Noisy Voice Notes

Noisy Voice Notes A small, hand-curated set of real-world voice notes recorded over roughly a year by a single speaker (me — Daniel Rosehill), captured in everyday environments rather than a studio. Most clips contain meaningful background noise: traffic, café chatter, kitchens, public transport, wind, kids, etc. They are deliberately not clean. The dataset is intended as a small evaluation / probing set, not a training corpus. Each clip ships with the original transcript plus a… See the full description on the dataset page: https://huggingface.co/datasets/danielrosehill/Noisy-Voice-Notes.

About This Project

Noisy-Voice-Notes is a small dataset published on Hugging Face under CC-BY-4.0, consisting of voice notes recorded over roughly a year by a single speaker in everyday settings such as streets, cafés, kitchens, and public transport rather than in a studio. Most clips contain meaningful background noise, and the recordings are released as MP3 without denoising, normalisation, or trimming. Each clip is paired with the original transcript produced by the voicenotes.com pipeline (ElevenLabs Scribe at the time of capture), which has not been hand-corrected, along with a metadata table covering DNSMOS P.835 scores, acoustic features such as SNR and crest factor, transcript statistics, capture context, and speaker-assigned note type labels.

The dataset is designed as a probing and evaluation set rather than a training corpus. It includes a PII screening process combining Microsoft Presidio, spaCy NER, and keyword heuristics, with conservative removal of flagged clips. The corpus is English with some Hebrew code-switching, and it is limited to a single speaker, so it does not represent multi-speaker or clean-speech benchmarks.

The dataset is relevant to researchers and engineers evaluating background-noise suppression and denoising models against consumer-grade audio, measuring automatic speech recognition degradation on imperfect phone recordings using a human-verified reference transcript, and exploring few-shot or fine-tuned classification of short spoken-intent audio. A planned second annotation pass will tag background noise types to allow slicing by acoustic environment.

Project Details

Tags

task_categories:automatic-speech-recognitiontask_categories:audio-classificationtask_categories:text-classificationlanguage:enlanguage:helicense:cc-by-4.0size_categories:n<1kdoi:10.57967/hf/8681region:usspeechvoice-notesnoisebackground-noiseasr-evaluationdnsmospersonal-recordings

Explore More Projects