Repository navigation
v3.2
Fast voice conversation with interruption (v3.2.0)
The Voice tab now answers with a natural neural voice and supports real conversation flow:
Neural TTS via Piper + paroli (Windows port): answers use Brazilian Portuguese Piper voices (faber, cadu, jeff), running locally on the CPU through paroli-stream. It is a Windows port of paroli that splits each voice into an encoder and a chunked decoder, so audio starts playing about 0.2 s after a sentence is ready. The default voice is downloaded automatically from caradepato/pt-br-voices-paroli. Windows SAPI remains available as a fallback.
Streaming speech: each sentence is spoken as soon as the LLM writes it, instead of waiting for the full answer.
Real barge-in: talking over the assistant stops the voice within milliseconds and cancels the generation. Your speech, recorded from its first syllable, becomes the next question. The history keeps only the part of the answer you actually heard.
Speech-to-text setup: whisper.cpp is now installed from inside the app into bin\whisper, with large-v3-turbo-q5_0 as the recommended model and NVIDIA Parakeet v3 as an experimental option.
The paroli runtime ships inside the installer (16 MB total). The port, the voice converter and paroli-server live in the separate paroli-windows project.