SynthioLabs/dose-benchmark

DOSE v1.0: Drug-name Oral Synthesis Evaluation

Can your voice model say retatrutide?

DOSE is a benchmark measuring how accurately text-to-speech (TTS) systems pronounce pharmaceutical drug names. It covers established brands, newly-approved names, and generic/INN names, each tested inside a real clinical sentence rather than spoken in isolation.

Version: DOSE v1.0. Live leaderboard and five public audio samples: https://synthiolabs.com/synthio-dose-benchmark

This public dataset is the held-out test set only: drug name, name type, and clinical sentence. Reference pronunciations, evaluation audio, and per-system scores are not released, so the set cannot be used as an answer key. Email rohit@synthiolabs.com to submit a model for scoring.

Why this matters

Mispronounced drug names in voice AI (IVR systems, clinical voice agents, pharmacy assistants) are a patient-safety concern analogous to LASA (Look-Alike/Sound-Alike) drug name confusion — a recognized WHO/ISMP/FDA error category. As voice AI is deployed further into clinical workflows, TTS mispronunciation of a drug name is a new, largely unmeasured extension of that same risk category.

Quickstart

from datasets import load_dataset

ds = load_dataset("SynthioLabs/dose-benchmark", split="train")

row = ds[0]
print(row["drug"], row["name_type"], row["sentence"])

Filter by name type:

generic_names = ds.filter(lambda r: r["name_type"] == "generic")
brand_names = ds.filter(lambda r: r["name_type"] == "brand")

Dataset

• 274 drug names: 128 traditional (established) + 146 new (recently approved) names. One row per name.
• Brand and generic/INN names (name_type column).
• 9 TTS systems were evaluated on these sentences (see leaderboard). Those recordings and scores are not in this download.
• Every system read the exact same clinical sentence containing the drug name, with no special pronunciation hints given to any system.

How DOSE is scored

Reference pronunciations. Each drug’s reference comes from public pharmaceutical sources, including labels, manufacturer materials, and established medical references. It is then checked against spoken clinical usage so the target reflects the accepted pronunciation — not a model’s guess. The reference strings themselves are held out.

Judge. Each system’s recording is scored 0–5 against that same reference by automated audio comparison. Several independent judgments are combined into one result; 4 or higher counts as a pass. Human reviewers are not in the scoring loop.

Columns

Column	Type	Description
drug	string	Drug name (brand or generic/INN)
name_type	class label	brand or generic
sentence	string	The clinical sentence to be read aloud

Leaderboard (overall pass rate, score >= 4)

Rank	System	Model	Pass rate
1	Synthio	Synthio RxPronounce	91.2%
2	Cartesia	sonic-3.6 (beta)	80.3%
3	ElevenLabs	eleven_v3	79.2%
4	OpenAI TTS	gpt-4o-mini-tts	77.4%
5	xAI	Grok TTS	77.0%
6	Google (Gemini TTS)	gemini-3.1-flash-tts-preview	74.5%
7	Deepgram	aura-2-thalia-en	72.6%
8	Cartesia	sonic-3.5	69.0%
9	Microsoft Azure	DragonHD Neural	63.1%

Submit your model

Email us at rohit@synthiolabs.com to submit a model for evaluation.

Citation

@misc{synthio2026dosev1,
  title   = {DOSE v1.0: Drug-name Oral Synthesis Evaluation},
  author  = {Pari, Rohit Gangadhar and {Synthio Labs}},
  year    = {2026},
  url     = {https://synthiolabs.com/synthio-dose-benchmark}
}