Better question is why the Google app exists. It's just a platform for kitchen sink projects that they could not be bothered to wrap a whole app around.
Because google will a/b test(maybe unintentionally) products and eventually kill or merge them, great example now is using voice to have google help you vs gemini essentially both voice helpers.
"Do one thing well", a common Linux design philosophy, was blatantly ignored here. Search is not the place for this, sibling comment is likely right about trying to reach wider audience.
Claude.ai and chatGPT have become my go to for translation over Google Translate, they provide better idiomatic context. They aren't perfect for code, but I can chat with them in target languages (outside of English) and they're pretty fluent.
I personally attribute very little of my English learning to formal learning. And even less to speaking English. My history as I understand it: I watched TV shows, played games, watched movies. Eventually I would write to other people on the Internet. I guess I was fourteen. Note that I made the jump to writing (“talking”) to people about something I was interested in (music) instead of seeking out English communication for its own sake.
Only as an adult in a very International setting (a local sport with many foreigners) did I discover that I sucked at speaking the language. But getting good at speaking it then only took tens of hours at the most. Then it would feel as effortless as writing it.
I’m trying to learn a language now (after failing to learn another language ten years ago) and I’m taking a similar approach: exposure to lots of input (cf. “output” like writing/talking) and trying to remember words with retrieval practices like flashcards. It will be hundreds of hours until I try to speak anything.
In a way, I've been using Google tools to improving my english for a while now. A while back I was way more active on youtube, recording videos of my projects speaking english. I would then see the auto generated captions to see if the model had some trouble understanding what I was saying. Most of the time it was ok. Since I saw the model misunderstanding native english speakers as well.
I tried something similar with Google Translate, but for a different language. I'd attempt to say a phrase in my target language and see if it translates to what I was trying to say in English.
Fascinating. I had only ever used the SMS version so that the whole thing was really centered around collecting voice samples is a double surprise to me.
Additional data points. Now Google also knows your vocal profile and any subtle stress factors -- or lack thereof -- emerging in the context of each topic/theme used in your practicing, as well as your proficiency level and your learning rate. Not bad.
That’s a bit too cynical. I agree the nefarious data guzzler of an ad machine is siphoning up that voice data but hey if that’s all they wanted they had Google Fi/Voice etc.
This definitely seems like a Google building products people will love. Let’s just hope they don’t kill it in 2 days/weeks/months/years
I'm considering adding a similar kind of practice to my vocabulary learning app: https://vocabuo.com
Translating sentences is problematic, because they can often include other words that users don't know and there are often many possible translations, which is difficult to check.
The task "say a sentence using the word [word]" and then correcting the sentence using some LLM seems like a very promising, especially when combined with spaced-repetition flashcards.
Really cool, but why on earth is this in Google Search and not Translate?!
Better question is why the Google app exists. It's just a platform for kitchen sink projects that they could not be bothered to wrap a whole app around.
Because google will a/b test(maybe unintentionally) products and eventually kill or merge them, great example now is using voice to have google help you vs gemini essentially both voice helpers.
To get wider audience? Don’t know why they want that. Some audio data collecting?
Yes. They want all of your data.
Search, Assistant, Google, Gemini, ... It's all gSiri
"Do one thing well", a common Linux design philosophy, was blatantly ignored here. Search is not the place for this, sibling comment is likely right about trying to reach wider audience.
Claude.ai and chatGPT have become my go to for translation over Google Translate, they provide better idiomatic context. They aren't perfect for code, but I can chat with them in target languages (outside of English) and they're pretty fluent.
It’s Unix, not Linux, philosophy fyi: https://en.m.wikipedia.org/wiki/Unix_philosophy
2 replies →
I personally attribute very little of my English learning to formal learning. And even less to speaking English. My history as I understand it: I watched TV shows, played games, watched movies. Eventually I would write to other people on the Internet. I guess I was fourteen. Note that I made the jump to writing (“talking”) to people about something I was interested in (music) instead of seeking out English communication for its own sake.
Only as an adult in a very International setting (a local sport with many foreigners) did I discover that I sucked at speaking the language. But getting good at speaking it then only took tens of hours at the most. Then it would feel as effortless as writing it.
I’m trying to learn a language now (after failing to learn another language ten years ago) and I’m taking a similar approach: exposure to lots of input (cf. “output” like writing/talking) and trying to remember words with retrieval practices like flashcards. It will be hundreds of hours until I try to speak anything.
https://en.wikipedia.org/wiki/Input_hypothesis
I used to use google's voice recognition a lot when learning Chinese. I would whip myself when it didn't understand me.
So my horrible pronunciation is probably in the training data.
In a way, I've been using Google tools to improving my english for a while now. A while back I was way more active on youtube, recording videos of my projects speaking english. I would then see the auto generated captions to see if the model had some trouble understanding what I was saying. Most of the time it was ok. Since I saw the model misunderstanding native english speakers as well.
Love this.
I tried something similar with Google Translate, but for a different language. I'd attempt to say a phrase in my target language and see if it translates to what I was trying to say in English.
I saw a language speaking quizz test in my replies to a Google search with the browser. Not in app. Probably part of an a/b testing.
And was a little bit pissed off to have my search results space wasted with a stupid widget I didn't ask for and out of the blue.
That's one way to gather massive amounts of high quality audio data to train their own Whisper AI.
I like it. It's a win win in my book.
Exactly as they did with GOOG-411, which was used to build a huge phoneme database.
https://en.wikipedia.org/wiki/GOOG-411
Fascinating. I had only ever used the SMS version so that the whole thing was really centered around collecting voice samples is a double surprise to me.
Why would you want to collect training data from English learners?
Because a lot of English is spoken by non-native speakers.
1 reply →
Improved transcription accuracy from non-native English I bet.
This doesn't make any sense. Youtube has trillions of hours of audio
... they own youtube
What is its inevitable shut down deadline?
2026-06-09
Beautiful, practice acting when?
Additional data points. Now Google also knows your vocal profile and any subtle stress factors -- or lack thereof -- emerging in the context of each topic/theme used in your practicing, as well as your proficiency level and your learning rate. Not bad.
That’s a bit too cynical. I agree the nefarious data guzzler of an ad machine is siphoning up that voice data but hey if that’s all they wanted they had Google Fi/Voice etc.
This definitely seems like a Google building products people will love. Let’s just hope they don’t kill it in 2 days/weeks/months/years
> Let’s just hope they don’t kill it in 2 days/weeks/months/years
Foo me once... Foo me 50 times...
And what's the commercial application of all this stressed syllable data?
The obvious application would be better tricking John Connor on the phone :)
https://youtu.be/MT_u9Rurrqg?si=xCEHZhDyGxucmXtW
Cross-device tracking
4 replies →
I'm considering adding a similar kind of practice to my vocabulary learning app: https://vocabuo.com
Translating sentences is problematic, because they can often include other words that users don't know and there are often many possible translations, which is difficult to check.
The task "say a sentence using the word [word]" and then correcting the sentence using some LLM seems like a very promising, especially when combined with spaced-repetition flashcards.