fishisfish0614/hervoice

hervoice is a small tool that tries to capture not just what you say, but how you say it, so that an AI assistant you connect it to can respond to your tone rather than only your words. You hold down a button and speak, and the tool listens for both the content of your speech and the emotional cues behind it. Under the hood it works in a few steps. First it transcribes your speech into text using Whisper, an existing speech to text system. Then it analyzes the audio itself using a library called librosa to measure things like pitch, loudness, pauses, and speaking speed. Finally it sends both the transcribed text and those acoustic measurements to a large language model, which decides which of eight emotions best fits the moment: happy, sad, angry, tired, tender, excited, anxious, or neutral. Importantly, hervoice does not generate any reply itself. Its only job is to listen and understand. Once it has judged the emotion, it logs the result and sends it, through a webhook you configure, to whatever AI assistant or notification system you have set up on your own server, and that system decides how to respond. The README places real weight on privacy. By default, the recorded audio is deleted right after analysis, keeping only the transcribed text and the emotional judgment, though you can choose to keep the audio if you want. All API keys are meant to be stored as environment variables rather than hardcoded, and no data is sent to any third party beyond whichever transcription and language model services you choose to connect. Running it requires a Python environment, a Groq API key for the transcription step, an API key for a language model, and ffmpeg installed for audio conversion, it starts a small local web server you open in a phone browser and hold to talk. This project was originally built for personal use and later open sourced, so it fits developers or hobbyists building their own personal AI assistant who want it to notice emotional tone, not general consumers looking for a ready-made app. It is released under the MIT license.