I am trying to switch from Siri to local Assist. Whisper works in German but often misunderstands things, especially when the Atom Echo is a few meters away. I saw there is a new official option vosk now. How well does this work? What other options are there?
#What is the best speech-to-text model for German these days?
cold granite
cold granite
What is the best speech-to-text model for German these days?
cold granite
So I just found out that you can use ollama as conversation agent and let the LLM correct misspelled words. This works quite realiably.
wide swan
So I just found out that you can use ollama as conversation agent and let the LL...
yeah llm "correction" based on text works pretty well depending on the model etc...
in my experience it smoothens things out nicely. but my experience is basically only in english.
cold granite
Whisper in English is fine. In German it‘s pretty bad, but since there is nothing else available, I will have to live with it. The LLM is able to correct even the worst garbage input it seems. Would be nice if something better came along. The main wyoming dev is German, but I guess training a speech model is out of his capacities as well.
wide swan
Whisper in English is fine. In German it‘s pretty bad, but since there is nothin...
its about available training data too.
cold granite
Well it has already been done for gpt4. OpenAI just doesn’t release their cash cow.
Sadly Germany is not great at AI stuff. We don’t even have one original model. France at least has Mistral. Their situation for piper is even worse than the German one though
But apart from all the lamenting, ollama as conversation agent is pretty amazing. You can even tell it stuff like „a little brighter“ and „go back to the previous color“ etc
wide swan
But apart from all the lamenting, ollama as conversation agent is pretty amazing...
yeah, I use qwen3 with ollama myself.
floral pike
What hardware are you running your ollama on, and which exact model are you guys using? @wide swan & @cold granite
onyx root
yeah, I use qwen3 with ollama myself.
same it’s peak
vagrant bramble
yeah, I use qwen3 with ollama myself.
I used to use ollama, but llama-server is so much faster
onyx root
I used to use ollama, but llama-server is so much faster
ooh i didnt know that
wait what makes it “faster”? shouldnt it be fully dependant on the model and hardware
-# mostly
wide swan
What hardware are you running your ollama on, and which exact model are you guys...
Its running on a 5060ti (16gb) on one of the servers in my rack 🙂
vagrant bramble
wait what makes it “faster”? shouldnt it be fully dependant on the model and har...
No idea, but I get 10-25 TPS higher on llama servee
llama-server --host 0.0.0.0 --port 10600 -fa 1 -ngl 99 --ctx-size 16384 --jinja --cache-ram -1 --threads -1 --temp 0.7 --top-p 0.95 --min-p 0.01 --repeat-penalty 1.0 -b 4096 -ub 4096 -hf unsloth/Qwen3-4B-Instruct-2507-GGUF:Q8_K_XL
I use it with https://github.com/mike-nott/mcp-assist but that's not related to this issue
cold granite
What hardware are you running your ollama on, and which exact model are you guys...
I am actually using gpt-oss:120b in the Ollama cloud. Not a hundred percent private then, but it‘s for free, very responsive and understands complex commands. You can even tell it to correct words that whisper regularly misunderstands.
Although I think even a 1B model would be sufficient for English, which you could run on an Intel A310. They struggle at other languages though.
I have my own home server in a microATX case