On 29 january, ji-ji reported that the previous day, the search dog input method had been officially declared fully ai, with an upgrade of the 20. 0 version of ai and a full upgrade of the three models: ai voice, ai translation and ai typing。
This is an extremely large update of the dog search method in 20 years:
In the area of voice, the main type is “faster, more accurate and less vocal” and the addition of a “spoken-to-written” capability supports the automatic conversion of the content of the scattered language into a standard written language and optimizes the replacement of common confusing words, such as “he or she”。
In the area of translation, a large-scale translation model has been introduced to support “input-in-translation” in more than 30 languages in real time。
In the area of ai typing, the upgrade of the model enhances the understanding of the language, and the vocabulary is upgraded at the sky level to support multiple high-impact additions to the facade。
For advertising problems that bother users, the mobile side adds a " one-click off keyboard advertising" feature, which is expected to go online in february。
The previous version of the dog search process has been used on line ai, ai search, ai synthesizing expression, graph recognition, magma search, etc., with hundreds of millions of search requests。
At present, the dog-searching system is home to hundreds of millions of ai users, with almost 2 billion daily voice usages and now number one in the industry。
So, what's the "most powerful version of ai" going up? We actually experienced it once。
I. Voice typing: soft sound and verb colour, but long sections and chimpanzee are clearly short board
In the midday break of the office, we touch the microphone above the keyboard and enter it in a way that is significantly less than the normal volume of speech: “the summary of the meeting needs to be delivered tomorrow morning”。
Under such soft input conditions, dog searches allow for steady identification, text proofing is not required and is indeed available and suitable for ad hoc handling of information in a quiet environment。
And then we deliberately tested the text of a confused, jump-starting speech: “does anyone say that if you don't go, i won't go there?”
Under the automated “port-to-written” function, the system organizes the original sentence in its entirety, completes the markers and adjusts the order of the language, and produces a relatively smooth written expression。
At the same time, the problem with dog-to-dog input was mainly in the context of long voice and dialect。
We've tried to read the original text of the sms for dozens of seconds, and the system has made multiple errors in terms of cut-offs, words and the identification of names, and the overall stability is inadequate。
The final output is only one or two sentences and is not very accurate, even if it is transposed in the voice setting to the chimpanzee mode, and set for an open-ended voice input. After several adjustments, experience has not improved significantly。
Overall, the dog-searching process has made some progress in soft input and short-term speech organization, but there is still significant room for improvement in the recognition of dialects such as long-term continuous voice and chili。
Ii. Translation modules: typing, voice and reproduction, but the chain is still long and the process needs to be sharpened
In the foreign language chat scene, we click on the translation portal on the keyboard, turn on the "fast translation" feature and enter "you are the brightest star in the world!"。
At this point, the chinese content entered is translated into korean in real time during the input process, the translation results appear almost in sync with typing, and the overall experience is close to " type-to-drop translation " , which is suitable for sending foreign language content directly into chat software。
In a simulation of overseas travels, we try to translate in real time by voice and say to the starter: “is 150 baht okay? If so, i'll buy one more!”。
The system will go aheadVoice recognitionThe re-export of the corresponding thai language version would theoretically complete the “on-site simultaneous transmission”。
In practice, however, this process is not smooth: it needs to be manually selected in “fast translation” and then clicked on the microphone to start the input, with the occasional replacement of the language mode and the need for re-selection, and stability in its use。
In reading the content of foreign languages, we tested the "clipboard translation" feature。
The sentence “c'est le temps que tu as perdu” in the original french version of “the little prince” is an example of a copy of the text that allows the dog-searching input to recognize the contents of the clipboard, but does not directly give the translation results。
Users need to click on the " unopen " button on the right side of the text to paint sentences, then click on the " translator " at the top right corner, before the system jumps to the dog-search translation page to show the chinese interpretation, and eventually they need to copy the translation manually back to the original interface。
As a result, there are no obvious problems with the quality of translation per se, the meaning is accurate and the tone is natural, but access to different modes of translation is more dispersed on the operational path: translation by hand, translation by voice, translation by clipboard is independent, and some steps require multiple confirmations or jumps, which is somewhat cumbersome for use。
Overall, the current ai translation of dog search input is more like a position of capability, but interaction is still at the functional assembly stage, and there is some room for optimization from the “hand-on” state。
At the end:
In this update of the dog search input method, we can also have a complete experience of the ai typing function。
During these three days, we used it as a dominant input method, and in practical terms, the ai model's ability to recognize spelling abbreviations or its reliance on individual input habits, with a certain degree of variability in the combined effect, although the overall performance was stable and relatively “handy”。
Overall, during the current experience, voice input has performed well in terms of “low-speak” and “short-speech writing”, but its ability to transliterate and recognize dialects is not yet mature。
While the translation function is diverse, multiple entry points are more fragmented and parts of the process rely on manual leapfrogging and reproduction and are not sufficiently integrated。
It should be noted that this test is based on the ios end, and that at the anzor end, the search dog input method has come online to include more new ai features, such as “mixable smart replacement”, with actual experiences or differences。
According to the dog search input team, the subsequent version will continue to be highly iterative。
