Eranio
Color theme

Listening: Dictionary

Eranio Published on: 22. November 2026

Listening: Dictionary

This is where the bot learns how you speak. The page belongs to Listening and sits directly below it in the menu: Bot → Listening: Dictionary. It appears under the same conditions as Listening, so only after an admin has released it for you. Twitch only.


What the dictionary is for

Speech recognition does not understand every word correctly, for example "Tote increase" instead of "Tode increase". The dictionary remembers such mishearings for your channel so the command works next time anyway. Everything here applies to your own channel only, even if the bot serves several channels.

Layout of the page

Area What you see there
How the bot learns A short explanation of how learning and rating work
Hit rate How many of your rated commands were recognised correctly in the last 7 days
Recognition model Which model is active right now (you choose it on the Listening page)
Dictionary How many learned words are active, candidate or blocked
Reaction time How long it takes from speaking on stream to the reaction, average, median and maximum for the last 24 hours and 7 days
Own word list Names, game terms and inside jokes you give speech recognition as a hint
Recognised voice commands The latest executed voice commands to rate
Learned words The dictionary itself with counter, status and the option to block or delete

How learning works

  1. You say for example "Cactus increase", but speech recognition understands nothing that fits a command. The bot remembers the word briefly.
  2. If you or a mod triggers the matching command in chat within a minute, for example !tode, that counts as one hit for the pair "Cactus → tode".
  3. After three identical hits the word becomes active. Until then it is listed as a candidate.
  4. From then on "Cactus increase" is enough and the counter goes up.

Learned words only work directly next to "increase", "decrease" or "command". If you say "Cactus" in a normal sentence, nothing happens.

Rating

Every recognised voice command has three buttons:

  • Correct: The bot understood you correctly. This feeds into the hit rate.
  • Wrong: The command was triggered by mistake. The heard word is blocked for this command and no longer triggers it.
  • I meant ...: You pick the command you actually meant. The heard word then applies to it immediately, and is blocked for the wrong one.

Commands that arrived too late show up in the list as Dropped, too late. The bot no longer executes voice commands and calls when the audio is too old (more than 3 minutes on most channels), so nothing is ever triggered long after you spoke.

Managing learned words

In the Learned words list you can activate, block or delete every entry. A blocked pair is never picked up by automatic learning again.

Own word list

With Edit own words you enter names, game terms and inside jokes that speech recognition often gets wrong, for example a regular viewer or a game character. The words are passed to speech recognition as a hint, after the bot names and your commands. If you add too many, the last ones are cut off.

Recognition model, small or large

On the Listening page you choose with Choose recognition model between:

  • Small (fast): The default, copes with many channels at once.
  • Large (more accurate, German): Recognises German more accurately but needs more computing power.

If the large model cannot keep up, the service switches to the small model automatically for a while (60 minutes by default) and returns by itself afterwards. You see this at the top of the Listening page, for example "The large model could not keep up" with the time. The hit rate helps you decide which model suits you better.

Privacy and retention

  • Transcripts on the Listening page are deleted automatically after a few days as before, after 3 days by default.
  • The dictionary, the ratings and the hit rate stay, so you can see later what was learned and from what.
  • Only the recognised command, the heard words around the command, the time and your rating are stored, never whole transcripts. These entries are deleted after one year.
  • The reaction time measurements are plain numbers without text.

FAQ

Do I have to maintain the dictionary? No. It fills itself when you or your mods trigger commands in chat after speech recognition misheard something. Rating and blocking are optional.

Can a learned word trigger nonsense? Only directly next to "increase", "decrease" or "command" and only after three hits. If it still triggers something wrong, rate the command as Wrong and the pair is blocked.

Does my dictionary also work for other streamers? No. Dictionary, word list and model choice always apply to your own channel only.

Why was my command dropped? The audio was too far behind because speech recognition lagged. Simply say the command again.