Voice in the Wild

An open-source resource hub for builders working on voice AI that holds up outside the lab - noisy rooms, bad mics, slurred speech, overlapping speakers, hostile environments.

Originally curated for the Big Berlin Hack 2026 "Voice Interface in the Wild" track, but useful for anyone shipping voice apps to the real world.

What's in here

File	What it covers
TOOLKIT.md	Categorized list of open-source repos: audio understanding, speech quality, ASR/VAD, DSP, augmentation, agent platforms
quality-dashboard/README.md	How to wire up a "quality dashboard" (DNSMOS + WER + VAD-miss + LUFS) in under an hour
examples/README.md	Six concrete project ideas with the right tools picked out for each
DATASETS.md	Noise, RIR, and stress-test datasets you can break your pipeline with
CONTRIBUTING.md	How to add a tool, fix a broken link, or suggest a new use case

The challenge (from the track brief)

Voice-driven apps work flawlessly in the lab but fall apart the moment they meet the real world. Background chatter, low-quality microphones, slurred speech, and unpredictable environments kill the user experience. Most teams discover these failures only after shipping.

Build a voice-driven application that survives real-world harsh audio conditions. Pick your own adventure - agent, game, hardware, robot - and prove your app holds up when the environment fights back.

If you only install one thing

VERSA - a unified evaluation CLI that wraps DNSMOS, UTMOS, PLCMOS, PESQ, STOI, VISQOL, WER, and speaker similarity behind a single command. Gets you 80% of the way to a quality dashboard.

Recommended starter stack

A "quality dashboard" you can wire up in under an hour. Each row maps to a different failure mode you'll want to measure.

Signal	Tool	Notes
Non-intrusive MOS	DNSMOS	De-facto ref-free quality metric (SIG/BAK/OVRL)
MOS decomposition	NISQA	Noise / coloration / discontinuity / loudness
ASR degradation	faster-whisper + jiwer	WER vs. clean reference
VAD miss-rate	silero-vad	Tiny, pip-installable, torch-hub callable
Loudness / usability	pyloudnorm	LUFS + true peak (ITU-R BS.1770)
Unified runner	VERSA	Single CLI for all of the above + more

See quality-dashboard/README.md for the wiring guide.

How to use this repo

• Browsing for ideas? Start with examples/.
• Need a specific tool? Jump to TOOLKIT.md and Ctrl-F.
• Building a benchmark? Read quality-dashboard/.
• Need to break your pipeline on purpose? See DATASETS.md and the augmentation section of TOOLKIT.md.

Contributing

PRs welcome. See CONTRIBUTING.md. Broken links, missing tools, and better one-liners are all fair game.

License

MIT - see LICENSE. Curated descriptions and structure are MIT-licensed; linked tools retain their own licenses (check before shipping).