README.md

metadata

license: mit
dataset_info:
  - config_name: default
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits: null
    download_size: 0
    dataset_size: 0
  - config_name: noisy-bg-snr-10
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: noisy-bg-snr-30
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: noisy-bg-snr-50
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: denoised-bg-snr-10
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: denoised-bg-snr-30
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: denoised-bg-snr-50
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: noisy-short-snr-10
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: noisy-short-snr-30
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: noisy-short-snr-50
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: denoised-short-snr-10
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: denoised-short-snr-30
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: denoised-short-snr-50
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: noisy-gauss-amp-0.001
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: noisy-gauss-amp-0.009
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: noisy-gauss-amp-0.017
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: denoised-gauss-amp-0.001
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: denoised-gauss-amp-0.009
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: denoised-gauss-amp-0.017
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: original
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
  - config_name: denoised-original
    features:
      - name: md5_text
        dtype: string
      - name: file_name
        dtype: string
      - name: md5_audio
        dtype: string
      - name: duration
        dtype: float32
      - name: text
        dtype: string
      - name: audio_language
        dtype: string
      - name: text_language
        dtype: string
      - name: session_id
        dtype: string
      - name: speaker
        dtype: string
      - name: type_concept
        dtype: string
      - name: recording_context
        dtype: string
      - name: medical_entities
        dtype: string
      - name: filename
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: test
        num_bytes: 250000000
        num_examples: 500
    download_size: 212500000
    dataset_size: 250000000
configs:
  - config_name: noisy-bg-snr-10
    data_files:
      - split: test
        path: noisy-bg-snr-10/test-*
  - config_name: noisy-bg-snr-30
    data_files:
      - split: test
        path: noisy-bg-snr-30/test-*
  - config_name: noisy-bg-snr-50
    data_files:
      - split: test
        path: noisy-bg-snr-50/test-*
  - config_name: denoised-bg-snr-10
    data_files:
      - split: test
        path: denoised-bg-snr-10/test-*
  - config_name: denoised-bg-snr-30
    data_files:
      - split: test
        path: denoised-bg-snr-30/test-*
  - config_name: denoised-bg-snr-50
    data_files:
      - split: test
        path: denoised-bg-snr-50/test-*
  - config_name: noisy-short-snr-10
    data_files:
      - split: test
        path: noisy-short-snr-10/test-*
  - config_name: noisy-short-snr-30
    data_files:
      - split: test
        path: noisy-short-snr-30/test-*
  - config_name: noisy-short-snr-50
    data_files:
      - split: test
        path: noisy-short-snr-50/test-*
  - config_name: denoised-short-snr-10
    data_files:
      - split: test
        path: denoised-short-snr-10/test-*
  - config_name: denoised-short-snr-30
    data_files:
      - split: test
        path: denoised-short-snr-30/test-*
  - config_name: denoised-short-snr-50
    data_files:
      - split: test
        path: denoised-short-snr-50/test-*
  - config_name: noisy-gauss-amp-0.001
    data_files:
      - split: test
        path: noisy-gauss-amp-0.001/test-*
  - config_name: noisy-gauss-amp-0.009
    data_files:
      - split: test
        path: noisy-gauss-amp-0.009/test-*
  - config_name: noisy-gauss-amp-0.017
    data_files:
      - split: test
        path: noisy-gauss-amp-0.017/test-*
  - config_name: denoised-gauss-amp-0.001
    data_files:
      - split: test
        path: denoised-gauss-amp-0.001/test-*
  - config_name: denoised-gauss-amp-0.009
    data_files:
      - split: test
        path: denoised-gauss-amp-0.009/test-*
  - config_name: denoised-gauss-amp-0.017
    data_files:
      - split: test
        path: denoised-gauss-amp-0.017/test-*
  - config_name: original
    data_files:
      - split: test
        path: original/test-*
  - config_name: denoised-original
    data_files:
      - split: test
        path: denoised-original/test-*
task_categories:
  - text-to-speech
language:
  - en
tags:
  - dataset
  - audio
  - speech
  - asr
  - denoising
pretty_name: Speech Brain Noise Evaluation Dataset
size_categories:
  - 10K<n<100K

Denoising Impact Evaluation Dataset

Dataset Description

The ekacare/denoising-impact-evaluation-dataset is a comprehensive benchmark dataset designed to evaluate the effects of speech enhancement on automatic speech recognition (ASR) systems in medical speech contexts. It includes paired noisy and denoised audio subsets under controlled acoustic conditions to support systematic analysis of denoising performance.

Source Data

Base Dataset: ekacare/eka-medical-asr-evaluation-dataset

Selection Criteria: 500 English audio samples with duration >= 15 seconds

Domain: Medical speech containing clinical terminology and pharmaceutical names

Language: English

Noise Augmentation Pipeline

The Noisy subsets in this dataset were synthetically augmented using the audiomentations library with three distinct noise types at varying intensity levels:

Noise Types & Configurations:

Background Noise

SNR levels: 10 dB, 30 dB, 50 dB

Short Noise

SNR levels: 10 dB, 30 dB, 50 dB

Gaussian Noise

Amplitude levels: 0.001, 0.009, 0.017

Denoising

All audio samples underwent enhancement using SpeechBrain's SpectralMaskEnhancement framework with the pre-trained MetricGAN-plus-voicebank model

Dataset Structure

Subsets

Each subset in the dataset follows a structured naming convention that encodes both the type of acoustic corruption applied and its corresponding denoised counterpart. All noisy and denoised subsets are derived from the same underlying set of 500 original medical speech recordings, ensuring one-to-one correspondence across conditions.

Noisy Subsets

Subsets prefixed with noisy- are created by applying synthetic noise to the original clean audio using the audiomentations library. The subset name specifies:

• the noise type (bg for background noise, short for short noise, gauss for Gaussian noise), and
• the noise severity, expressed either as signal-to-noise ratio (SNR, in dB) or noise amplitude.

For example:

• noisy-bg-snr-10 contains audio corrupted with continuous background noise at 10 dB SNR.
• noisy-short-snr-30 contains audio with short noise events at 30 dB SNR.
• noisy-gauss-amp-0.009 contains audio with additive Gaussian noise of amplitude 0.009.

Denoised Subsets

Subsets prefixed with denoised- are generated by applying the same speech enhancement model to their corresponding noisy subsets. Specifically, each denoised subset is obtained by processing the paired noisy audio using SpeechBrain’s SpectralMaskEnhancement framework with the pre-trained MetricGAN-plus-voicebank model.

For example:

• denoised-bg-snr-10 is produced by denoising the audio in noisy-bg-snr-10.
• denoised-gauss-amp-0.017 corresponds directly to noisy-gauss-amp-0.017.

Original and Denoised-Original

In addition to synthetically corrupted audio, the dataset also includes:

• original: the unmodified clean medical speech recordings.
• denoised-original: the output of applying the same enhancement pipeline directly to the original audio.

This structure enables controlled, paired comparisons between noisy and denoised audio under identical conditions, facilitating systematic evaluation of the impact of speech enhancement on ASR performance.

Hence, we have the following subsets

• noisy-bg-snr-10: 500 samples
  • test: 500 samples
• noisy-bg-snr-30: 500 samples
  • test: 500 samples
• noisy-bg-snr-50: 500 samples
  • test: 500 samples
• denoised-bg-snr-10: 500 samples
  • test: 500 samples
• denoised-bg-snr-30: 500 samples
  • test: 500 samples
• denoised-bg-snr-50: 500 samples
  • test: 500 samples
• noisy-short-snr-10: 500 samples
  • test: 500 samples
• noisy-short-snr-30: 500 samples
  • test: 500 samples
• noisy-short-snr-50: 500 samples
  • test: 500 samples
• denoised-short-snr-10: 500 samples
  • test: 500 samples
• denoised-short-snr-30: 500 samples
  • test: 500 samples
• denoised-short-snr-50: 500 samples
  • test: 500 samples
• noisy-gauss-amp-0.001: 500 samples
  • test: 500 samples
• noisy-gauss-amp-0.009: 500 samples
  • test: 500 samples
• noisy-gauss-amp-0.017: 500 samples
  • test: 500 samples
• denoised-gauss-amp-0.001: 500 samples
  • test: 500 samples
• denoised-gauss-amp-0.009: 500 samples
  • test: 500 samples
• denoised-gauss-amp-0.017: 500 samples
  • test: 500 samples
• original: 500 samples
  • test: 500 samples
• denoised-original: 500 samples
  • test: 500 samples

Usage

Load specific subset and split:

from datasets import load_dataset

# Load specific subset and split
dataset = load_dataset('ekacare/speech-brain-noise-evaluation-dataset', 'noisy-bg-snr-10', split='test')

# Load everything
dataset = load_dataset('ekacare/speech-brain-noise-evaluation-dataset')

Dataset Creation

This dataset was created using StreamableDatasetManager on 2025-10-13T21:45:58.155717.

Data Fields

The dataset includes the following columns:

• md5_text: String data
• file_name: String data
• md5_audio: String data
• duration: Float32 data
• text: String data
• audio_language: String data
• text_language: String data
• session_id: String data
• speaker: String data
• type_concept: String data
• recording_context: String data
• medical_entities: String data
• filename: String data
• audio: Audio data (16kHz sampling rate)

Technical Details

• Total samples: 10,000
• Shard length: 100
• Number of subsets: 20
• Number of splits: 2
• Audio format: 16kHz sampling rate