README.md

metadata

license: cc-by-4.0
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: duration
      dtype: float64
    - name: reference
      dtype: string
    - name: RobotsMali/stt-bm-quartznet15x5-v0
      dtype: string
    - name: RobotsMali/stt-bm-quartznet15x5-v1
      dtype: string
    - name: RobotsMali/soloba-ctc-0.6b-v0
      dtype: string
    - name: RobotsMali/soloba-ctc-0.6b-v1
      dtype: string
    - name: RobotsMali/soloni-114m-tdt-ctc-v0
      dtype: string
    - name: RobotsMali/soloni-114m-tdt-ctc-v1
      dtype: string
    - name: RobotsMali/stt-bm-quartznet15x5-v2
      dtype: string
    - name: soloni-114m-tdt-ctc-v2
      dtype: string
  splits:
    - name: test
      num_bytes: 20690456
      num_examples: 45
  download_size: 19975985
  dataset_size: 20690456
task_categories:
  - automatic-speech-recognition
language:
  - bm
tags:
  - speech
  - asr
  - bambara
  - low-resource

Nyana-Eval Dataset

Dataset Description

Nyana-Eval is a compact, stratified evaluation subset for benchmarking Automatic Speech Recognition (ASR) models in Bambara. It consists of 45 audio recordings totaling approximately 3.03 minutes, carefully selected to represent real-world linguistic and acoustic challenges in low-resource Bambara speech. This dataset is derived from the larger RobotsMali/Bam_ASR_Eval_500 corpus and is optimized for quick, reproducible human evaluation.

Nyana-Eval is ideal for:

• Rapid evaluation of Bambara ASR models (e.g., WER/CER computation on diverse conditions).
• Human-assisted qualitative analysis (e.g., semantic fidelity, code-switching handling).
• Testing models on low-resource settings gaps: dialectal variations, noise, proper names, and code-mixing with French.

Key Statistics:

• Total Samples: 45 (balanced: 15 per source subset).
• Total Duration: ~3.03 minutes (average ~4.0 seconds per sample).
• Audio Format: Mono-channel WAV files at 16 or 44.1k kHz sampling rate.
• Languages: Primary: Bambara (Bamana); Secondary: French code-switching (~15% of samples).
• License: CC-BY-4.0 License (open for research, commercial use with attribution).

Compiled by Robots Mali AI4D Lab, this dataset powers the human-comparative analysis in the [Bambara ASR Models Evaluation Report].

Dataset Structure

Nyana-Eval is a single-split dataset (default: test), with each entry including raw audio, duration, transcriptions (reference) and models transcriptions.

Features/Columns

Column	Type	Description	Example Value
audio	Audio	Raw audio waveform (array + sampling rate: 16 or 44.1k kHz) or file path.	{"path": "1.1.wav", "array": [...], "sampling_rate": 16000}
duration	Float64	Length of the audio clip in seconds (range: 0.62s – 15s).	3.45
references	String	Bambara text	"nɔgɔ ye a ka tɔɔrɔ ye"
'8 * models transcriptions'	String	ASR provised transcriptions

Splits

• Default Split: Full 45 samples (test for evaluation).
• Subsets by Source: Balanced 15 samples each from the three parent subsets (see Sources below).

To load in Python (via Hugging Face Datasets):

from datasets import load_dataset
dataset = load_dataset("RobotsMali/nyana-eval", split="test")
print(len(dataset))  # Output: 45
print(dataset[0])    # Example: {'audio': ..., 'duration': 3.45, 'transcription': 'adama dusukasilen ye a sigi'}

Sources and Compilation

Nyana-Eval is a balanced subsample (15 per subset) from the full 500-sample RobotsMali/Bam_ASR_Eval_500 corpus (~36.69 minutes total). Selection criteria ensured diversity: voice variety (age/gender/accents), acoustic challenges (noise/volume/overlaps), and linguistic phenomena (code-switching, proper names etc.)

Parent Subsets Breakdown (15 samples each in Nyana-Eval):

• Ref. 1: RobotsMali/kunkado (Hugging Face) – 15 audios (~1.96 minutes scaled).
  Semi-supervised interviews and spontaneous discourse. Source: RobotsMali/kunkado. Focus: Dialectal variations and natural flow.

• Ref. 2: jeli-ASR street interviews subset – 30 audios (~1.85 minutes).
  Street interviews Subset from the jeli-asr project. Source: jeli-asr

• Ref. 3: Readings of Excerpts from An Bɛ Kalan app (RobotsMali) – 220 audios (~20.06 minutes).
  User-generated readings and interactions from the mobile app for Bambara learning, captures learner speech with occasional errors or pauses. source: RobotsMali-AI/an-be-kalan

Metadata

General Metadata

• Creator: Robots Mali AI4D Lab
• Version: 1.0 (November 2025).
• Creation Date: Derived November 2025 from Bam_ASR_Eval_500.
• Update Frequency: Static (expansions via parent dataset).
• Download Size: ~25 MB (audios + metadata).
• Ethical Notes: Ethically sourced/anonymized; focuses on public-domain cultural speech. For research; cite Robots Mali.

Challenges Represented:

• Code-switching: samples (e.g., "Segou ville").
• Proper names (e.g., "Sunjata," "Traoré").
• Noise/Overlaps: (e.g., low-volume interviews, multi-speaker).

Related Resources

• Parent Dataset: RobotsMali/Bam_ASR_Eval_500 (full 500 samples).
• Models: Test with RobotsMali ASR models
• App: Collect similar data via An Bɛ Kalan.

This README is self-contained; explore the attached report PDF for detailed human annotations and model rankings on these exact 45 samples!

Citation

@dataset{robotsmali_nyana_eval_2025,
  author       = {RobotsMali AI4D Lab},
  title        = {Nyana-Eval: 45-sample Human-Evaluated Bambara ASR Test Set},
  year         = {2025},
  url          = {https://huggingface.co/datasets/RobotsMali/nyana-eval},
  note         = {Stratified subset of Bam_ASR_Eva_500 used for human + WER evaluation}
}