Can you understand what cats are saying by wearing an AI collar that costs 800 yuan?

Which pet owner doesn't want to understand what their cat or dog is trying to say when it makes a sound, or to make them understand human language?

A company named "Moe Translator" in Hangzhou recently launched a product. For just 800 yuan, it can achieve two - way translation between us and pets, with an accuracy rate of 94.6%.

Through an AI collar with both sound - recording and playback functions, combined with a mobile app, the AI collar will convert the sounds of cats and dogs into text and display it in the dialog box of the app. Users can send messages in the dialog box, and the collar will emit "meows" or "barks" to make pets "understand" human words.

It sounds particularly unreliable. After all, there are a large number of similar products if we simply search on WeChat mini - programs. Some directly state "For entertainment only, don't take it seriously", and some also use the banner of AI, mainly analyzing emotions through AI based on recordings.

The AI pet language translation shown in the picture is driven by the Qwen - Omni large model.

On the other hand, we have no way to verify what cats and dogs are saying. The translator can simply use some general scenarios that are less likely to be wrong, such as expressions like "I'm hungry", "I want to go out", "I'm not feeling well", "Someone is coming".

And translating what we say into cat or dog language may also fail the verification due to the limited cognition of pets.

But for such a "mysterious" thing, there are still indicators to measure it, and it has achieved 94.6%.

PettiChat official website: pettichat.com

The AI collar launched by Moe Translator has also attracted a lot of attention on X. Netizens are discussing this Chinese AI pet translation startup. Some netizens directly said, "The 95% accuracy rate is based on the premise that you can verify what they say, but you simply can't. So this is pure nonsense, haha."

Despite some controversies, PettiChat is still very popular. It has successfully raised funds from 863 supporters on the crowdfunding platform Kickstarter, with a subscribed amount of 140,000 Hong Kong dollars.

During the crowdfunding stage, the product was sold at $119, approximately 800 yuan. After the crowdfunding ended, the current overseas price is $149.

According to the records in the WeChat store, 190 people have purchased the product at the pre - sale price of 799 yuan. In the product review section, some users posted pictures of themselves with the product, saying "It's interesting to occasionally hear what my furry baby is thinking."

The translated pet languages are also quite user - friendly. There are not only a large number of modal particles like "ma~", "na", "wei", "yi", "heihei", "yaya", "huhu", but also expressions full of emotions like "You're welcome" and "Don't forget about me".

Can pets really understand so much information?

So, how is the accuracy rate of PettiChat measured? Are these similar products a form of IQ tax?

Pet translation devices that have been constantly suspected and updated

In 2002, the Japanese toy company Takara launched BowLingual, a "emotion translation" project for dogs.

Its working method is very simple: the microphone records the sound, and then classifies the dog barks into several emotional states, such as "happy", "anxious", "angry". The principle is close to a gimmick, but it was actually sold, and it even won the Ig Nobel Prize. The comment said, "It achieved peaceful communication between humans and dogs, so it won the Peace Prize."

Twenty years later, similar AI tools have emerged in an endless stream. The mini - programs mentioned at the beginning, as well as dedicated applications, have started using machine learning to analyze cat meows and label each meow.

At the CES exhibition at the beginning of the year, Traini also launched an AI collar, positioned as a one - way "human - to - dog" translation. When we speak, it converts the speech into acoustic signals that dogs can understand.

The desire to establish communication with pets seems to have never been extinguished. With the evolution of technology, we are becoming more and more convinced that it is somewhat possible to communicate with cats and dogs.

PettiChat, which has attracted everyone's attention this time, has done one more thing than previous products: it has presented a set of test data.

In terms of appearance, PettiChat weighs 27 grams and can be clipped onto the collar, which indeed does not impose an additional burden on pets.

Inside this small device, there is an edge - computing chip that directly processes audio, with a minimum latency of 40 milliseconds. It doesn't need to be continuously connected to the Internet and only briefly uses cloud resources when analyzing sounds. It also has other capabilities such as IP65 waterproofing, supporting 1000 translations on a single charge, and 100 - hour GPS tracking.

According to their promotion on the crowdfunding platform, the acoustic model they use is based on more than 1.5 million pet bark samples, combined with peer - reviewed research in animal behavior. The final result is that the accuracy of recognizing emotional states based solely on sound patterns reaches 91 - 92%. After adding the dimension of posture monitoring, the comprehensive accuracy rate reaches 94.6% under laboratory conditions.

Over 5 million pet voiceprint data

There are two somewhat unclear promotional pictures on the crowdfunding page, listing the benchmarks on which these data were tested. After looking closely, we found that two papers were mentioned.

One is DogSpeak, a canine vocalization classification dataset from the top multimedia conference MM 2025.

In this paper, the author proposed a large - scale dog bark dataset called DogSpeak, aiming to study whether it is possible to determine the gender, breed, and even the specific dog just based on its barks.

The data source is dog videos on social media platforms such as YouTube and TikTok. The author first searched for videos using five breeds: Husky, Chihuahua, German Shepherd, Pit Bull, and Shiba Inu. Then, the identity, gender, and breed of the dogs were confirmed based on channel information, titles, comments, etc.

Finally, 156 dogs, 5 breeds, 77,202 dog bark sequences, and 33.162 hours of pure dog barks were obtained. These data were not labeled with what the dogs were doing for different sounds, nor were any contextual information added.

The main experimental task in the paper is to determine the gender, breed, and identify the specific dog through the sound sequences. The experimental results show that these tasks are not as easy as expected. It is very difficult to perfectly solve the problem of dog bark recognition in a real and complex environment relying solely on "pure acoustic features".

Facial expressions can be considered.

At the end of the paper, the author suggested that future research should break out of the comfort zone of traditional audio technology and explore more advanced structural, rhythmic, and even potential "linguistic" features of dogs.

The other paper is also from MM, a dataset and classification method for urban sound research in 2014. This is a classic dataset paper in the field of urban environmental sound classification. Its core contributions are the UrbanSound8K dataset and the urban sound classification method.

The author classified urban sounds into major categories such as human voices, nature, machinery, and music, and then further divided them into specific sound sources such as dog barks, car horns, sirens, drilling, air - conditioners, and street music.

PettiChat cited the datasets of these two papers for testing and, in combination with the College of Animal Sciences of Zhejiang University, accumulated over 5 million pet voiceprint data, approximately 1.5 million labeled data. At the same time, the environmental sounds from UrbanSound were added to the dataset to ensure robustness in real - world environments.

The model used by PettiChat is based on the Tongyi Qianwen large model of Alibaba Cloud. The models involved in the testing include Qwen2 - Audio without pre - training, Qwen2.5 - Omni - 7B, Qwen3 - Omni - 30BA3B, and Xiaomi Mino - V2 - Omni.

They created a large independent test set based on these voiceprint data and UrbanSound 8K, which contains "audio samples of pet barks with background noise superimposed". For example, a dog bark or a cat meow is superimposed with TV sounds, traffic sounds, household appliance sounds, street sounds, etc. to create a mixed audio closer to real life.

There are also various noise samples, which may be background sounds without pet sounds, used to test whether the model will misjudge ordinary noise as pet barks.

These samples include both pet barks with noise and pure noise/non - pet sounds, which are only used to test whether the model can accurately identify pet sounds in a complex real - world environment. Under this test, the Petti model achieved an average accuracy rate of 98