ElevenLabs Officially Launches Scribe v2 Medical… 18% Reduction in Clinical Audio WER

ElevenLabs has officially launched Scribe v2 Medical on ElevenAPI. This model is fine-tuned specifically for clinical audio and reduces the WER (word error rate) of clinical audio by 18% compared to the base Scribe v2.

Benchmark Performance

• Omi Health Benchmark: Recorded the lowest overall WER of 5.88% among 30 systems, with dosage recognition accuracy improved to 86.2%, surpassing the base model (79.8%).
• Eka Medical Benchmark: Achieved a lower WER than all publicly available results.
• Multilingual Performance: Achieved an approximate 36% relative reduction in WER compared to the base model on Corti's MedDictate benchmark, consistently maintained across three languages.

Limitations and Solutions

Recognizing isolated single-word clips without context remains a challenging task. The WER for isolated terms is 14.3%, higher than the in-sentence term WER (7.5%). For example, errors occur where 'etodolac' is misrecognized as "It'll do the luck." ElevenLabs stated that these errors can be reduced by emphasizing drug names or medical phrases through the keyterm prompting feature.

Data Security and Usage

Medical fine-tuning does not degrade non-medical speech performance, maintaining the same 5.3% WER as the base model in CommonVoice tests. For HIPAA compliance, signing a BAA and enabling Zero Retention Mode (ZRM) ensures that audio and text outputs are immediately deleted, preserving no data. Users can specify the scribe_v2_medical model ID in the Speech to Text API and call it via the batch endpoint.