Cut, YouTube
.vtt subtitles — drop into HTML <track> for web video

Tips for best accuracy

Record in a quiet space; background noise hurts every model.
Aim the microphone at the speaker. Built-in laptop mics are usable but headset mics are dramatically better.
Specify the language explicitly if you know it — faster and slightly more accurate than auto-detect.
For specialist vocabulary (medical, legal, technical), step up to small or larger — bigger models have more world knowledge baked in.

Common use cases

Podcasters — generate show notes and subtitles without uploading masters to a third-party service
Journalists — transcribe interviews on a train, on a plane, anywhere offline
Students — turn a lecture recording into searchable notes
Researchers — interview data that needs to stay on the device for ethics / IRB compliance
Content creators — instant SRT captions for YouTube, TikTok, Instagram, podcasts
Customer support — voice memos turned into tickets

FAQ

Is it really free?

Yes. No signup, no daily limit, no watermark, no API key. The cost of running this is your CPU and the one-time model download.

Does it work offline?

After the first transcription (when the model downloads and caches in IndexedDB), yes. Try it: load the tool, run an audio file, then disconnect your network.

Will my audio be uploaded?

No. Your microphone recording or uploaded file is decoded by the Web Audio API into PCM samples and handed straight to the Whisper model in your browser. The only network request after page load is the one-time model download from Hugging Face's CDN.

How long can the audio be?

Whisper handles arbitrarily long audio by chunking — the tool will happily transcribe an hour-long lecture. Each minute of audio takes a few seconds of CPU on a modern laptop with the default base model.

Why is the first run slow?

The model (~75–150 MB) downloads once on the first transcription and the WebAssembly runtime warms up. Every subsequent file uses the cached model and is much faster.

Can I transcribe video files?

Yes. The browser's audio decoder pulls the audio track out of MP4 / MOV / WebM containers. Pure-video formats without an audio track will produce an empty transcript.

Word-level timestamps?

The default mode produces segment-level timestamps. Whisper supports word-level timestamps via DTW post-processing — we'll wire this up as an option in a future update.

Related tools

Background Remover — also runs an AI model locally in the browser
Markdown Editor — clean up the transcript into shareable notes
Diff Viewer — compare two transcripts side-by-side