The dialect that still spells English as sound
An Apache local speech plugin that keeps the Arabic dialect and still phoneticizes the English inside it.
- Category
- AI
- Who it is for
- Someone who sends Hermes voice notes in Levantine, Gulf, Egyptian, Iraqi, or Maghrebi Arabic.
- Possible model
- Open source
- Difficulty
- Hard
- Inspired by
- Lahja · Lahja
- Original launch
- Watch the launch · Nov 8, 2026
Difficulty is an editorial estimate of technical and operational complexity, not a score.
Startup idea analysis by Startup Ideas. Last reviewed Nov 8, 2026. How we read a launch.
The startup idea
Tarek posted Lahja on November 8. Hermes did not understand his Arabic, so he built it for a Nous Research plugin hackathon. It keeps the dialect, keeps English as English, and learns new words with /stt-fix. Fully local. Word errors from 41% to 29%. The repository is Tarekkharsa/hermes-lahja. Lahja is Apache 2.0. The dialect Whisper model is Apache 2.0 as well. A voice note goes through ffmpeg at 16 kHz, Silero VAD, and whisper.cpp on Metal. No API keys. It stores no audio and no transcripts. The known limits say English words inside Arabic sentences come out phonetically, and you fix the ones you use with /stt-fix. Numbers are written as words, not digits. The 29% is the remaining word error against stock Whisper, same settings, across 13 dialects. The post says English stays English. The readme says those words come out as sound until you correct them. This page did not send a voice note.
The opportunity is the dialect, with the English list you maintain. A smaller product is one local note, the 29% still printed, and /stt-fix for the words that came out wrong. A recognizer that already keeps every English word is not this readme.
Do not hide the 29%. Do not say English is already English.
The problem
A voice note in dialect comes back as formal Arabic, or as the wrong language.
Who would pay for this?
- Primary
- Someone who sends Hermes voice notes in Levantine, Gulf, Egyptian, Iraqi, or Maghrebi Arabic.
- Secondary
- A person who will add the English words they mix in, one /stt-fix at a time.
- Early adopter
- One local note, the phonetic-English line, no transcript stored.
How could it make money?
Possible models. Nothing here is a claim about what Lahja charges.
- No price. Apache 2.0.
- No API key.
- The model is a separate Apache download.
What would the MVP look like?
- A Hermes plugin.
- Local whisper.cpp.
- The 29% on the page.
- /stt-fix for the English that came out as sound.
How to validate the idea
- Ask five people whether the English word in the sentence came back in English.
- If they say yes, they read the post.
- Do not store the audio and keep the line.
How to differentiate
- The dialect is the point.
- English inside it is a list you fix.
- Nothing is stored.
- Do not become a translator.
Original product launch
This launch is inspiration and evidence that someone shipped, not a partnership. The idea above is our reading of the broader opportunity. Lahja did not write it.
- Product
- Lahja
- Source
- Lahja
- Film
- Lahja, English inside Arabic comes out as sound
- Posted
- Nov 8, 2026 · 1:28
Product that inspired this idea
Tarek's post says Hermes did not understand his Arabic, so he built Lahja for a Nous Research plugin hackathon. It keeps the dialect, keeps English as English, and learns new words with /stt-fix. Fully local. Word errors from 41% to 29%. The repository is Tarekkharsa/hermes-lahja. Lahja is Apache 2.0. The dialect model is Apache 2.0 as well. It runs whisper.cpp locally, with no API keys, and stores no audio and no transcripts. The known limits say English words inside Arabic sentences come out phonetically, and you fix the ones you use with /stt-fix. Numbers are written as words, not digits. The 29% is still a word error against stock Whisper on 13 dialects. The post says English stays English. The readme says those words come out as sound. This page did not send a voice note.
Posted by @tarekkh1997. Watch the original launch. No separate product site is listed unless the post itself is the source.
- speech
- arabic
Related startup ideas
- A coworker that stays on after the chat
An agent with its own computer that keeps a project page, and can be reached from the chat app a team already uses.
- A self-hosted assistant for personal apps
An assistant a person can host, pointed at the apps they already use, including computer use.
- A worker that lives in a folder
Give each worker a folder on the machine, a department, and a job, and say on the first screen whether it asks before it acts.
- Ask the computer, then approve the change
Put the agent on the desktop the person already uses, and stop it before it changes a file.
- Memory that survives the next coding session
Sell a memory layer that learns how one developer works, then carries that into the next agent run.
- One model talks, another moves
A character in the browser where the voice and the body are different models, and the film says which voice was scripted.
Looking for more ideas?
The full list is startup ideas for 2026, each one tied to a launch you can watch.
Browse startup ideas for 2026