README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: id
      dtype: string
    - name: transcription
      dtype: string
    - name: transcript_edit
      dtype: string
    - name: evaluated_language
      dtype: string
    - name: environment
      dtype: string
    - name: device
      dtype: string
    - name: domain
      dtype: string
    - name: accuracy
      dtype: float64
    - name: orthography
      dtype: string
    - name: orthography_issues
      dtype: string
    - name: meaning
      dtype: float64
    - name: meaning_loss
      dtype: string
    - name: errors
      dtype: string
    - name: error_examples
      dtype: string
    - name: performance
      dtype: string
    - name: audio_url
      dtype: string
    - name: created_at
      dtype: string
    - name: user_id
      dtype: string
    - name: duration
      dtype: float64
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
  splits:
    - name: train
      num_bytes: 1258868651.29
      num_examples: 1154
  download_size: 1081849693
  dataset_size: 1258868651.29
license: apache-2.0
task_categories:
  - automatic-speech-recognition
language:
  - af
  - am
  - ha
  - ig
  - ln
  - lg
  - om
  - sn
  - sw
  - wo
  - xh
  - yo
  - bem

Dataset Card for asr-africa/ASR_Evaluation_dataset

Dataset Overview

Languages Covered: Afrikaans, Amharic, Bemba, Hausa, Igbo, Kinyarwanda, Lingala, Luganda, Oromo, Swahili, Wolof, Xhosa, Yoruba

Source: Transcriptions evaluated by native/advanced speakers of the languages.

Dataset Description

This dataset provides human evaluations of automatic speech recognition (ASR) outputs across 13 African languages. For each audio sample, it includes the model-generated transcription, human-corrected transcript (if any), numeric quality ratings, and structured feedback.

Dataset Structure

Field	Description
transcription	ASR model prediction
transcript_edit	Evaluator’s correction (if any)
accuracy	1–5 rating of transcription accuracy
meaning	1–5 rating of preserved meaning
orthography	Quality of transcription use the standard orthography (including accents, diacritics, special characters)
errors	Types of transcription errors
performance	Free-text performance summary
environment	Recording environment
device	Recording device used
domain	Topic/domain of the speech
duration	Audio duration in seconds
audio_url	File path

Dataset Statistics

Duration Summary (All Languages)

• Mean: 14.73 seconds
• Median: 9.59 seconds
• Min: 1.5 seconds
• Max: 90.66 seconds
• Std Dev: 13.27 seconds

Language Breakdown by Duration (Seconds)

Language	Count	Mean	Min	Max	Median	Std Dev
Afrikaans	116	24.27	3.54	63.18	22.71	14.53
Amharic	132	37.99	13.62	90.66	34.07	14.62
Bemba	116	16.92	7.02	29.88	16.53	5.20
Hausa	60	8.21	2.58	18.54	7.23	4.16
Igbo	55	30.42	29.38	31.77	30.40	0.49
Kinyarwanda	103	12.23	7.40	20.60	11.93	2.50
Lingala	30	7.90	5.36	11.99	7.65	1.74
Luganda	110	6.84	2.03	11.82	6.81	1.66
Oromo	116	5.10	1.50	27.18	2.40	5.91
Swahili	132	6.16	2.14	12.96	6.00	1.98
Wolof	53	9.44	4.86	16.26	9.96	2.37
Xhosa	8	15.44	3.14	33.54	6.40	14.73
Yoruba	122	6.11	2.16	11.24	5.94	2.17

Accuracy and Meaning Ratings (Per Language)

Language	Avg Accuracy	Avg Meaning
Swahili	4.96	4.97
Luganda	4.70	4.78
Amharic	4.65	4.82
Lingala	4.63	4.70
Hausa	4.58	4.97
Oromo	4.52	4.50
Bemba	4.39	4.86
Yoruba	4.22	4.48
Wolof	3.98	4.13
Kinyarwanda	3.75	4.81
Xhosa	3.62	3.38
Afrikaans	3.59	4.10
Igbo	2.25	2.15

Device Used Per Language

Language	Dedicated Mic	Laptop Mic	Mobile/Tablet
Afrikaans	51	59	5
Amharic	1	73	55
Bemba	0	2	114
Hausa	0	4	56
Igbo	0	0	55
Kinyarwanda	2	2	99
Lingala	0	1	29
Luganda	0	82	27
Oromo	0	115	0
Swahili	0	65	65
Wolof	4	6	43
Xhosa	0	3	5
Yoruba	0	28	91

Recording Environment Used Per Language

Language	Noisy Background	Quiet Room	Studio
Afrikaans	14	102	0
Amharic	2	117	10
Bemba	0	61	55
Hausa	0	60	0
Igbo	0	55	0
Kinyarwanda	0	103	0
Lingala	0	30	0
Luganda	7	102	0
Oromo	0	115	0
Swahili	5	118	5
Wolof	15	38	0
Xhosa	1	7	0
Yoruba	65	56	0