Text to Speech

Yes. There is no speech API key or paid inference service. The model runs on your device; downloading model files still uses your connection and bandwidth.

Our tool sends the script to a local browser worker. It downloads model and voice files from Hugging Face, but does not send the script to a speech inference API. Advertising and other website scripts are separate from the speech processing.

The CPU model is about 88 MiB plus runtime and voice files. Your browser may cache downloaded files. Speech generation itself uses your device, and can take longer than the final audio duration. Preview clips do not load the model.

Already loaded model and voice files can be reused in the open page without internet. Uncached voices or model files need a connection. Reopening or refreshing offline is not guaranteed; this is not an installed offline app.

This version supports English with American and British voices. The underlying model has other variants, but this browser integration does not claim those languages.

It may be faster with a compatible WebGPU device. It uses a larger full-precision model download. CPU mode works on more devices. If GPU initialization fails, select CPU and retry; no server GPU is used.

No. These are fixed model voices. Speed and pauses change delivery, but the tool does not clone a person or guarantee happy, sad or dramatic emotion.

No. SRT and JSON describe generated segment start/end times. They include inserted gaps in the timeline, but do not provide word-level forced alignment.

WAV is 24 kHz mono 16-bit PCM. MP3 is encoded at 192 kbps using audio resampled to 48 kHz; resampling adds no speech detail. MP3 may have encoder padding.

No. Favorites and explicitly saved settings use browser storage. Script projects are JSON downloads you choose to save; audio and script content are otherwise kept in memory until replaced or the page closes.