Editorially curated voice and metadata for advanced, reliable AI systems.
Every Depinti Data package is a complete, structured data product.
WAV files · 44.1 kHz · 24-bit · Mono. One utterance per clip with natural trim, clean room-tone margins, and consistent audio levels. No compression, reverb, or pitch correction applied.
JSON and CSV formats with dataset-level provenance. Per-clip fields for transcript, tone, pace, emotion, intent, emphasis words, quality status, and editorial notes.
Depending on your requirements, datasets can include News (authoritative, measured), Explainer (analytical), Conversational (natural, engaged), and Instructional (composed, professional) or other delivery styles.
Includes README, Dataset Card, Delivery Specification, Speaker Consent Confirmation, Evaluation Licence, FAQ, and SHA-256 Checksums.
Produced and human-verified by the journalist and broadcast voice professional who recorded the audio.
Human-verified spoken text exactly as recorded, paired with primary vocal register labels (authoritative, explanatory, warm).
Delivery speed classifications alongside affective quality markers distinct from vocal tone.
Communicative purpose annotations and listening-reviewed vocal emphasis word tags.
Internal QC approval confirming pass on five-point technical audio quality checklist.
Rigorous room-tone, noise floor, peak level, and clipping verifications per audio clip.
Contextual rationale describing delivery approach, register choices, and emphasis intent.
Every dataset comes with a tiered licence agreement written specifically for AI speech model development.
For teams assessing dataset fit, internal ingestion testing, and non-production benchmarking.
For internal model development, fine-tuning, R&D experiments, and internal prototyping.
For datasets contributing to a deployed commercial product or service under agreed scope.
Individually negotiated terms for large organisations, multi-product use, or custom volumes.
Fine-tune text-to-speech systems on consistent, multi-register RP British English speech data.
Benchmark speech recognition performance across news, conversational, and instructional styles.
Prototype model architectures and annotation pipelines before committing to larger training runs.
Compare your annotation approach against a professionally produced human-annotated reference.
Build British English voice interfaces with authentic, structured broadcast audio data.
Test ingestion tooling and metadata parsing against clean, verified dataset files.
Depinti RP Voice Evaluation Dataset v1.0 contains 20 recorded and annotated clips across all four delivery registers for internal review. A full dataset version of v.1.0 can be provided upon request, with different licenses available.
Custom datasets and licence is available.