---
dataset_info:
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 16000
  - name: id
    dtype: large_string
  - name: text
    dtype: large_string
  - name: speaker_id
    dtype: int64
  - name: language
    dtype: large_string
  - name: gender
    dtype: large_string
  - name: date_of_birth
    dtype: large_string
  - name: occupation
    dtype: large_string
  - name: educational_background
    dtype: large_string
  - name: marital_status
    dtype: large_string
  - name: native_district
    dtype: large_string
  - name: native_state
    dtype: large_string
  - name: current_city
    dtype: large_string
  - name: years_spent_in_current_district
    dtype: int64
  - name: income
    dtype: large_string
  - name: device_manufacturer
    dtype: large_string
  - name: device_model
    dtype: large_string
  - name: audio_length_s
    dtype: float64
  splits:
  - name: test
    num_bytes: 1786718598.23
    num_examples: 2102
  download_size: 1569700186
  dataset_size: 1786718598.23
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
pretty_name: Voice Arena Monsoon en-IN (public test)
language:
- en
task_categories:
- automatic-speech-recognition
size_categories:
- 1K<n<10K
tags:
- open-asr-leaderboard
- speech
- asr
- benchmark
- conversational
- spontaneous-speech
- english
- fairness
- bias-evaluation
license: cc-by-4.0
---

# Voice Arena Monsoon en-IN (public test)

Part of the [Open ASR Leaderboard](https://huggingface.co/spaces/hf-audio/open_asr_leaderboard), in the main board's default column set, so it contributes to the headline Average WER for every model listed.

A conversational Indian English ASR test set that records **who was speaking**, not only what
was said. Every clip carries twelve speaker attributes — gender, age, native district
and state, education, occupation, income band, handset — so a difference between two
systems can be traced to a group of speakers instead of disappearing into one number.

This is the **public** split, released for self-scoring. A speaker-disjoint private
split is withheld to limit benchmark-specific optimisation.

## At a glance

| | |
| :--- | :--- |
| Clips | 2,102 |
| Duration | 5.62 h |
| Speakers | 1,444 |
| Clip length | 9.6 s mean / 10.4 s median |
| Male / female clips | 50 / 50 |
| Native districts | 428 |
| States / UTs | 24 states + 6 union territories |
| Distinct handset models | 556 |
| Current cities | 641 |

Small in hours, large in speakers: more than half the speakers contribute exactly one
clip, and the ten largest contributors account for under 7 per cent of total duration.
A result here is an average over hundreds of voices rather than a few talkers recorded
at length.

## Loading

```python
from datasets import load_dataset

ds = load_dataset("VoiceArena/MonsoonASR-Open-ASR-leaderboard-en-IN", split="test")
print(ds[0]["text"], ds[0]["native_district"], ds[0]["gender"])
```

Audio is 16 kHz.

## Scoring

Reference transcripts are plain strings. Score with word error rate after the
[Open ASR Leaderboard](https://huggingface.co/spaces/hf-audio/open_asr_leaderboard) normaliser, and report the **aggregate** over the
split — `sum(ins + del + sub) / sum(reference words)` — not the average of per-clip
rates.

## Fields

| Group | Fields |
| :--- | :--- |
| Segment | `id`, `audio`, `audio_length_s`, `language` |
| Reference | `text` |
| Speaker | `speaker_id`, `gender`, `date_of_birth` |
| Background | `occupation`, `educational_background`, `marital_status`, `income` |
| Geography | `native_district`, `native_state`, `current_city`, `years_spent_in_current_district` |
| Recording | `device_manufacturer`, `device_model` |

## More

Collection, quality control, the transcription protocol and a worked regional
analysis: [https://huggingface.co/blog/open-asr-leaderboard-global-south](https://huggingface.co/blog/open-asr-leaderboard-global-south).