batyrme/medical-speech-dataset

Kazakh Medical Speech Dataset

A dataset of Kazakh medical-domain speech paired with transcriptions, for automatic speech recognition (ASR), enriched with medical and language metadata.

⚠️ Schema status — not final. The column schema below is not finalized and may change in future revisions. The current schema is the production version of the dataset.

🔒 Privacy. Clips whose transcript was flagged as containing personally identifiable information (names, addresses, phone numbers, etc.) have been removed from this dataset (3,669 clips). The remaining clips are intended to be privacy-safe, so no PII columns are included.

Overview

• 38,456 clips, ~160 hours
• Language: Kazakh (kk), with occasional Russian code-switching (natural in the source).
• Domain: medicine / healthcare — doctor consultations, health TV programs, medical podcasts, anatomy and pharmacology lectures.
• Audio: 16 kHz, mono, WAV.
• Splits: train (37,092) · dev (645) · test (719) — dev/test are held-out, speaker-disjoint episodes (see Evaluation splits).

Columns

Source columns

column	type	description
audio	Audio (16 kHz, mono)	the speech clip
transcription	string	verbatim transcript (original casing & punctuation)
duration	float32	clip length in seconds (rounded to 2 decimals)
episode	string	source episode / program name the clip was taken from

Enrichment columns (LLM-generated — see caveats)

column	type	description
medical_notations	string	comma-separated medical terms/entities appearing in the transcript, in order of appearance (surface forms, not lemmatized); empty if none
medical_domain	string	comma-separated medical domain(s)/specialties from a fixed taxonomy (e.g. Cardiology, Infectious Diseases, Pharmacology & Medications); empty if the clip is non-medical
dominant_language	string	dominant language of the clip: Kazakh or Russian
is_code_switched_russian	bool	true if the clip mixes Kazakh and Russian (code-switching)

Duration distribution

Range 4.0 – 30.0 s, mean ≈ 15.0 s, median ≈ 14.9 s, std ≈ 4.2 s.

Usage

from datasets import load_dataset
ds = load_dataset("batyrme/medical-speech-dataset", split="train")
row = ds[0]
print(row["transcription"], row["duration"], row["episode"])
print(row["medical_domain"], row["dominant_language"])
audio = row["audio"]   # {'array': ..., 'sampling_rate': 16000}

Notes & caveats

• Transcriptions were human-corrected but are not guaranteed perfect; they are verbatim (casing + punctuation preserved) so you can normalize as needed.
• Clips are cut at word/sentence boundaries; non-speech segments (music, ads, untranscribed speech) are excluded, so each clip's audio is covered by its transcription.
• Medical-term density is roughly constant across clip lengths (~6 terms / 100 words); shorter clips less often contain an explicit term simply because they have fewer words.
• The enrichment columns (medical_notations, medical_domain, dominant_language, is_code_switched_russian) were generated automatically with a large language model (Qwen3.5-122B). They are provided as convenience metadata and may contain errors.
• Privacy filtering was performed automatically from the transcripts; while flagged PII clips were removed, automated detection is best-effort and not a guarantee that every trace of sensitive information has been eliminated.

Evaluation splits

dev and test are whole held-out episodes chosen so their speakers do not appear in train — verified with ECAPA-TDNN speaker embeddings — to avoid speaker-overlap bias in evaluation.

• dev (645 clips): episodes 177, 226, 249
• test (719 clips): episodes 183, 225, 233, 237
• All other episodes (including the part-1 rows) are in train.