Earshot: an offline talking reader and assistant

Player

Nothing loaded yet. Paste text, open a file, or fetch a page below. You can also load the welcome sample.

Pause holds your place. Stop goes back to the start.

Speed

Who reads

Earshot's neural voice (Kokoro). Load it under Load voices first.

Your device's voice (Web Speech)

VoiceOver's own voice (Earshot moves focus, your screen reader reads)

Neural voice

Device voice

Now reading

Nothing yet.

Text to read

Paste or type

Text

Open a file

Choose a text, Markdown or web page file (.txt, .md, .html)

Read a web page

Web address

This only works when the site allows other pages to read it (called CORS). Wikipedia articles and GitHub project pages work. For other sites, copy the page text and paste it above.

The text

No text loaded.

Talk to Earshot

Press Talk and speak. Press it again when you're done, or just stop talking. Try “read”, “pause”, “faster”, “what is this about?”, or ask a question about the text.

Listening uses your device's speech recognition, which may send audio to Apple or Google. Load Whisper under Load voices to keep it on your device.

Conversation

Ask about the text

How answers work right now: Earshot finds the passage that best matches your question and reads it to you. It's a keyword search on your device, so it never makes things up, but it can't explain in its own words. Turn on smarter matching or answers in its own words under Load voices.

Answer

Ask something once a text is loaded.

Other matching passages

Load voices

These models run on your device. The first time, your browser downloads them from Hugging Face. After that they're saved in your browser and work offline. Nothing you read or say is uploaded.

Choose what to download

Neural voice: Kokoro 82M, 28 English voices (about 93 MB)

Speech recognition: Whisper

Whisper model

Smarter question matching by meaning: MiniLM (about 23 MB)

Answers in its own words: Granite 4.0 350M (about 350 MB, needs WebGPU)

Download size: about 150 MB. Use Wi-Fi. Phones may need a minute to get ready.

Settings

Keyboard shortcuts

Status updates for screen readers (short, polite, and held back while Earshot is talking)

Stop listening when I stop talking

Play a short tone when listening starts and stops

Highlight and scroll to the paragraph being read

Keyboard shortcuts

Shortcuts work when you're not typing in a text box. Escape works everywhere.

K: Read or pause
S: Stop
Escape: Stop talking or listening
J and L: Back or forward one sentence
P and N: Previous or next paragraph
- and =: Slower or faster
T: Talk (start or stop listening)
A: Jump to the question box
W: Where am I?
?: Jump to this list

About Earshot

Earshot is a mashup of open-source projects from my GitHub stars. Neural speech is Kokoro 82M. Listening is OpenAI's Whisper or Distil-Whisper. Both run in your browser through Transformers.js and ONNX Runtime Web. Kokoro's pronunciation comes from eSpeak NG.

Honest limits: the neural models are big for phones. On an iPhone, the first sentence can take a few seconds to start. Your device's voice and VoiceOver mode work instantly with no download. Web pages only load when the site allows it.

Starred credits and licenses