DATA_DICTIONARY.md

# Data Dictionary — `metadata.csv`

One row per clip. Multi-select columns are semicolon-separated (`a;b;c`). Booleans are the literals `true` / `false`. Empty cells mean "not annotated yet".

Schema version: **`noisy-voice-notes/1.2.0`** (column `schema_version` in every row).

## Identity

| Column | Type | Notes |
|---|---|---|
| `id` | string | Primary key. Stable across releases. Joins to `audio/<id>.mp3` and `transcripts/<id>.md`. |
| `uuid` | string | Secondary stable identifier (UUIDv4). Useful when re-keying. |
| `title` | string | Human title from the source app (voicenotes.com). May be empty. |
| `speaker` | string | Always `Daniel Rosehill`. Single-speaker dataset — included so downstream code never has to assume. |
| `recorded_at` | ISO-8601 | When the clip was originally recorded. |
| `duration_s` | float | Clip duration in seconds. |

## Audio quality — model scores (DNSMOS P.835)

Reference-free perceptual proxies. 1–5 scale. Computed locally with the Microsoft ONNX models.

| Column | Type | Notes |
|---|---|---|
| `BAK` | float | Background-noise quality. **Lower = noisier.** |
| `SIG` | float | Speech-signal quality. |
| `OVRL` | float | Overall MOS-style score. |
| `noise_level` | int 1–5 | Bucketed BAK. 5 = noisiest, 1 = cleanest. Thresholds: <1.5, <2.0, <2.5, <3.0, ≥3.0. |

## Audio quality — human

| Column | Type | Notes |
|---|---|---|
| `audio_quality_rating` | int 1–3 \| empty | Subjective: 1 = bad, 2 = neutral, 3 = best. Empty = unrated. |
| `transcription_quality` | enum \| empty | One of `excellent`, `good`, `fair`, `poor`, `unusable`. Reflects the source ASR transcript, not the audio. |

## Annotation — note content

| Column | Type | Notes |
|---|---|---|
| `note_types_multi` | multi-select | Broad family labels. Keys: `idea_brainstorm`, `note_to_self`, `todo_list`, `email_draft`, `prompt`, `development`, `technology_topic`, `health`, `meeting`, `feedback_bug`, `research`, `logging_record`, `content_draft`, `ai_agent`, `geo_cultural`, `misc`. |
| `note_categories_multi` | multi-select | Leaf categories under those families (e.g. `journal-entry`, `to-do-list-home`, `podcast-prompt`, `mwp-prompt`, `bug-report`). Full set lives in `note_types.py` in the source repo. |
| `subject_matter` | string | Short freehand topic (e.g. "rsync to NAS"). |
| `mwp_prompt` | bool | `true` if the clip is a "Morning Writing Prompt" / MWP-style prompt. Speaker shorthand. |
| `free_notes` | string | Catch-all freehand annotation. |

## Annotation — audio defects (multi-select)

`audio_defects` keys (semicolon-separated):

`background_music`, `background_conversations`, `crying_baby`, `traffic_sounds`, `wind_noise`, `echo_reverb`, `phone_notification`, `hvac_fan`, `vehicle_interior`, `cafe_ambience`, `street_outdoor`, `multiple_speakers`, `breathing_mouth_noise`, `mic_handling`, `distortion_clipping`, `pet_animal`, `doors_dishes`, `mechanical_construction`, `keyboard_typing`, `tv_radio_broadcast`, `poor_quality_general`.

This is the column that will grow most during the second annotation pass (background-noise tagging).

## Annotation — non-intended audio (multi-select)

`non_intended_audio` keys: `side_conversation`, `speaker_to_other`, `speaker_to_pet`, `false_start`, `environmental_speech`, `background_laughter`, `interruption`, `self_correction`, `thinking_aloud`, `filler_only`.

Captures audio that ended up in the recording but wasn't *meant* to be the note (radio in the background, false starts, addressing someone else, etc.).

## Annotation — language

| Column | Type | Notes |
|---|---|---|
| `languages` | multi-select | Languages actually spoken by the speaker. Keys: `english`, `hebrew`. Defaults to `english`. |
| `hebrew_usage` | enum \| empty | Fine-grained Hebrew usage. One of `none`, `place_names`, `single_words`, `code_switching`, `mostly_hebrew`, `all_hebrew`. |
| `background_languages` | multi-select | Other languages audible in the background. Keys: `bg_arabic`, `bg_russian`, `bg_french`, `bg_spanish`, `bg_german`, `bg_yiddish`, `bg_other`. |

## Annotation — capture context

| Column | Type | Notes |
|---|---|---|
| `microphone` | enum \| string | `phone` (default), `samson_q2u`, `other`, `unknown`. May also contain a freehand string in future revisions. |
| `capture_location` | enum \| string | `unknown` (default), `home`, `shuk_mahane_yehuda`, or a freehand location string. |
| `preprocessed_audio` | bool | `true` if the source app applied denoise / echo-cancel before this audio was exported. |

## Transcript stats (auto-computed)

| Column | Type | Notes |
|---|---|---|
| `transcript_chars` | int | Character count of the cleaned transcript body. |
| `transcript_words` | int | Whitespace-tokenised word count. |
| `wpm` | float | Words ÷ (duration / 60). Wall-clock — see `active_wpm` for a more honest figure. |

## Acoustic features (auto-computed by `enrich_hf_dataset.py`)

| Column | Notes |
|---|---|
| `sample_rate`, `channels`, `bitrate_kbps`, `codec`, `file_size_bytes` | File-level audio properties. |
| `rms_dbfs`, `peak_dbfs`, `crest_factor_db`, `dc_offset` | Loudness / dynamics. |
| `clipping_ratio` | Fraction of samples at ≥ -0.1 dBFS. |
| `silence_ratio`, `speech_ratio` | Energy-VAD silence vs speech proportion. |
| `active_speech_s` | `speech_ratio * duration_s`. |
| `active_wpm` | Words ÷ active speech seconds. More honest than wall-clock WPM. |
| `snr_db_estimate` | Median active-frame dBFS minus median silent-frame dBFS. |
| `spectral_centroid_hz`, `spectral_bandwidth_hz`, `spectral_rolloff_hz` | Mean spectral descriptors. |
| `zero_crossing_rate` | Mean ZCR. |
| `hnr_db_proxy` | Harmonic-to-noise proxy from HPSS energy ratio. |

## Filesystem pointers

| Column | Notes |
|---|---|
| `audio_relpath` | Path to the MP3 relative to the dataset root (`audio/<id>.mp3`). |
| `transcript_relpath` | Path to the transcript markdown relative to the dataset root (`transcripts/<id>.md`), or empty if no transcript was released. |
| `schema_version` | Pinned schema string. Consumers should branch on this. |