Beta — Free to useMore features coming soon

Amharic Speech
to Text.

Dink is a fine-tuned speech recognition tool built specifically for Amharic — a language that standard AI models struggle to understand.

The Problem

Standard AI ignores low-resource languages

Amharic is spoken by over 50 million people, yet it remains an edge case for most speech AI. General-purpose models like Whisper — trained predominantly on high-resource languages — fail to recognize Amharic when used out of the box.

Whisper (base) on Amharic

Input: Amharic speech audio

"בליליום לראסו, אין גדינסטי " — garbled output

The model hallucinates, repeats filler words, or produces completely wrong text.

Dink (fine-tuned) on Amharic

Input: same audio

ሰላም፣ ዛሬ ምን እንሰራለን?

Accurate Amharic text — readable, usable, correct.

How It Works

Built through iteration

Dink is powered by Whisper-Small, fine-tuned specifically for Amharic through multiple training iterations — each one improving accuracy on this low-resource language.

~130h
Training Data
29%
Word Error Rate
v4
Model Version

Features

Simple by design

🎙

Record or Upload

Use your microphone directly in the browser or upload any audio file. Supports all common formats.

Fast Transcription

Get your Amharic transcription in seconds, powered by a dedicated inference endpoint.

📋

Copy & Use

One-click copy to grab your transcribed text and use it anywhere you need it.

Beta

This is just the beginning

Dink is currently in beta. The model is still being improved and new features are on the way — transcription history, better accuracy, and support for more dialects.

Try it now — it's free