danielrosehill/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 layer of personal annotations.

What it is useful for

Three use cases drove the curation:

1. Background-noise / denoising evaluation — clips with realistic, varied noise floors and known DNSMOS-style quality scores, to probe how denoisers and noise-suppression models behave on consumer-grade voice notes.
2. ASR evaluation on imperfect material — speech-to-text systems are typically benchmarked on relatively clean speech. These clips let you measure degradation on the kind of audio people actually capture on phones, with a human transcript already in place for reference.
3. Voice-note classification — the speaker's own annotation of what each note is (to-do list, note-to-self, diary entry, email draft, blog idea, podcast prompt, outline, etc.) is the basis for a personal classification project. The same labels can be used by anyone exploring few-shot or fine-tuned classifiers over short spoken-intent audio.

Source

All recordings come from my own personal archive on voicenotes.com, a voice-note app I've been using and recommend. Transcripts in this dataset are the ones generated by voicenotes.com's pipeline (ElevenLabs Scribe under the hood, at the time of capture) — they have not been hand-corrected. They reflect what a strong commercial ASR system produced on this audio, which is itself part of what makes the dataset useful for ASR comparisons.

Audio is released as MP3 in its original form (no denoising, no normalisation, no trimming).

Layout

audio/         # <id>.mp3
transcripts/   # <id>.md   (original voicenotes.com transcript)
metadata.csv   # one row per clip

Each row in metadata.csv joins to audio/<id>.mp3 and transcripts/<id>.md via audio_relpath / transcript_relpath.

Schema (selected columns)

The full column-by-column reference lives in DATA_DICTIONARY.md. The summary below is a quick orientation.

Column	Meaning
id, uuid	stable identifier
title, speaker, recorded_at, duration_s	basic metadata (speaker is always Daniel Rosehill)
BAK, SIG, OVRL	DNSMOS P.835 scores (1–5; lower BAK = noisier background)
noise_level	bucketed BAK
audio_quality_rating	1–5 star human rating
audio_defects	list (clipping, wind, distortion, etc.)
languages	spoken languages present
hebrew_usage, background_languages	code-switch / ambient-language flags
non_intended_audio	TV, music, other speakers, etc.
note_types_multi, note_categories_multi	speaker-assigned type/category labels
subject_matter	short topic descriptor
mwp_prompt	boolean — speaker's "Morning Writing Prompt" tag
transcription_quality	human judgement of the source transcript
microphone, capture_location	capture context (location supports freehand entries)
acoustic features	rms_dbfs, peak_dbfs, crest_factor_db, clipping_ratio, silence_ratio, speech_ratio, snr_db_estimate, spectral & ZCR features, HNR proxy
transcript stats	transcript_chars, transcript_words, wpm, active_wpm

DNSMOS is reference-free, run from the Microsoft ONNX models locally; it is a proxy for perceived quality, not ground truth.

Annotation status

Annotation is in two stages, and the dataset card will be updated as the schema stabilises:

1. Release validation (current). Each candidate clip is reviewed against my voicenotes.com archive to confirm it is OK to release publicly. Clips that touch on private people, sensitive topics, or PII flagged by an automated screen are withheld. The corpus on disk is therefore deliberately a subset of what I've recorded.
2. Background-noise triggers (next). A second annotation pass will tag the type of background noise present (traffic, café, kitchen, wind, HVAC, music, other speakers, etc.) so the dataset can be sliced by acoustic environment. The schema for those tags is still firming up.

PII screening combines Microsoft Presidio + spaCy NER with keyword heuristics for medical / financial / relationship / credential mentions. It is conservative — false positives are dropped from the public set rather than risk releasing something I shouldn't.

What this is not

• Not a clean-speech corpus.
• Not multi-speaker — it's one speaker, one accent, one set of recording habits.
• Not a hand-corrected transcript benchmark — the transcripts are the ASR system's output, kept as-is.
• Not large. It is a probing / evaluation set; treat it accordingly.

Citation / attribution

If you use this dataset, please cite or link back to the Hugging Face page (danielrosehill/Noisy-Voice-Notes). Released under CC-BY-4.0.

Contact

Issues, corrections, or "please remove this clip" requests: open an issue on the Hugging Face dataset page or contact daniel@danielrosehill.co.il.