Appen Research

@AppenResearch

Human data for frontier AI. Research and insights from Appen.

Washington, USA

appen.com

Replies Reposts Media

Appen Research
@AppenResearch
Sep 25
A physician dictates "hyponatremia." The model writes "hypernatremia." One prefix. The opposite diagnosis. MedTerm-90 scored 7 speech-to-text engines on 90.5 hours of real dictation. Best recall: @ElevenLabs Scribe v2, 83.35%. Where does yours rank? We'll run it.
00:00
5
Appen Research
@AppenResearch
Sep 21
Heading to @WeAreDevs in San José this week? Appen’s Jeanine Sinanan-Singh, Director of GenAI Research, and Alexi Schwartzkopff, Senior Strategic Account Executive, will be there. Building or evaluating AI? Come say hello and talk training data, evals, benchmarks, and more.
Jeanine S and Alexi Schwartzkopff
Appen Research
@AppenResearch
Sep 16
📊 2,400+ expert hours. 12 weeks. Two parallel fine-tuning tracks. That’s the human preference layer Appen delivered for PANDA Plus - Cohere’s fine-tuning program for its enterprise LLM, Command. Domain-qualified contributors with hands-on LLM experience compared live Command
2
Appen Research
@AppenResearch
Sep 10
New benchmark: MedTerm-90. We tested 7 speech-to-text systems on 90.5 hours of real physician dictation across 14 specialties, using clinical entity recall instead of WER. One finding: on 28% of recordings, no model reached 85% recall. Full benchmark and methodology in the
5
Appen Research
@AppenResearch
Sep 10
New @AppenResearch benchmark: MedTerm-90. 7 speech-to-text systems 1,413 real physician dictations 90.5 hours 14 specialties 30,017 clinician-validated entities We evaluate clinical entity recall rather than WER and manually adjudicate near-miss transcriptions. Top-line
2