ekacare/eka-medical-asr-evaluation-dataset

Eka Medical ASR Evaluation Dataset

Dataset Overview and Sourcing

The Eka Medical ASR Evaluation Dataset enables comprehensive evaluation of automatic speech recognition systems designed to transcribe medical speech into accurate text—a fundamental component of any medical scribe system. This dataset captures the unique challenges of processing medical terminology, particularly branded drugs, which is specific to the Indian context. The dataset comprises over 3,900+ curated audio recordings featuring medical terminology delivered in various speaking styles, including isolated medical entities, narrated medical sentences, and impromptu medical conversation. The dataset includes approximately 3,600 English recordings and 320 Hindi recordings. We intend to keep improving and growing this dataset for different languages and scenarios.
For details read this blog.

Data Collection and Quality Assurance

All audio recordings capture natural speech patterns, ensuring realistic evaluation scenarios. A significant portion of the dataset originates from EkaCare's internal team members through narrated medical text sessions and recorded EkaScribe demonstration sessions with our internal medical professionals. Additional high-quality content was sourced from speakers across four different medical colleges, providing diverse regional accents and speaking styles representative of India's medical education landscape.

Target Applications

This dataset is valuable for developers building and evaluating voice-enabled healthcare applications, and medical documentation systems that rely on speech-to-text functionality. Healthcare institutions implementing AI-powered scribe solutions will find this dataset essential for evaluating system performance across diverse Indian linguistic contexts.

Usage

Load specific subset and split:

from datasets import load_dataset

# Load specific subset and split
dataset = load_dataset('ekacare/eka-medical-asr-evaluation-dataset', 'en', split='test')

# Load all splits from a subset
dataset = load_dataset('ekacare/eka-medical-asr-evaluation-dataset', 'en')

# Load everything
dataset = load_dataset('ekacare/eka-medical-asr-evaluation-dataset')

Performance Benchmarks (english-only)

The following table shows the performance of various ASR models evaluated on the Eka Medical ASR Evaluation Dataset. All metrics are reported on the test set.

Models	WER	CER	semWER	kwWER
AWS transcribe	0.183	0.074	0.111	0.122
GPT-4o	0.161	0.097	0.116	0.117
Gemini 2.0 Flash	0.175	0.082	0.105	0.101
Gemini 2.5 Flash	0.148	0.055	0.072	0.068
Eleven Labs (Scribe V1)	0.186	0.087	0.102	0.093
Whisper V3 large	0.157	0.056	0.089	0.085
Bhashini ASR	0.199	0.093	0.123	0.114
Parrotlet-a-en-5b	0.109	0.047	0.072	0.062

Parrotlet-a-en-5b is an open-weight model released by EkaCare for english language ASR in medical domain

Evaluation Metrics

• WER (Word Error Rate): Measures the percentage of words that are incorrectly transcribed
• CER (Character Error Rate): Measures the percentage of characters that are incorrectly transcribed
• semWER (Semantic Word Error Rate): Evaluates transcription accuracy considering semantic/phonetic equivalences
• kwWER (Keyword Word Error Rate): Focuses on the accuracy of medical keywords and terminology

For more details on semWER and its importance in medical ASR evaluation:

• Description: Beyond Traditional WER: The Critical Need for Semantic WER in ASR for Indian Healthcare
• Implementation: ASR Semantic Metrics - KARMA OpenMedEvalKit

Dataset Structure

Subsets

This dataset includes the following subsets:

• en: 3,619 samples
  • test: 3,619 samples
• hi: 320 samples
  • test: 320 samples

Data Fields

The dataset includes the following columns:

• md5_text: String md5 of the ground truth text
• file_name: String filename of the audio file
• audio: Audio data (16kHz sampling rate)
• md5_audio: String md5 of the audio file
• duration: Float32 duration of the audio
• text: String Ground truth text
• audio_language: String language of the speech
• text_language: String language in the text (could differ from the speech in case of translation task)
• session_id: String identifier of the session
• speaker: String speaker id
• type_concept: String type of medical concept
• recording_context: String context of the recording (single entity narration, sentence narration or conversation)
• medical_entities: String Offsets of medical entities along with their type

Technical Details

• Total samples: 3,939
• Shard length: 500
• Number of subsets: 2
• Number of splits: 2
• Audio format: 16kHz sampling rate

Contributors / Annotators list

• Dr Anushree Rana
• Dr Rajshree Badami
• Dr Arun Kumar R
• Neha Ramesh Badge
• Dr Kashika Singh
• Dr Rishi Srivathsav
• Dr Arun Kumar
• Dr Sanjana SN

License

This dataset is released under the MIT License, enabling broad use while maintaining attribution requirements.