Feature Request: Jinja2 post processing of Voice Assistant Response

User87 · July 6, 2025, 9:35pm

Feature Request: Voice Assistants using LLMs often respond with markdown formatting of headlines, unordered bullet lists, etc. This is extremely problematic for smaller LLMs that have limited ability to follow formatting instructions like Gemma3n:e2b. The “Asterisk Asterisk” markdown formatting is killing the flow of Piper TTS. It would be nice to use Jinja2 to remove these characters directly in the voice assistant pipelines.

llm_response | replace (“**”,“”)

mchk · July 7, 2025, 10:52am

Update piper, in the latest versions the removal of * and # has been added

User87 · July 12, 2025, 11:31am

Thanks for the recommendation. I run Piper in a docker container. How did you configure this? Looking for documentation changes so I can implement.

mchk · July 13, 2025, 10:28am

Add this parameter for 1.6.x, that’s enough

github.com/rhasspy/wyoming-piper

wyoming_piper/main.py

4ffad3ce6


      
          parser.add_argument(
              "--auto-punctuation", default=".?!", help="Automatically add punctuation"
          )
          parser.add_argument("--samples-per-chunk", type=int, default=1024)
          parser.add_argument(
              "--max-piper-procs",
              type=int,
              default=1,
              help="Maximum number of piper process to run simultaneously (default: 1)",
          )
          parser.add_argument(
              "--streaming",
              action="store_true",
              help="Enable audio streaming on sentence boundaries",
          )
          #
          parser.add_argument(
              "--update-voices",
              action="store_true",
              help="Download latest voices.json during startup",
          )