---
language:
- hy
license: cc-by-4.0
tags:
- automatic-speech-recognition
- armenian
- fleurs
pretty_name: FLEURS Armenian Reference Corrections
size_categories:
- n<1K
dataset_info:
  features:
  - name: recording_id
    dtype: string
  - name: file_name
    dtype: string
  - name: fleurs_id
    dtype: int64
  - name: original_transcription
    dtype: string
  - name: corrected_transcription
    dtype: string
  splits:
  - name: test
    num_examples: 314
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*.parquet
---

# FLEURS Armenian Reference Corrections

This private dataset contains manual reference corrections used by
[ArmBench ASR](https://huggingface.co/spaces/Metric-AI/ArmBench-ASR) for the
Armenian (`hy_am`) test split of [`google/fleurs`](https://huggingface.co/datasets/google/fleurs).

Only rows whose reviewed transcription differs from the original FLEURS
`raw_transcription` are included. Audio is not duplicated; use `recording_id`
or `file_name` to join these rows to the upstream FLEURS test split.

## Fields

- `recording_id`: stable recording identifier, equal to the WAV filename stem.
- `file_name`: upstream FLEURS WAV filename.
- `fleurs_id`: original FLEURS sentence ID. This is not unique per recording.
- `original_transcription`: original upstream `raw_transcription`.
- `corrected_transcription`: manually reviewed reference used by ArmBench ASR.

## Scope

- Upstream dataset: `google/fleurs`
- Configuration: `hy_am`
- Split compared: `test`
- Changed references: 314
- Unchanged reviewed entries excluded: 78

Corrections were checked by exact comparison with the upstream transcription.
This repository is a correction manifest, not a replacement copy of FLEURS.