Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Ichigo: Local real-time voice AI

github.com

Post reply on HN
Print view View original

41–45 of 45 posts

Re: Ichigo: Local real-time voice AI

#41

Post by itake »

emreckartal wrote: post #27
itake wrote: post #23

Earlier quoted context omitted.

I just tried to use the demo website for live translation. The AI always responded in English, either ignoring my request to only respond in French or Lao, or preface the translation with english ("I can translate that to French. the translation is: ..."). I'm trying to use chatgpt for ai translation, but the other big problem I run into is TTS and SST on non-top 40 languages (e.g. lao). Facebook has a TTS library, b…

Oh, I see. We've limited it to English for simplicity for the demo. More languages are planned for future releases.

what is the limiting factor to all llama languages for stt or tts?

Re: Ichigo: Local real-time voice AI

#42

Post by gardenmud »

fooqux wrote: post #36
BugsJustFindMe wrote: post #34

Earlier quoted context omitted.

> I'm struggling to understand how a word can mean two numbers Ichi is the word for 1. Go is the word for 5.

/smacks forehead. Can't believe I fell for that.

I mean, it wasn't really a trick.

It's truly the exact same as someone saying "onefive can be read as (one five), but it's not (fifteen)" - to a non-English speaker I mean - I don't read 'prank' in that statement

Re: Ichigo: Local real-time voice AI

#43

Post by readyplayeremma »

lostmsu wrote: post #40
emreckartal wrote: post #25

Earlier quoted context omitted.

Ah, I see now. To clarify, while you can enable transcription to see what Ichigo says, Ichigo's design skips directly from audio to speech representations without creating a text transcription of the user’s input. This makes interactions faster but does mean that the user's spoken input isn't transcribed to text. The flow we use is Speech → Encoder → Speech Representations → LLM → Text → TTS. By skipping the text ste…

I understand that. The problem is that in many scenarios users would want to see transcripts of what they said alongside the model output. Like if I have a chat with a model about choosing a place to move to, I would probably also want to review it later. And when I review it, I will see: me: /audio record/ AI: 200-300m. No easy way to see at glance what the AI answer was about.

You can just run whisper on the conversations as a background job populating the text versions of all the user inputs, so it doesn't interfere with the real-time latency.

Re: Ichigo: Local real-time voice AI

#44

Post by lostmsu »

readyplayeremma wrote: post #43
lostmsu wrote: post #40

Earlier quoted context omitted.

I understand that. The problem is that in many scenarios users would want to see transcripts of what they said alongside the model output. Like if I have a chat with a model about choosing a place to move to, I would probably also want to review it later. And when I review it, I will see: me: /audio record/ AI: 200-300m. No easy way to see at glance what the AI answer was about.

You can just run whisper on the conversations as a background job populating the text versions of all the user inputs, so it doesn't interfere with the real-time latency.

It's not going to match what model hears.

Re: Ichigo: Local real-time voice AI

#45

Post by silverliver »

gnuly wrote: post #30
emreckartal wrote: post #13

Emre here from Homebrew Research. It's great to see Ichigo on HN! A quick intro: We're a Local AI company building local AI tools and training open-source models. Ichigo is our training method that enables LLMs to understand human speech and talk back with low latency - thanks to FishSpeech integration. It is open data, open weights, and weight initialized with Llama 3.1, extending its reasoning ability. Plus, we are…

any plans to share progress on open channels like matrix.org or even irc?

I second this request. Any publicly indexable channel would be fine.

I think Matrix is not publicly indexable unless the channel is unencrypted and set to public.

Post reply on HN