Synthetic medical speech dataset for fine-tuning Whisper-based ASR models on clinical and nursing terminology. Contains 101,475 audio-text pairs totaling 184.1 hours of speech at 16 kHz mono, generated using Kokoro-82M TTS with 19 voices across three English accent groups.

This is v2 -- a companion to the v1 dataset (125,500 samples, ~257 hours). v2 focuses on terms from additional data sources (RxNorm API, FDA openFDA, LOINC, CMS/HCPCS, medical abbreviations, hand-curated nursing terms) that were not covered in v1. All v2 sentences were deduplicated against the full v1 sentence set (179,637 sentences, case-insensitive).

Metric	Value
Samples	101,475
Duration	~184.1 hours
Sample rate	16,000 Hz (mono)
Audio format	WAV (float32) embedded in Parquet
TTS model	Kokoro-82M (`hexgrad/Kokoro-82M`)
TTS speed	0.94x
Voices	19 (12 American, 5 British, 2 Indian English)
Source terms	78,706 unique medical terms

Data Sources

All terms were collected from public/licensed medical terminology databases. No patient data (PHI) was used.

Source	API / Database	Terms Collected	License	Notes
RxNorm	`rxnav.nlm.nih.gov/REST/`	69,769	Public Domain (US Gov)	Drug names, dosage forms, branded + generic
FDA openFDA	`api.fda.gov/drug/`	7,673	Public Domain (US Gov)	Drug labels, active ingredients, brand names
LOINC	Local JSON (top 500 panels)	862	Regenstrief License (research use)	Lab test names and observation codes
Medical abbreviations	Local CSV (104K abbreviation file)	104	Public Domain compilation	Common clinical abbreviations (BID, PRN, etc.)
Nursing terms	Hand-curated	258	Original (IntelMedica)	Nursing assessment, vitals, wound care, triage
CMS/HCPCS	Curated procedure codes	40	Public Domain (CMS)	Healthcare procedure codes

Total unique terms: 78,706

What's New vs v1

Dimension	v1	v2
Samples	125,500	101,475
Duration	~257 hours	~184.1 hours
Drug terms	ICD-10, UMLS, FDA NDC	RxNorm API (69K), FDA openFDA (7.7K)
Lab tests	None	LOINC top 500 (862 terms)
Procedures	None	CMS/HCPCS (40 terms)
Nursing-specific	Minimal	258 hand-curated terms (vitals, wound care, triage, SBAR)
Abbreviations	From UMLS expansions	104 common clinical abbreviations
Combo sentences	None	Drug+side-effect combos, vitals+assessment combos
Term traceability	No `term` column	`term` column links every sentence to its source term
Deduplication	Self-deduplicated	Deduplicated against all v1 sentences (179,637)

Combined v1+v2: 226,975 samples, ~441 hours of medical speech.

Known Gaps and v3 Coverage Plan

This section documents what v2 does NOT cover well, and how v3 will address each gap.

Category Imbalance

v2 is 92% drug-related (93,359 of 101,475 sentences). This reflects the dominance of RxNorm/FDA as term sources. Non-drug clinical language -- conditions, procedures, body structures, patient assessments -- is underrepresented.

Gap	Current State (v2)	v3 Plan
Conditions / diagnoses	Only via ICD-10 in v1	SNOMED CT (~350K active concepts covering conditions, procedures, body structures)
Radiology terms	None	RadLex (~68K terms from RSNA, free with BioPortal registration)
Lab tests	Only top 500 LOINC	Expanded LOINC (full ~100K observations, free with Regenstrief account)
Oncology / pathology	None	NCI Thesaurus (~180K concepts, public domain, NCI REST API)
Clinical narratives	Template-based only	PubMed abstract mining (~~35M abstracts, NLM E-utilities) + openFDA FAERS adverse event narratives (~~25M reports)
Drug relationships	Flat drug names only	DailyMed (~140K drug labels with indications, contraindications, dosing in natural language)
Procedure codes	40 HCPCS terms	HCPCS Level II expanded (~7,500 codes, CMS annual download)
Clinical trials language	None	ClinicalTrials.gov (~480K trials with condition/intervention text, v2 API)
Drug class hierarchies	None	RxClass (ATC, VA, EPC, MoA drug classifications via NLM API)

Accent Diversity

v2 has only 3 accent groups (American 63%, British 26%, Indian 11%) using 19 Kokoro preset voices. Real hospital workforces include Filipino, Nigerian, Arabic, Spanish, Chinese, Korean, Caribbean, and many more accents.

Gap	v3+ Plan
Filipino/Tagalog accent	Voice cloning with F5-TTS or Fish Speech from volunteer recordings (5-10 min reference audio)
Nigerian/West African accent	Voice cloning pipeline (same approach)
Spanish (Mexican) accent	Voice cloning pipeline
Arabic accent	Voice cloning pipeline
Chinese (Mandarin) accent	Voice cloning pipeline
Caribbean accent	Voice cloning pipeline

Sentence Quality

v2 uses template-based sentence generation (~12 templates per category). While functional, these sentences sound formulaic and don't capture the full range of how nurses actually speak during documentation.

Gap	v3 Plan
Robotic/template feel	LLM-generated contextual clinical scenarios (Qwen 3.5 or similar) with varied sentence structures
No multi-sentence context	Generate full clinical vignettes: chief complaint -> assessment -> plan, not isolated sentences
No conversational patterns	Mine MedDialog (300K doctor-patient dialogues) and MIMIC-III clinical notes (2M notes) for natural phrasing
Limited nursing workflow	Generate scenario-specific sentences: shift handoff (SBAR), medication admin (5 rights), wound assessment (Braden scale), fall risk (Morse)

Full v3 API Source List

These are the APIs and databases planned for v3 term collection:

Source	Est. Terms/Concepts	License	Access
SNOMED CT (Snowstorm API)	~350,000	UMLS License (we have NLM-10000066889)	Ready
DailyMed (NLM REST)	~140,000 drug labels	Public Domain	Ready
openFDA FAERS (deep)	~25,000,000 events	Public Domain	API key (free)
PubMed E-utilities	~35,000,000 abstracts	Public Domain	API key (free)
RadLex (BioPortal)	~68,000	Free (RSNA)	BioPortal account (free)
MeSH (NLM)	~30,000 descriptors	Public Domain	Ready (UMLS)
NCI Thesaurus (REST)	~180,000	Public Domain	Ready
ClinicalTrials.gov v2 API	~480,000 trials	Public Domain	Ready
HCPCS Level II (CMS)	~7,500	Public Domain	Ready
LOINC expanded	~100,000	Regenstrief License	Account (free)
RxClass (NLM)	Drug class hierarchies	Public Domain	Ready
MED-RT (NCI/NLM)	~18,000 relationships	Public Domain	Ready
DrugBank Open	~14,000	CC BY-NC 4.0	Registration (free)

Generation Method

The v2 pipeline has three stages:

Term Collection (collect_terms_v2.py) -- Pull terms from RxNorm REST API, FDA openFDA API, LOINC JSON, local abbreviation CSV, hand-curated nursing terms, and CMS/HCPCS codes. Deduplicate across sources.
Sentence Generation (generate_sentences_v2.py) -- Wrap each term in clinical sentence templates (~12 templates per category). Templates simulate nursing documentation language (medication administration, assessment, vitals, handoffs). Deduplicate against all v1 sentences.
Audio Synthesis (synthesize_audio_v2.py) -- Kokoro-82M TTS on GPU at 0.94x speed, native 16 kHz output. Round-robin voice assignment across 19 voices with accent-weighted distribution (63% US, 26% UK, 11% IN).

Columns

Column	Type	Description
`audio`	Audio (16kHz)	WAV audio embedded in Parquet
`text`	string	Transcription text (the sentence spoken)
`term`	string	Source medical term the sentence was generated from (for traceability)
`speaker_id`	string	Kokoro voice identifier (e.g., `af_sarah`, `am_adam`)
`duration`	float	Audio duration in seconds
`category`	string	Term category (drug, lab_test, vitals, etc.)
`source_api`	string	Data source (rxnorm, fda, loinc, nursing_curated, etc.)
`accent`	string	Accent group (en-us, en-gb, en-in)
`sample_rate`	int	Always 16000

Category Distribution

Category	Count	%
drug	93,359	92.0%
side_effect	2,122	2.1%
lab_test	1,668	1.6%
vitals_assessment_combo	1,425	1.4%
drug_side_effect_combo	1,000	1.0%
warning	643	0.6%
active_ingredient	555	0.5%
abbreviation	191	0.2%
assessment	138	0.1%
procedure	118	0.1%
vitals	80	0.1%
medication_admin	50	<0.1%
sbar_handoff	40	<0.1%
wound_care	40	<0.1%
route	26	<0.1%
triage	20	<0.1%

Accent Distribution

Accent	Voices	Count	%
American English (en-us)	af_sarah, af_nova, af_sky, af_bella, af_jessica, af_nicole, am_adam, am_echo, am_eric, am_liam, am_michael, am_onyx	64,092	63.2%
British English (en-gb)	bf_emma, bf_isabella, bm_george, bm_lewis, bm_daniel	26,703	26.3%
Indian English (en-in)	hf_alpha, hm_omega	10,680	10.5%

Total: 19 voices (12 American, 5 British, 2 Indian)

License

CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International)

You may share and adapt this dataset for non-commercial purposes with appropriate attribution.

Attribution: Junaid Farooq, MD / IntelMedica LLC

Component licenses:

RxNorm / FDA openFDA: Public Domain (US Government)
LOINC: Regenstrief Institute License (research use permitted)
Kokoro-82M TTS model: Apache 2.0
Medical abbreviations: Public Domain compilation
UMLS License: NLM-10000066889 (for abbreviation expansions)

Citation

@dataset{farooq2026intelmedica_medical_tts_v2,
  author = {Farooq, Junaid},
  title = {IntelMedica Medical TTS Dataset v2 (16kHz)},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/intelmedica/medical-tts-parquet-2-16khz},
  organization = {IntelMedica LLC},
  organization_url = {https://intelmedica.ai}
}

Author

Junaid Farooq, MD IntelMedica LLC Physician-Led Open-Source Medical AI

Disclaimer

This dataset is for research purposes only. It is NOT a medical device, NOT clinical decision support software (SaMD), and NOT intended for direct patient care. All audio is synthetically generated from public medical terminology databases -- no real patient data (PHI) is included. The dataset is designed for training and evaluating automatic speech recognition models in medical/nursing domains.

Downloads last month: 14

Total file size:

21.2 GB