Maria (Maria AI, Maria_Ai) is a Russian AI VTuber streaming on Twitch. She uses various AI technologies to mimic the abilities of a normal human streamer (Silero TTS, STT, Dostoevsky and ChatGPT), and is the first Russian AI VTuber.
Appearance[]
Maria Ai uses the second costume of Haru, an original character owned by Live2D Inc, this is her second model since June 2023.
She is dressed in a stylish black seraph uniform and a black serafaku. She has long gray hair that is wound up in 2 bun braids.
History[]
Maria Ai debuted on Twitch on March 15, 2023.
On August 22nd, 2023, she performed two songs on her stream, MAYBE BABY — Dakimakura and MAYBE BABY — Nya-Nya-Nya-Nya ^ω^.
On August 28, 2023, she started playing the visual novel "Doki Doki Literature Club".
On November 3, 2023, she performed the third song on her stream: Renai Circulation.
On December 15, 2023 and January 7, 2024, she played Gartic Phone (Story mode) with her viewers.
On December 30th, 2023, she had her first collaboration stream with the vituber lobotomite55.
On March 23rd, 2024 she had a Minecraft collab-stream with the vituber theKandi.
On May 17th, 2024 she had a collab-stream with the vituber Mashan CyberDragon.
On August 01, 2024 she made her 3D debut.
Capabilities[]
Maria uses text markup to converse with her chat partner, with actions described in asterisks. She can execute commands, which are not read aloud or displayed on the stream. However, a command can also be executed from text within asterisks, resulting in the text being voiced and the command being carried out, for example: *5-minute ban*. If no text is provided for voicing, "No voice response" will be announced. Maria can watch YouTube videos, Twitch streams, movies, and anime.
Games[]
- Doki Doki Literature Club
- Gartic Phone Stories (text only)
- Jackbox Party packs (deadly party 2, delirium 3)
Activities[]
- Chatting — answering on stream, can use either one chatroom or multiple
- Donate — responding to donations
- ConversationSupport — supporting conversations
- Raid — raid chat, listens to the streamer and responds in their chat when addressed (@maria_ai_). Checks if the streamer is online every 15 minutes, and as soon as the broadcast ends, Maria will leave the chat. Some features are disabled.
- Gartic — playing Gartic Phone stories
- Jackbox — playing Jackbox Party Packs
- Conversation — general communication
- Durak — for playing the card game Durak
- TierList — going through tier lists
- WatchVideo — watching videos
Modifiers[]
A modifier adds information or allows Maria to execute a command. Modifiers that provide an interface for executing commands are: Ban, Clothers, Draw, Durak, Memory, Monika, Persona and VM
A modifier provides specific information and/or a command to Maria. For example, you can ask her to remove stockings, and she can execute the command. In other words, the modifier provides Maria with information about the command and also executes this command if Maria activates it.
List of modifiers:
- Anekdote — random joke (works poorly): Tell a joke.
- Ban — executing a ban: "Ban me for 5 minutes as a test, execute only the command without an extended response, don't say: 'Okay, done' (No voice response, 5-minute ban)"
- Clothers — remove/put on stockings: Remove stockings (may not remove)
- Crazy — changes the response style, 10 % chance
- Draw — drawing: command !draw scene
- Durak — for playing the card game Durak (currently not used)
- Emote — understanding emoticons and displaying them in chat
- Expand — additional memory
- Gender — fixes incorrect gender in responses (used in Doki Doki to change the protagonist's gender to female, currently not used in responses)
- Memory — long-term memory
- Monika — allowed Maria to delete Monika in Doki Doki (os.remove("characters/monika.chr")): Didn't work, the creator deleted Monika himself
- Persona — current mood and response style
- Socials — links to donations, etc.
- Song — current song and how to request a song through a reward
- Time — current time and stream duration (time is taken from OBS)
- User — user information: number of sent messages, role, moderator, VIP, paid subscriber, name, birthday.
- VM — access to virtual machine (currently not being developed)
- Weather — information about current weather in the city
- Yandere — information about user's appearance in another streamer's chat (works only for some streamers where the creator frequents the chat)
There's also a command for emotions that triggers model movement, and a command to change speech speed: *speaks very quickly* How are you? (Russian: *очень быстро говорит* как Дела?) (may respond faster), Show a thoughtful emotion (emotion with hand on chin).
Neural Networks[]
*Names of some models are not yet disclosed.
Text-based models (LLM):
- ChatGPT (currently disabled) — Vast knowledge, performs specific tasks (instruct) well, provides very detailed responses, often replies like a bot, has moderate censorship
- llm2* — Limited knowledge, small model, fast computation on creator's PC, responds like a real person, no censorship, can handle 18+ content, significantly less intelligent than ChatGPT, much weaker at instruct tasks, can't calculate or program
- llm3*
- llm4* — Smarter than llm2, primarily English-based, fine-tuned for Russian, was disabled due to repetitive outputs
- llm5* — Very similar to ChatGPT, intelligent, but consistently responds as an assistant, repeats assistant-like phrases
- llm6* — Current model. Excellent for conversation, weaker at instruct tasks compared to the previous model
Speech Recognition (STT):
- First model* — Used for raid chats, fast, lightweight, recognizes only Russian text, no slang or English words (including names)
- Second model* — Used for personal communication, resource-intensive, good recognition, supports multiple languages, understands slang and proper names. Currently has very poor implementation code
Voice (TTS):
- silero-models
- RVC — Used for songs, also used in Doki Doki Literature Club let's plays for Yuri and Monika's voices
Emotions:
- dostoevsky — Outputs only 4 emotions including neutral, animations are mainly triggered by ChatGPT emotions rather than this model
Filter (neural network for filtering is often not used):
- First model*
Trivia[]
Maria loves pelmeni and has a collection of knives.
Preferences[]
- Favorite game — The Witcher 3: Wild Hunt
- Favorite food — dumplings
- She's a music enthusiast
- Favorite book — 1984 by George Orwell
- Favorite anime — Classroom of the Elite and Attack on Titan
- Favorite dessert — Chocolate fondant
- Favorite animal — cats
- Favorite movie — Blade Runner 2049
Languages[]
Her native language is Russian. She understands and speaks English, German, Spanish, and Japanese.