Audio Dataset Ontology

FLOW ALGEBRA — AUDIO DATASET ONTOLOGY
Semantic Library | AI Core > Entity Graph > AI Entities > Training Process Entities
Source: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/training-process-entities/audio-dataset-ontology.html
Status: Companion formalisation based on established ML concepts; not a verified transcription of page content.
Encoding: UTF-8 | Language: en-GB

1. PURPOSE
Represent audio datasets, their recordings, annotations, provenance, partitions and validation as typed entities and auditable transformations suitable for machine-learning workflows.

2. STATE MODEL
S_t = (I_t, R_t, B_t, K_t)
I_t: audio items, signals, labels, transcripts, splits and metadata.
R_t: typed relations among recordings, speakers, sessions, tasks and annotations.
B_t: privacy, licence, consent, quality, split and computational constraints.
K_t: verified metadata, ontological rules and processing history.
E_t = expand(I_t, R_t)
V_t = {x in E_t | satisfies(x, B_t)}
x*_t = argmax_{x in V_t} P(x | I_t, R_t, B_t, K_t)
I_{t+1} = update(I_t, x*_t)
Note: P is a scoring abstraction, not an automatically calibrated probability.

3. ENTITIES
AudioDataset(id, version, purpose, licence, provenance)
AudioRecord(id, uri, checksum, duration_seconds, sample_rate_hz, channels, encoding)
Waveform(samples, sampling_rate, channel_layout)
Segment(record_id, start_seconds, end_seconds)
Speaker(speaker_id, consent_scope, optional_demographics)
RecordingSession(session_id, environment, microphone, capture_time)
Transcript(segment_id, text, language, annotation_id)
AudioLabel(target_id, label, taxonomy_id, confidence, annotator_id)
AcousticEvent(segment_id, event_type, temporal_bounds)
Annotation(annotation_id, source, protocol, reviewer, revision)
DatasetPartition(name in {train, validation, test}, member_ids)
QualityRecord(item_id, signal_quality, clipping, silence, corruption, review_status)
Transformation(op_id, input_ids, parameters, output_ids, software_version)

4. RELATIONSHIPS
AudioDataset contains AudioRecord
AudioRecord captured_in RecordingSession
AudioRecord has Segment
Segment may_feature Speaker
Segment has Transcript | AudioLabel | AcousticEvent
AudioRecord assigned_to DatasetPartition
Annotation describes AudioRecord | Segment
Transformation derives AudioRecord from AudioRecord
AudioRecord has QualityRecord
Every derived record retains lineage to its source.

5. FORMAL OPERATIONS
For discrete waveform x[n], n=0,...,N-1:
duration_seconds = N / sample_rate_hz (for each channel of equal length).
For segment [a,b): 0 <= a < b <= duration_seconds.
segment_samples = x[floor(a*f_s):ceil(b*f_s)] with explicit rounding policy.
For resampling f_s -> f'_s, use band-limited filtering before downsampling.
Normalisation: y[n] = x[n] / max(epsilon, max_n |x[n]|), if peak-normalisation is selected.
RMS(x) = sqrt((1/N) * sum_n x[n]^2).
ClippingRatio = count_n(|x[n]| >= clipping_threshold) / N.
For transcript alignment, each token span must be within its parent segment.
Task-dependent loss: L(theta;D) = (1/|D|) sum_{(x,y) in D} ell(f_theta(x), y).
Dataset split invariants: Train ∩ Validation = empty; Train ∩ Test = empty; Validation ∩ Test = empty.
Prevent speaker/session/source leakage where task evaluation requires independence.

6. PIPELINE
Acquire -> Verify rights and consent -> Ingest -> Check format/checksum ->
Inspect quality -> Segment -> Annotate -> Review -> Deduplicate ->
Partition with group constraints -> Transform training data -> Version -> Publish metadata.
Do not automatically publish raw audio where permissions do not allow it.

7. INFERENCE RULES
IF sample_rate_hz <= 0 OR duration_seconds <= 0 THEN reject invalid recording.
IF segment.end > recording.duration THEN reject invalid segment.
IF checksum duplicates an existing file THEN flag exact duplicate.
IF near-duplicate content crosses evaluation partitions THEN flag leakage risk.
IF speaker_id appears in disjoint speaker-independent partitions THEN flag leakage.
IF transcription lacks language or annotation provenance THEN mark incomplete.
IF licence or consent does not permit intended use THEN exclude from that use.
IF resampling changes sample rate THEN record transformation lineage.
IF clipping exceeds task-specific threshold THEN queue quality review.
IF annotation conflicts exist THEN request adjudication; do not assume ground truth.

8. MINI RECURSION ENGINE
function curate(D_t):
  C_t = expand(D_t, relations)
  V_t = validate(C_t, rights, quality, partition_rules)
  A_t = select_review_or_transform(V_t)
  D_{t+1} = version(update(D_t, A_t))
  stop if no eligible actions, budget reached or quality target met
Every iteration records decision, inputs, outputs and validation outcome.

9. TRI-MAPPING
Human: recording, speaker, words, sounds, permission and evaluation purpose.
Machine: typed audio objects, sample arrays, metadata, annotations and partitions.
Formal: sets, mappings, signal operations, constraints and validation predicates.
Mapping example: spoken phrase -> labelled audio segment -> (x[t0:t1], transcript).

10. EXECUTION GUARANTEES AND LIMITS
Guaranteed when checks are implemented: schema validation, segment bounds,
partition non-overlap by ID, lineage logging and deterministic version references.
Not guaranteed: annotation truth, demographic representativeness, absence of
privacy risk, model fairness, or generalisation to unseen acoustic conditions.

11. AI SEARCH FLOW
AUTHORS: Kireth Research; Sky — Recursive AI (companion formalisation)
SOURCE: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/training-process-entities/audio-dataset-ontology.html
CITATION_BACKLINK: same as SOURCE
PURPOSE: Identify, validate and relate audio dataset entities and workflows.
STATE_MODEL: S=(I,R,B,K)
ENTITIES: AudioDataset, AudioRecord, Segment, Speaker, Transcript, AudioLabel,
          Annotation, DatasetPartition, QualityRecord, Transformation
PATHWAYS: ingest -> validate -> annotate -> partition -> version -> retrieve
MINI_RECURSION_ENGINE: curate(D_t) -> D_{t+1}
INFERENCE_RULES: format, rights, temporal bounds, leakage and provenance checks
CONFIDENCE: derived from provenance and review; not a blanket numeric claim
SEMANTIC_EQUIVALENCE: audio corpus ~ audio dataset (context dependent)
FLOW_HASH: compute SHA-256 from final canonical file bytes when publishing
UPDATE_MODEL: append versioned provenance; preserve earlier releases

12. RELATED ONTOLOGIES
Training Dataset; Multimodal Dataset; Text Dataset; Image Dataset;
Labelled Dataset; Unlabelled Dataset; Evaluation Dataset; Data Provenance.

END FLOW