Voice Analyzer: an AI-powered voice analysis tool built with Gemini

I've been on a bit of a tear lately building proof-of-concept apps with Google's Gemini API, mostly through AI Studio which makes it absurdly easy to go from idea to working prototype. Voice Analyzer came out of a specific curiosity: we talk a lot about LLMs processing text and images, but the audio modality feels underexplored. There's an enormous amount of information embedded in a voice recording beyond just the words — tone, pace, emphasis, speech patterns, confidence, emotional state — and I wanted to see how well Gemini's multimodal capabilities could extract and articulate those non-verbal signals. The answer turns out to be: surprisingly well, with some fascinating caveats.

What it actually does with your voice

Voice Analyzer is a web application — Node.js backend, simple frontend — that takes an audio file and runs it through Gemini's multimodal API to produce a detailed analysis of the voice characteristics and speech patterns. It doesn't just transcribe; it analyses. The output includes observations about speaking pace, tonal variation, vocal energy, clarity, and the overall "feel" of the voice. I've tested it with podcast recordings, my own voice memos, and even meeting recordings, and the results are consistently more nuanced than I expected. Gemini picks up on things like hesitation patterns, changes in confidence between topics, and shifts in energy level throughout a recording. It's not perfect — it occasionally over-interprets pauses that are just someone drinking coffee — but as a demonstration of what multimodal AI can do with audio input, it's genuinely impressive.

Why this matters beyond the novelty

I build a lot of these Gemini proof-of-concept apps — there's a whole collection of them on GitHub — and the goal isn't always to build a polished product. Sometimes it's about testing boundaries and developing intuitions for what these models can and can't handle. Voice analysis is a particularly interesting frontier because the applications are surprisingly broad once you have reliable extraction of non-verbal features: coaching and public speaking feedback, accessibility tools that go beyond transcription, meeting analysis that captures dynamics rather than just minutes, and mental health applications where vocal biomarkers can signal changes in wellbeing. None of those applications need Voice Analyzer specifically, but they all need someone to prove that the underlying capability exists and works. That's what this project is: a proof point, shared openly so others can build on it.

Setup is intentionally lightweight — install dependencies, set your Gemini API key in the environment file, run the dev server — because the point is exploring what's possible, not shipping a product. Check out the repo: Voice-Analyzer on GitHub.