4 ms·

It really does. My native Japanese friends use the app all the time to have it explain things in Japanese and they say it speaks perfect Japanese. The accent, t

It really does. My native Japanese friends use the app all the time to have it explain things in Japanese and they say it speaks perfect Japanese. The accent, tone, everything. Sometimes I have it explain something in English for me and Japanese for him and vice versa. It can switch up the languages in the same context no problem. All kinds of phrases in both English and Japanese that are hard to explain to each other ChatGPT has no problem helping us out. Even this seems like low hanging fruit for Google to update their translation infrastructure with this technology - what's going on? Or near real time translation should be totally possible now with this technology - are there any services out there providing it?

sanxiyn 2y ago
I think what's going on is Google Translate is much cheaper.
xanderlewis 2y ago
Have you tried the voice chat feature in Japanese? It speaks in a way that’s weirdly incredibly fluent and yet simultaneously completely unnatural in its intonation. It sounds completely American, presumably because the voice model was trained on mostly American English and a bit of Japanese and that’s the result — Japanese with a Californian accent.
- bottlepalm 2y ago
  Yea I’m talking about the voice chat on mobile. My friend says it’s perfect Japanese. He doesn’t have to slow down to talk to it or anything.
numpad0 2y ago
Because real world performance is about the same across the board between various 13B models, GPT-3.5 and 4, DeepL, and Google Translate, while computational cost wildly varies. It's kind of interesting/odd that GPT translations are sometimes lauded as significantly more accurate, I think it's just that GPTs are just dropping what won't work in English to better adhere the language, and therefore there are less to be confused about.
- bottlepalm 2y ago
  It’s not the same is what I’m saying. Google Translate was never good. GPT is literally perfect translation.
  - numpad0 2y ago
    > GPT is literally perfect translation. It literally is not. GPT as MT to users is just a near-SOTA alternate implementation of machine translation, though things like open-weight 13B models at edge has huge potentials so "just" might not be the right word in that sense. I've tried round tripping a line from your comment with ChatGPT 3.5 and Google Translate - I'll be honest, I did try couple different snippets to find one that breaks worse in GPT, but GT is sometimes just more accurate: OG: Sometimes I have it explain something in English for me and Japanese for him and vice versa. It can switch up the languages in the same context no problem. All kinds of phrases in both English and Japanese that are hard to explain to each other ChatGPT has no problem helping us out. GPT: 時々、私は彼には日本語で説明し、私には英語で説明してもらうことがあります。言語を同じ文脈で切り替えることもできます。ChatGPTは、英語と日本語の両方で説明が難しい様々なフレーズについても問題ありません。 GT: 私には英語で、彼には日本語で何かを説明してもらうこともあれば、その逆も同様です。同じコンテキスト内で言語を問題なく切り替えることができます。お互いに説明するのが難しい英語と日本語のあらゆる種類のフレーズを、ChatGPT が問題なく助けてくれます。 GPT->GPT: Sometimes, I explain things to him in Japanese, and he explains them to me in English. We can also switch languages within the same context. ChatGPT has no problem with various phrases that are difficult to explain in both English and Japanese. GT->GT: Sometimes I'll have him explain something in English and he in Japanese, and vice versa. You can safely switch between languages within the same context. ChatGPT will help you with all kinds of phrases in English and Japanese that are difficult to explain to each other. It's just looking more perfect from your end, which is an interesting phenomenon.
    - Manabu-eo 2y ago
      Your japanese phrases are quite easy for machine translation, as they have clear subjects stated and not much ambiguity. The main problem is the implicit subject due to omission of pronoms in regular japanese, among other ambiguous things that can only be correctly translated in context. Also ommision of gender, plural, etc. The huge decoder-only models so far do a much better job grasping/guessing that and staying consistent compared to small encoder-decoder MTL models. This simple phrase any beginner in japanese would understand: "山田さん、猫いますか？はい、います。" "Yamada-san, neko imasuka? Hai, imasu." Google Translate now: Yamada-san, do you have a cat? Yes, I'm here. When tested a few years ago: “Mr Yamada, does he have a cat? Yes, it has a cat” LLama-3-8b-instruct: "Yamada-san, is there a cat? Ah, yes, there is." GPT 3.5: "Does Yamada-san have a cat? Yes, he/she does." gemini-1.5-pro-api-0409-preview: Does Mr./Ms. Yamada have a cat? Yes, they do. LLama-3-70B-instruct: "Yamada-san, do you have a cat? Yes, I do." GPT 4, Command-R+: "Mr. Yamada, do you have a cat? Yes, I do." The two google translate ones are simply terrible. Technically, they are possible translations, but I would rate them as wrong. All LLMs gave possible translations, but the larger the LLM, more likely to be correct the translation became. Well, the top two ones chose a gender, while the -san thing is a clever way to avoid chosing a gender, but not a very pure translation. If I had to give only one translation, I would pick either of those. Also of note, while some even explained each element of their translation, none said that other possible translations were possible.

sanxiyn 2y ago

I think what's going on is Google Translate is much cheaper.

xanderlewis 2y ago

Have you tried the voice chat feature in Japanese? It speaks in a way that’s weirdly incredibly fluent and yet simultaneously completely unnatural in its intonation. It sounds completely American, presumably because the voice model was trained on mostly American English and a bit of Japanese and that’s the result — Japanese with a Californian accent.

bottlepalm 2y ago
Yea I’m talking about the voice chat on mobile. My friend says it’s perfect Japanese. He doesn’t have to slow down to talk to it or anything.

numpad0 2y ago

Because real world performance is about the same across the board between various 13B models, GPT-3.5 and 4, DeepL, and Google Translate, while computational cost wildly varies. It's kind of interesting/odd that GPT translations are sometimes lauded as significantly more accurate, I think it's just that GPTs are just dropping what won't work in English to better adhere the language, and therefore there are less to be confused about.

bottlepalm 2y ago
It’s not the same is what I’m saying. Google Translate was never good. GPT is literally perfect translation.
- numpad0 2y ago
  > GPT is literally perfect translation. It literally is not. GPT as MT to users is just a near-SOTA alternate implementation of machine translation, though things like open-weight 13B models at edge has huge potentials so "just" might not be the right word in that sense. I've tried round tripping a line from your comment with ChatGPT 3.5 and Google Translate - I'll be honest, I did try couple different snippets to find one that breaks worse in GPT, but GT is sometimes just more accurate: OG: Sometimes I have it explain something in English for me and Japanese for him and vice versa. It can switch up the languages in the same context no problem. All kinds of phrases in both English and Japanese that are hard to explain to each other ChatGPT has no problem helping us out. GPT: 時々、私は彼には日本語で説明し、私には英語で説明してもらうことがあります。言語を同じ文脈で切り替えることもできます。ChatGPTは、英語と日本語の両方で説明が難しい様々なフレーズについても問題ありません。 GT: 私には英語で、彼には日本語で何かを説明してもらうこともあれば、その逆も同様です。同じコンテキスト内で言語を問題なく切り替えることができます。お互いに説明するのが難しい英語と日本語のあらゆる種類のフレーズを、ChatGPT が問題なく助けてくれます。 GPT->GPT: Sometimes, I explain things to him in Japanese, and he explains them to me in English. We can also switch languages within the same context. ChatGPT has no problem with various phrases that are difficult to explain in both English and Japanese. GT->GT: Sometimes I'll have him explain something in English and he in Japanese, and vice versa. You can safely switch between languages within the same context. ChatGPT will help you with all kinds of phrases in English and Japanese that are difficult to explain to each other. It's just looking more perfect from your end, which is an interesting phenomenon.
  - Manabu-eo 2y ago
    Your japanese phrases are quite easy for machine translation, as they have clear subjects stated and not much ambiguity. The main problem is the implicit subject due to omission of pronoms in regular japanese, among other ambiguous things that can only be correctly translated in context. Also ommision of gender, plural, etc. The huge decoder-only models so far do a much better job grasping/guessing that and staying consistent compared to small encoder-decoder MTL models. This simple phrase any beginner in japanese would understand: "山田さん、猫いますか？はい、います。" "Yamada-san, neko imasuka? Hai, imasu." Google Translate now: Yamada-san, do you have a cat? Yes, I'm here. When tested a few years ago: “Mr Yamada, does he have a cat? Yes, it has a cat” LLama-3-8b-instruct: "Yamada-san, is there a cat? Ah, yes, there is." GPT 3.5: "Does Yamada-san have a cat? Yes, he/she does." gemini-1.5-pro-api-0409-preview: Does Mr./Ms. Yamada have a cat? Yes, they do. LLama-3-70B-instruct: "Yamada-san, do you have a cat? Yes, I do." GPT 4, Command-R+: "Mr. Yamada, do you have a cat? Yes, I do." The two google translate ones are simply terrible. Technically, they are possible translations, but I would rate them as wrong. All LLMs gave possible translations, but the larger the LLM, more likely to be correct the translation became. Well, the top two ones chose a gender, while the -san thing is a clever way to avoid chosing a gender, but not a very pure translation. If I had to give only one translation, I would pick either of those. Also of note, while some even explained each element of their translation, none said that other possible translations were possible.