Transcribe audio locally with Whisper in your browser

What you need before starting

Use a modern browser with WebAssembly support and enough available memory for your recording and model. Voice Cadence runs transcription in a browser worker; a dedicated GPU is not required. Longer recordings and larger models can take more time and memory.

Google sign-in unlocks transcription, voice-follow, and transcript editing in the studio. It does not upload or sync your recordings to Google. You can record, import audio, and edit by waveform without signing in or downloading a speech model.

Download and load the model

Expand On-device speech and choose a model. Whisper Tiny English is the default and downloads about 40 MB on first use. Choose Load speech model and watch the download and preparation status. You can also start from Transcribe locally in the audio editor.

Model files download directly from Hugging Face. Your audio is not sent there. When browser storage is available, model files are cached for reuse, although clearing browser data or cache eviction may require downloading them again.

Transcribe a recording

Record a take or import an audio file, open Edit audio, and choose Transcribe locally. Keep the tab open while it works. Voice Cadence shows progress and provisional text as processing advances, then finishes the word timing used by the editor.

The emerging text is a preview, not a guarantee that every word and boundary is final. Wait for completion before cutting audio by transcript. If you cancel, you can retry; initializing the model again may be necessary after some interruptions.

Understand what stays local

Your scripts, audio, and project edits live in this browser’s device storage. Transcription happens on the device. Downloading the model and signing in still involve internet requests, and the public site uses Google Analytics for site traffic.

Local processing does not mean the website is a fully installed offline app. Do not assume a future page load or a fresh sign-in will work without a connection. Download a project backup if you need a portable copy; browser storage and model caching are not permanent archives.

Choose accuracy and speed deliberately

Start with Tiny English for English speech and a short test recording. For another language, select an appropriate multilingual model offered in the app. Larger models require larger downloads and more device resources; there is no fixed completion time that applies to every browser and recording.

Clear speech and a quiet recording make review easier, but recognition can still miss names, accents, technical terms, or overlapping speakers. Check important wording against the recording. Word boundaries are approximate, so audition cuts before exporting.

If the download or transcription stalls

Check the model status first: downloading files and recognizing speech are separate stages. A blocked connection to Hugging Face can prevent model setup even while ordinary recording still works. Check your connection and retry loading the model.

If the browser becomes slow, close memory-heavy tabs and try a shorter recording or smaller model. Save a project backup before reloading during troubleshooting. For microphone problems, use Check microphone; speech-model settings cannot fix missing or silent audio.

How this differs from WhisperX

This app uses Whisper through Transformers.js in the browser; it does not run the Python WhisperX package. The transcript and word timing come from the browser transcription workflow. Speaker identification and guaranteed frame-accurate word alignment are not part of this feature.