Our model for the way clinicians speak.
1st of 12for finished text and medical terms
Sentences from the 44ai-clinical-de test set, exactly as Klara wrote them.
Off-the-shelf speech models aren't built for clinical work.
In the GP practice, on the ward and in the emergency department, people speak medicine, often in dialect. And when they dictate, they say the punctuation and line breaks out loud. General-purpose speech recognition is built for everyday language.

GP practice
Consultation notes, dictated between two patients

Ward
The progress note, dictated on the ward round

Emergency department
The initial assessment, once the patient is stable
That's why we built Klara ourselves.
One model for clinical speech.
Hear for yourself.
Press play. Both systems were given the same clip from the test set. Here is what each one wrote, with the red pen marking what a clinician would still have to correct before signing off.
0:00
Corti SymphonyCorti
Scribe v2 MedicalElevenLabs
gpt-4o-transcribeOpenAI
Dragon Medical OneMicrosoft
Scribe v2ElevenLabs
Nemotron 3.5 TranscribeNVIDIA
MAI-Transcribe-2Microsoft
Cohere TranscribeCohere
stt-async-v5Soniox
Whisper large-v3OpenAI
Nova-3Deepgram
Klara 44ai
misheardheard correctly, but not formatted as in the report
Four examples from the test set. Full results in the paper and on GitHub.
Less red ink.
We ran twelve speech recognition systems on three clinical dictation test sets and compared every transcript with the finished report.
The next-best system still leaves 20 corrections for every 100 words and punctuation marks. Klara leaves fewer than half as many.
9per 100
Klara
44ai
20per 100
Corti Symphony
Corti
23per 100
Scribe v2 Medical
ElevenLabs
25per 100
gpt-4o-transcribe
OpenAI
28per 100
Dragon Medical One
Microsoft
28per 100
Scribe v2
ElevenLabs
30per 100
Nemotron 3.5 Transcribe
NVIDIA
34per 100
MAI-Transcribe-2
Microsoft
44per 100
Cohere Transcribe
Cohere
46per 100
stt-async-v5
Soniox
47per 100
Whisper large-v3
OpenAI
57per 100
Nova-3
Deepgram
Formatted accuracy, %. The system's output compared with the finished report, including punctuation and formatting.
Medical term accuracy, %. Medications, diagnoses, procedures and anatomy, written correctly.
Word accuracy, %. Recognised words only, without formatting.
Word accuracy on dialect speakers, %. Word accuracy for Swiss, German and Austrian dialect speakers.
Everyone is good at textbook German.
On the ward, nobody speaks textbook German. Speech is mumbled, in dialect and unstructured.
Check it for yourself.
The scoring code, the normalisation rules and every transcript from all twelve systems are public.
$ git clone https://github.com/44ai-labs/44ai-klara-evaluation
$ cd 44ai-klara-evaluation
$ make install
$ make score
## All three test sets, weighted by audio duration
| System | Formatted | Word | Medical |
| | | | terms |
| ---------------------------- | --------- | ---- | ------- |
| Klara | 90.9 | 93.9 | 90.4 |
| Corti Symphony | 80.1 | 89.9 | 89.6 |
| ElevenLabs Scribe v2 Medical | 76.7 | 90.7 | 89.0 |
| gpt-4o-transcribe | 74.9 | 90.9 | 88.5 |
| Dragon Medical One\* | 72.2 | 80.4 | 76.7 |
| ElevenLabs Scribe v2 | 72.1 | 90.0 | 87.8 |
| Nemotron 3.5 Transcribe | 69.6 | 87.3 | 81.5 |
| MAI-Transcribe-2 | 66.0 | 90.6 | 84.6 |
| Cohere Transcribe\*† | 55.8 | 82.5 | 78.1 |
| Soniox stt-async-v5† | 53.8 | 88.0 | 83.6 |
| Whisper large-v3\*† | 53.3 | 82.1 | 78.3 |
| Deepgram Nova-3† | 42.6 | 80.0 | 73.3 |
Top quality, full data sovereignty.
You decide where Klara runs: on your own servers or in a sovereign cloud. Both with the same quality.
- On premises, in your own data centre
- Sovereign cloud
- Medical device
- ISO 13485:2016
- EU AI Act
- EU GDPR
- No US CLOUD Act
- ISO/IEC 27001Hosting provider
- BSI C5Hosting provider
Want Haidy® at your institution too?
Klara is the speech recognition inside Haidy®. We'll show you how it works with your reports and your workflows.
- ReportsHaidy® Reporter
- DictationHaidy® Dictate
- WorkflowsHaidy® Hive