My speech-to-phrase pipeline has Dolphin3 AI running locally for the conversation agent. My pipeline is set to prefer handling commands locally. Piper is used for text to speech. When I give it a controll command it says “done” when it finishes. When I type in a quesition using Assist I get a good answer in about 3 seconds. The question is: “What can you tell me about the Lusitania?”. When I use my Voice PE with the same pipeline I get no response. There’s nothing in the logs. Does my question get flushed if there’s nothing to control? If not, what am I doing wrong or am I just expecting more than the system provides?
This ASR is not suitable for working with LLMs
What is ASR?
Speech Recognition. What tool do you use for this?
Speech to Phrase. That’s the option that offers fully local use of the Voice PE. But I use my localy running AI as the “Conversation Agent” in the voice pipeline.
The default local setup does not involve the use of llm. Change the components used on your own. You can use vosk or whisper. If you have weak hardware, you will have to use cloud services. Read the information in the relevant topics.
It doesn’t respond with a pipeline using AI for the conversation agent, Nabu Casa cloud for speech to text and text to speech either. In this case it doesn’t even work typing my question into assist. I get “Unexpected error during intent recognition”.
To begin, review this documentation; it will help identify at which stage the error occurs. Without specific details, it’s difficult to provide advice.
I switched from Dolphin3 to Mistral Typing in questions works with a speech-to-phrase pipline and with a Nabu Casa cloud base pipeline. Talking to my voice pe Workis fine but slowly with the Nabu Casa cloud pipeline. It does not work with the speech-to-phrase pipeline. I presume that speech-to-phrase consumes the sentence before it gets to the AI conversation agent. Seems like a pretty severe limitation for the local case. And I don’t understand the problem with Dolphin 3 I thought it had worked in the past. Nevertheless many thanks for all the help.