Check out https://talkto.chat - talk to "DaVinci" model while keeping context. Let me know your thoughts!
#Talkto.chat - DaVinci based voice chat with context
rustic lance
hushed crest
Please add text input for users without a microphone
rustic lance
Havent thought about this - in this case ChatGPT would make more sense?
zenith fiber
This isn't working well for me, do I just click the microphone to end what I am saying?
rustic lance
Sorry to hear. Yes you need to tap/click to start and a second time to end. On Android it can automatically stop when silent. On IOS, the reason is that it apparently requires a User Interaction to play audio, therefore the second tap/click to stop. Thanks for the feedback though.
rustic lance
@zenith fiber which device/browser are you using?
zenith fiber
Avast Secure Browser, I can try Firefox
Doesnt work in my firefox, maybe I have settings against popups
Tried it in google chrome and it works fine
fickle abyss
Cool. Can I ask what kind of TTS and STT you are using?
inner musk
Works great!
rustic lance
@zenith fiber I've only tested in Chrome and Safari for both desktop and mobile version, if the demand picks up I will invest more time in other browsers. Has been a bit of a headache getting it to run in both Chrome and Safari so far.
@fickle abyss Sure! For TTS it's using Google Cloud text-to-speech which was easy to integrate with Google Cloud Functions/Firebase. For STT it's using the browser native Speech recognition which in the back is using services from Google and Apple depending on the browser (Chrome/Safari)
rustic lance
@zenith fiber On Firefox: I've briefly looked into it, unfortunately FF Desktop and Mobile don't support SpeechRecognition per default. Making it work using e.g. an external STT like OpenAI Whisper would be a bit more involved. Looks like there is a flag that can be enabled in FF but I can't test right now. https://developer.mozilla.org/en-US/docs/Web/API/SpeechRecognition
fickle abyss
<@387972437901312000> Sure! For TTS it's using Google Cloud text-to-speech which...
I Was not aware browsers had that capability 😮
flat osprey
@rustic lance I am interesting in working on projects like this. Do you have your code somehwere?
versed fjord
what language model is DaVinci based on?
rustic lance
@versed fjord it's a GPT3 based model, here's the documentation https://beta.openai.com/docs/models/gpt-3#:~:text=TRAINING DATA-,text-davinci-003,-Most capable GPT
@flat osprey I haven't release the source code yet. Is there anything specific you would like to know?
zenith fiber
<@193591036021309440> I've only tested in Chrome and Safari for both desktop and...
Right on, no worries. Currently working on my gpt app I'll publish once I work through the final CORS headaches on heroku 😃🫠
icy gust
@zenith fiber Hi! I've ben testing and works so good. One question, what configuration do you use in the request about temperature, echo, stop etc... I've tried to use de davinci-003 but it has a lot of lattency and doesn´t keep the context of the conversation. Tks
rustic lance
@icy gust I'm using the bare defaults, e.g. only prompt, "model": "text-davinci-003" and "max_tokens": 1024. No other parameters are being set/modified. The context of the conversation is kept by prefixing the prompt with the prior conversation and to modify the user input and to include history pairs of "I said: <user prompt>" and "You replied: <model completion>". The final line of the current prompt then "I say: <latest user prompt>, You reply: <leave empty for completion>"
@icy gust The latency varies, it's usually very quick, e.g. 1-3 seconds, but I'm also getting error messages "That model is currently overloaded with other requests. " currently after like 60s of waiting
icy gust
Tks @rustic lance
toxic mantle
This is pretty neat. The transcription is fast and accurate, but I'm getting an issue where anything I say is repeated twice.
rustic lance
@toxic mantle thanks, can you post or dm a screenshot and let me know your browser and device?
hushed crest
Text input please
rustic lance
You can now rickroll your friends with opening prompts 🙂 https://talkto.chat/l/qj4k
solemn cloak
a french version pls he cannot understand my beautifull accent 😅
wild mauve
I don't know if this is a bug. Sometimes it will stop recording before I stop talking.
rustic lance
@solemn cloak If you visit the site from France it should default to French when you speak
@wild mauve which browser are you using?
Ok, use 300 words to reword Martin Luther Kings I have a dream in the slang of generation Z -> https://talkto.chat/l/G2Pv
rustic lance
@hushed crest I've added text input and mute
Could some of you try the new character impersonation feature and give me your feedback? https://talkto.chat/@KITT or https://talkto.chat/?@Koyomi Araragi for anime Koyomi Araragi
you can choose any character you want by going to https://talkto.chat/@<character name>
hushed badge
@rustic lance is there a way to talk in a Dutch language and to receive Dutch answers?
I understand the voice input only accepts English languages.
Does it use Google's API voices?
hushed badge
<@391326197876916224> If you visit the site from France it should default to Fre...
I visit the site from the Netherlands
Still only English possible.
hushed badge
Can you add NL to this?
rustic lance
@hushed badge I've added NL but haven't tried
hushed badge
<@689011959663689746> I've added NL but haven't tried
Okay. Will check it. Hold 👍
hushed badge
<@689011959663689746> I've added NL but haven't tried
Works. It adjusted the voice to NL. Also the recognition works to translate it's text to NL. 👌
Is it possible to use another Google Voice?
rustic lance
This could be useful for people who are learning English. Apply it to that specific sector.
surreal ingot