RakhatM/kk-ru-pharma-tts

Kazakh/Russian Pharmaceutical TTS Corpus

A synthetic speech corpus of pharmaceutical / clinical phrases in Kazakh (kk) and Russian (ru), synthesized with a multilingual Orpheus TTS model. Designed for ASR auto-adaptation experiments: the splits cover seen / unseen speakers and matched / unseen evaluation conditions for benchmarking domain and speaker generalization in low-resource medical ASR.

Splits

Split	Clips	Purpose
train	27,182	Training. RU voices: Elena, Alexey. KK voices: Marzhan, Madina, Saule, Akzhol, Madi, Ali, Berik.
dev	3,152	Validation. Same speaker pool as train.
test_matched	3,366	In-domain test, same speakers as train.
test_unseen_spk	3,366	Speaker-generalization test. RU voices: Darya, Denis (held out from train).

Total: 37,066 clips, ~31 hours of audio at 24 kHz mono.

Schema

Each row contains:

Field	Type	Description
clip_id	string	Unique clip identifier, e.g. train_000001_r1.
audio	Audio	24 kHz mono PCM, embedded in parquet (HF Audio feature).
text	string	Original (un-normalized) prompt text — the transcript.
lang_code	string	kk or ru.
lang_label	string	kaz, rus, or mixed (intra-sentence code-switching).
drug	string	Drug name mentioned in the text (when applicable).
voice	string	Speaker name; see Speaker Pool below.
gender	string	male / female.
emotion	string	neutral, angry, sad, fearful, surprised.
rendering	int	1 or 2; multiple TTS renderings of the same source utterance.
duration_seconds	float	Audio duration.

Speaker pool

Code	Voices in train/dev/test_matched	Voices in test_unseen_spk
kk	Marzhan, Madina, Saule, Akzhol, Madi, Ali, Berik	(subset of the same 7 voices)
ru	Elena, Alexey	Darya, Denis (held out)

The Russian held-out split lets you measure cross-speaker generalization; the Kazakh held-out split shares the speaker pool but uses different source utterances.

Notes on TTS quality

• Numbers: source plan texts containing digits (8, 10, В12) were normalized to spelled-out forms before TTS to avoid mispronunciation:
  • RU: 8 часов → восемь часов, 10 дней → десять дней, В12 → Б 12
  • KK: В12 → Б он екі The text column in this dataset preserves the original, un-normalized form for downstream ASR training/evaluation. The audio reflects the spelled-out pronunciation.
• Mixed-language clips (lang_label="mixed") intentionally embed Russian medical terminology in Kazakh sentence structure (or vice versa) to model real clinical code-switching.
• Per-utterance variation: each source utterance is rendered with two voice/emotion combinations (suffixes _r1, _r2) for acoustic diversity.

Synthesis pipeline

• TTS model: a multilingual Orpheus-style 11-language voice TTS, run locally via vLLM batched inference.
• Audio codec: SNAC 24 kHz hierarchical RVQ (3 levels) → 24 kHz mono PCM.
• Sampling: temperature 0.6, top-p 0.95, repetition penalty 1.1, min 28 audio tokens.

Intended uses

• ASR fine-tuning / adaptation for Kazakh+Russian medical speech.
• Benchmarking speaker generalization with the held-out test_unseen_spk split.
• Studying performance on intra-sentence code-switching.

Limitations

• Synthetic (TTS) audio: prosody and acoustic distribution may not match real clinical recordings. Models trained only on this corpus are not guaranteed to transfer to real medical dictation.
• Voice quality on RU Elena/Alexey is somewhat lower than on Darya/Denis — drug-name fidelity may be slightly degraded.
• Only 7 unique digit patterns appear in the source data; numeric coverage is narrow.

License

Synthetic dataset released under CC-BY-NC-4.0. See the underlying TTS model's license for any further restrictions on the synthesized audio.