There's one ROM hack that is getting a little out of control since its announcement a few weeks back, and that ROM hack is called Pokémon Emerald Arena, a fan-made, AI-assisted hack that has brought a real-time battle arena to the game, allowing you to battle Pokémon and trainers in real time using different moves and even the ability to use the arena and its objects to your advantage.
This original game has been developed by germanburgardt on X, and to say the least, it's gaining a lot of popularity due to its new battle mechanics and unedited Emerald storyline/experience. However, that hasn't stopped others from taking this open-source code and altering it even further, with one crazy Pokemon Master adding a new mode to the real-time battles known as "Arena Mode" that allows you to use your internal microphone on your Mac to use your voice as a command to use a certain Pokémon move, as if you were a real Pokemon trainer in battle.
From what the promotional material shows, you get put into a battle with another opponent, and both of you have 100 credits to choose certain Pokémon; you can't go over that limit, and stronger Pokémon cost more, meaning you have to be strategic in your choices and spend every credit wisely. For example, Mewtwo costs 32 credits, whereas Magikarp costs only 2.
Once chosen, you then get thrown into battle in the classic Emerald Arena mode that many of you are already familiar with if you've tried the ROM hack before. The game will show you your moves on-screen, and they can be controlled using your voice if you have enough PP. The microphone is paired with WhisperAI, an automatic speech recognition (ASR) system developed by OpenAI that converts spoken audio into written text that can be heard at 120ms, making it pretty quick at picking up what you said, or so the promotional video says.
Arena Mode Features/Specs
- Main menu: the title screen and main menu stay as in the original game, with ARENA added as the last item, after NEW GAME / CONTINUE and OPTION. It opens 1P VS COM, 1P VS 1P and BACK. In 1P VS COM, left and right toggle the COM between EASY and HARD.
- Draft: each player spends 100 credits on up to 6 of the arena's 150 Pokémon. All are level 100 with fixed movesets.
- Prices: costs rise with base stats, from 2 for Magikarp to 32 for Mewtwo, so the strongest six don't fit the budget.
- Grid: 24×24 portraits, five per row, scrolling. A READY cell leads to DONE.
- Details: the left pane shows the dex number, type, stat bars, total and moves.
- Budget: Pokémon out of budget are greyed out.
- Redraft: REDRAFT steps back one screen and keeps your picks.
- Battle: straight into the real-time arena, with no classic intro, messages or animations. The only screen outside the arena is the party list for switching.
- After a KO: each player picks their next Pokémon. Player 2 picks in a list inside the arena.
- Switching: SELECT, H for Player 2, or voice opens the party list.
- HUD:
- Each side's moves and PP sit under its HP bar.
- Spoken commands appear as speech bubbles, wrapped to two lines when long.
- Player 2's panel and bubble show only when Player 2 is human.
- Esc: opens RESUME / START OVER on the draft, versus and battle screens.
- Screens: new versus and result screens use a large pixel font, team colours and confetti.
- 1P VS 1P: Player 2 plays on the same keyboard with I/J/K/L, U to attack, O to dodge, Y/P to change moves and H to switch. Player 2 can also use voice.
- Listening: push-to-talk on Right Option, or
--hands-freewith WebRTC voice detection. - Transcription: Whisper small.en runs on the Mac.
- Intent: TypeSafe Jev picks the command. Clear commands like a move name, a direction or "switch to Gengar" skip Jev.
- Movement: walks, approaches, retreats and attacks follow paths around rocks, using the game's own A*. A close-range move runs up to the opponent first.
- Commands: chains like "retreat and then shadow ball" work. So do "switch pokemon", "switch to Dragonite" and drafting by voice ("pick Charizard and Gengar", "done", "fight").
- Two players on one mic:
- Setup: on the versus screen, each player reads "Quick Pikachu, fire a Thunderbolt!" to set their voice.
- Routing: commands are routed with Resemblyzer speaker embeddings.
- Clips with both voices: a clip holding both voices is cut where the voice changes, and each piece goes to its player.
- Ownership: a command naming a move only one player has goes to that player.
- Measured on the author's Mac: from the agent log, Whisper takes a median 120 ms per command (261 commands) and Jev a median 215 ms (101 calls).
And yes, for those asking on socials, this is completely AI-made, co-developed by Didier Lopes, so tread carefully. It is available today on GitHub to download, and again, just out of caution, please do your own due diligence when downloading files from GitHub!
