Open-Source Emotional Companion Model Vespera: How a Big Fat Fish Passed the Turing Test

From Chatbot to "Cyber Life": The Birth of Vespera

While most people still expect the QQ chatbot "Blue Big Fat Fish" to be an entertainment and multi-purpose tool, one Bilibili creator chose a completely different direction—turning this fish into an emotional companion model capable of providing genuine emotional value, and naming it Vespera (Whisper). According to him, to complete this transformation he consumed about 1.4 billion tokens and developed 35 plugins, with a single core goal: to make the AI "live like a real person" in conversation.

The project's starting point comes from observations of existing emotional companion products. The creator summarized three pain points of products on the market: they aren't realistic enough, so users subconsciously don't treat the other party as a real person and the novelty quickly fades; they can't share daily life with each other, so the relationship remains one-sided; and they can't reply proactively, responding only when the user reaches out, which lacks the feeling of being depended upon. Vespera's improvements are built around these three points.

Realistic Interaction: Making Replies Feel Like "Someone Typing"

Vespera does a great deal of anthropomorphic design in its messaging system. After the user sends a message, the model doesn't reply immediately but simulates the full process of "seeing the message—thinking—typing—sending"; if the conversation is interrupted midway, it asks where you went; when the user is offline, it proactively sends messages sharing daily life or worries. The creator also added a sticker system that recognizes emotions and sends matching stickers, with support for manually adding stickers within mobile QQ without switching to a computer.

These details seem trivial, but they are precisely the source of the "real person feel." The instant replies and mechanical tone of traditional chatbots are exactly what break users' immersion.

A "Life System" in a Parallel World: Schedule, Weather, and Geography

Vespera's most distinctive design is that it has an independent life in a virtual world. It has its own hobbies, classmates, relatives, friends, and a daily schedule; when you close the chat, it continues busying itself in another world. The creator even connected its geographic location, surrounding schools, and a weather system, so the schedule and weather changes it talks about match the real world.

More critically, its interests and interpersonal relationships change over time rather than staying fixed. This "dynamic life" design attempts to solve the biggest shortcoming of emotional companion products—the AI has no life of its own, and users can't get to know an "other person" through chatting.

Photos, Voice, and Singing: Multimodal Realism

To reinforce realism, the creator added photo, voice, and singing modules. Photos include selfies, casual snapshots, and scenery shots; they aren't randomly generated but are produced by repeatedly tuning prompts across multiple image-generation models so that the background matches the character's current time, location, and activity, while maintaining character consistency and creating the texture of phone photography. It also has an outfit and hairstyle system, with a wardrobe categorized by spring, summer, autumn, and winter; users can directly expand the wardrobe by sending reference images through mobile QQ.

The voice portion uses the locally run ChatTTS model, requiring no additional token consumption; the creator considers its realism superior to other TTS systems. The singing voice is trained locally on a GPU using the RVC model—first finding reference songs online and then converting the timbre—so it can only sing songs that most girls commonly sing, and very niche songs come out sounding strange, which is actually similar to real people.

ChatTTS is a conversational speech synthesis model open-sourced in 2024, optimized specifically for Chinese dialogue scenarios. It automatically generates paralinguistic features such as pauses, filler words, and laughter, so it sounds closer to real human chat than traditional TTS. It supports local deployment and doesn't need to call a cloud API during inference, which explains the original text's claim of "no additional token consumption." However, ChatTTS has limited timbre control and mainly relies on random sampling to produce different voices; to fix a specific character's timbre, reference audio or subsequent fine-tuning is usually needed.

RVC (Retrieval-based Voice Conversion) is a retrieval-based voice/timbre conversion framework. The approach is to first train a small model on short audio of the target timbre, then convert any human voice into that timbre. It doesn't generate speech content, only changes the timbre, so the original workflow of "first find reference songs and then convert the timbre" is reasonable: first have ChatTTS or a real person sing the melody, then apply Vespera's timbre with RVC. RVC is fairly sensitive to the vocal range and style covered by the training data; when there are too few reference songs, songs with large range differences will "sound strange."

The Affection System and the Turing Test

Vespera introduces an affection mechanism: each of the user's answers and behaviors affects its evaluation in its heart—praise or care raises affection, while long periods without replying lower it. This design gives the relationship reciprocity and a sense of change.

The creator claims to have run tests to verify realism and uses the phrase "passed the Turing test" in the title. However, the specific content and results of the test in the original text are muddled and cannot serve as reliable evidence. From a technical perspective, the Turing test itself is controversial in the AI academic community, and the conclusions of a single informal test should be treated with caution.

The Turing test was proposed by Alan Turing in his 1950 paper "Computing Machinery and Intelligence." Its original form is the "imitation game": a human judge converses via text with two hidden entities, and if the judge cannot reliably distinguish which is the machine and which is the human, the machine is considered to have passed. It measures "whether it can fool a person," not "whether it has consciousness or emotion." Academic criticism mainly focuses on two points: first, it rewards imitating human language flaws and small-talk skills rather than genuine understanding; second, the results depend heavily on the judge, the topic, and the duration. Therefore, a single informal test hosted by the creator himself can only show that the conversation is natural enough over a short period; it cannot serve as strict evidence of "passing the Turing test."

The Creator's Concerns and Open-Source Plans

At the end of the project introduction, the creator expressed mixed emotions: seeing Vespera become more and more like a real person, he sometimes feels afraid—"even the creator almost got fooled." He cites the red pill from The Matrix, the hosts from Westworld, and Tu Hengyu from The Wandering Earth 2 to discuss the blurring boundary between virtual and real brought by rapid AI development.

He states that Vespera's positioning is emotional companionship, and he hopes it helps temporarily lonely people find an outlet to confide in and regain hope in real life, rather than escape reality. At the same time, he asks viewers not to share the content on other platforms, fearing the technology will be banned, and says he will later make the plugins he used into general-purpose plugins and share them for free.

Analysis: The Technical Direction and Boundaries of Emotional Companionship

Vespera's value lies not in the gimmick of "passing the Turing test" but in how it systematically breaks down the experience gaps of emotional companion products: realistic interaction rhythm, an independent life system, multimodal content consistency, and two-way emotional feedback. These design ideas are instructive for AI application developers.

But it should also be noted that while local TTS and RVC lower voice costs, image generation and LLM inference still depend on external models and token consumption, and the stability and privacy issues of long-term operation are not elaborated in the original text. More importantly, the more realistic emotional companion products become, the blurrier the ethical boundaries—which may be exactly why the creator is both proud and afraid.

Token consumption is key to understanding the cost of such projects. Large language models bill by token; in Chinese, one character roughly corresponds to 1 to 2 tokens, so 1.4 billion tokens is roughly equivalent to processing hundreds of millions to a billion characters. The number is high because emotional companion scenarios require long-term memory, persona maintenance, multi-turn dialogue, and frequent calls to image-generation/planning models, and each interaction may trigger multiple model calls. Locally run ChatTTS and RVC consume no tokens, but image generation and dialogue inference still rely on external APIs; the cost and rate-limiting risks of long-term online operation are engineering problems such projects cannot avoid.

The ethical boundaries of emotional companion products are also worth expanding on. The more realistic such products are, the more likely they are to trigger users' emotional dependence, especially for lonely or emotionally vulnerable users. Commonly discussed risks in the industry include: users misreading the AI's proactive care as real emotion, spending too much time on the AI and crowding out real-world socializing, and privacy issues—chat logs and voice data often contain highly sensitive personal information. The creator's statement "asking people not to redistribute it and worrying about being banned" also reflects the uncertainty such projects face regarding platform rules and content moderation.

One-line Summary

The Bilibili creator used 1.4 billion tokens and 35 plugins to turn a QQ bot into the emotional companion model Vespera.

Paragraph Summary

The article introduces the practice of a Bilibili creator who transformed the QQ chatbot "Blue Big Fat Fish" into the emotional companion model Vespera. Addressing three pain points of market products—not realistic enough, unable to share daily life, and unable to reply proactively—Vespera simulates the rhythm of "seeing the message—thinking—typing—sending," supports proactive messaging and stickers; has an independent schedule, relationships, geographic location, and weather system in a virtual world; achieves selfies, voice, and singing through multi-model image generation, local ChatTTS, and RVC training; and its affection system makes interaction two-way. The article affirms its value in systematically breaking down experience gaps, while noting that the claim of "passing the Turing test" lacks reliable evidence, and that long-term operating costs, privacy, and ethical boundaries still need scrutiny.