DOSE v1.0
Bring your
model to DOSE.
Put your voice model to the test.
Get in touch to submit it for evaluation.
.svg)
.svg)
.svg)
Oops! Something went wrong while submitting the form.
Drug-name Oral Synthesis Evaluation
Can your voice model say retatrutide?
Nine voice models read the same clinical sentences. DOSE scores the drug names and lets you hear the difference.
Nine models, ranked by drug-name pronunciation. A score of 4/5 or higher counts as a pass.
Filter by difficulty tier.
Ranked by overall accuracy
Synthio Labs leads by 10.9 points
Rank / Model
SHARE OF UTTERANCES SCORING 4/5 OR HIGHER
Pass rate
01
Synthio Labs
Synthio Labs RxPronounce
91.2%
02
Cartesia
sonic-3.6 (beta)
80.3%
03
ElevenLabs
eleven_v3
79.2%
04
OpenAI TTS
gpt-4o-mini-tts
77.4%
05
xAI
Grok TTS
77.0%
06
Google (Gemini TTS)
gemini-3.1-flash-tts-preview
74.5%
07
Deepgram
aura-2-thalia-en
72.6%
08
Cartesia
sonic-3.5
69.0%
09
Microsoft Azure
DragonHD Neural
63.1%
STRATIFIED · ERA
Every model scores lower on newly approved names. The chart connects each model’s traditional and new-name pass rates.
87.0%
Synthio Labs on newly approved names
96.1%
on traditional names
+15.1 pts
lead on new names
Pass rate by drug era
NEWLY APPROVED
TRADITIONAL
50
60
70
80
90
100%
Synthio Labs
RxPronounce
Cartesia
sonic-3.6
OpenAI TTS
xAI
ElevenLabs
Cartesia
sonic-3.5
Deepgram
Google (Gemini TTS)
Microsoft Azure
STRATIFIED · NAME TYPE
Compare brand-name results with generic and INN drug names, including the gap in percentage points for each model.
87.8%
Synthio Labs on generic / INN names
94.4%
on brand names
+12.2 pts
lead on generic names
Pass rate by name type
GENERIC / INN
BRAND
50
60
70
80
90
100%
Synthio Labs
RxPronounce
Cartesia
sonic-3.6
ElevenLabs
xAI
OpenAI TTS
Google (Gemini TTS)
Deepgram
Cartesia
sonic-3.5
Microsoft Azure
Listen to the benchmark
Explore five sample drug names. Play each model’s recording or scrub its waveform. The full dataset is available on Hugging Face with access approval.
Request the full dataset on Hugging Face ↗Drug name
Rybelsus
"We can start the patient on Rybelsus, an oral GLP-1 receptor agonist that helps stimulate glucose-dependent insulin secretion."
Reference pronunciation
rye
BEL
sus
Model
Pronunciation
Result
Synthio Labs
Synthio Labs RxPronounce
rye
BEL
sus
MATCH
Cartesia (Sonic-3.6)
sonic-3.6 (beta)
rye
BEL
sus
MATCH
Cartesia (Sonic-3.5)
sonic-3.5
rye
BEL
sus
MATCH
ElevenLabs
eleven_v3
rye
BEL
sus
MATCH
OpenAI TTS
gpt-4o-mini-tts
rye
BEL
sus
MATCH
Google (Gemini TTS)
gemini-3.1-flash-tts-preview
rye
BEL
sus
MATCH
Deepgram
aura-2-thalia-en
RYE
buhl
sus
STRESS + VOWEL
Microsoft Azure
DragonHD Neural
rye
BEL
sus
MATCH
xAI
Grok TTS
rye
BEL
sus
MATCH
Phonetic readings describe the generated audio. Match means a pronunciation score of at least 4/5.
Open an example to hear the original recording alongside the other models.
Wegovy
Deepgram
HEARD
weh-THEH-vee
TRUTH
wee-GOH-vee
Replaces the middle “g” consonant with a “th” sound.
Xofluza
xAI
HEARD
EKS-OH-EFF-EL-YOO-ZED
TRUTH
zoh-FLEW-zuh
Spells the name letter by letter instead of pronouncing it.
Farxiga
ElevenLabs
HEARD
FARK-see-goh
TRUTH
far-SEE-guh
Moves stress to the first syllable and adds a “k” sound.
Each drug name appears in a clinical sentence. Open a sample to compare how the models read it.
Rybelsus
rye-BEL-sus
"We can start the patient on Rybelsus, an oral GLP-1 receptor agonist that helps stimulate glucose-dependent insulin secretion."
Tryptyr
TRIP-teer
"We will prescribe Tryptyr drops twice daily because targeting TRPM8 thermoreceptors helps stimulate your eyes to produce natural tears."
274 drug names
Counts on a shared scale: 0–274
By era
By difficulty
1
real clinical sentence per drug
4/5
minimum score required to pass
53.3%
of names are newly approved
4
Azure synthesis failures
01
Choose the sentences
274 drug names cover established brands, generic and INN names, and newly approved drugs. Each appears in one clinical sentence.
02
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.
03
Generate the audio
All nine systems read the same clinical sentence for each drug, allowing their pronunciation of the name to be compared.
04
Score the pronunciation
Each system’s recording is scored 0–5 against the 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.
BibTeX
COPY
@misc{synthio2026dosev1,
title= {DOSE v1: Drug-name Oral Synthesis Evaluation},
author= {Pari, Rohit Gangadhar and {Synthio Labs}},
year= {2026},
url= {https://www.synthiolabs.com/synthio-dose-benchmark}
}
Put your voice model to the test.
Get in touch to submit it for evaluation.
Explore the dataset and evaluation details.
Full dataset access is subject to approval.
