Skip to content
Hohe Labs

About

Voice AI for languages that general AI treats as an afterthought

Hohe Labs builds speech recognition, speech synthesis, voice cloning and subtitle generation for Amharic, with Afaan Oromoo in training. The models are open, small enough to run in a browser, and sold on prepaid credits.

ሆሄ/two vowel orders of one Ge'ez consonant

What we build

The speech layer, end to end

Listening, speaking, speaking as someone specific, and putting words on a screen. Each is its own product and they share one account and one balance.

Why it matters

A language needs infrastructure before it needs apps

Speech is how most people already use a phone. When recognition, synthesis and captioning work well in a language, everything built on top of them works in that language too: notes, search, accessibility, media. When they do not, speakers are asked to type in someone else's language, or to go without.

General-purpose AI is built first for the languages with the most data behind them, and the rest receive whatever is left over. Amharic and Afaan Oromoo need models trained for them, measured honestly, and priced and sold in a way their speakers can actually pay for.

That is the whole company: make the speech layer exist, make it open enough to build on, and make it reachable on the phones and payment methods people really have.

ha, seven orders
  1. 1ሀha
  2. 2ሁhu
  3. 3ሂhi
  4. 4ሃha
  5. 5ሄhemark
  6. 6ህh
  7. 7ሆhomark
One consonant and the seven vowel orders it is written in. The wordmark ሆሄ is the seventh order and the fifth of the same letter. A Ge'ez character carries a consonant and a vowel together, which is why an Amharic line of the same length holds more speech than a Latin one, and why the text measure and the caption timing on this site are set separately for Ethiopic.

How we work

Three positions we are willing to be held to

  • The weights are open

    Hohe ASR can run on your own machines, or in a browser tab with no GPU. That removes the main reason to hesitate before depending on a small company, and it means the hosted API has to earn its place on convenience rather than on lock-in.

  • The benchmark is public, and so is the weak spot

    We published an Amharic benchmark and scored six open models against it, our own included. The figures are for read speech. On real telephone audio the model is substantially worse, and we say so rather than leaving you to find out.

  • You pay in advance, in birr if you need to

    Prepaid credits, no subscription, nothing that renews. Pay by card or by bank transfer in birr, because a recurring card is a payment method many of the people who need this do not have.

6

open models scored

0

GPUs required

2

languages

Tell us what you are trying to build

If it involves Amharic or Afaan Oromoo speech, we will tell you honestly whether the models are good enough for it yet.

Email usRead the quickstart