README.md

---
license: cc-by-sa-4.0
language:
- en
task_categories:
- question-answering
size_categories:
- 10K<n<100K
tags:
- multi-hop-qa
- retrieval-augmented-generation
- speech-recognition
- asr-robustness
- accented-speech
configs:
- config_name: hotpotqa
  data_files:
  - split: test
    path: hotpotqa.jsonl
- config_name: 2wikimultihopqa
  data_files:
  - split: test
    path: 2wikimultihopqa.jsonl
- config_name: musique
  data_files:
  - split: test
    path: musique.jsonl
---

# Spoken Multi-hop QA: ASR Transcripts Across Four English Accents

ASR transcriptions of 3,000 multi-hop QA questions, each spoken in four
English accents and transcribed with Whisper-large-v3. Released as the
data companion to [*Better Retrieval, Worse Robustness: How Multi-hop RAG
Amplifies Upstream ASR Errors*](https://arxiv.org/abs/2608.22872)
(EMNLP 2026, Main Conference).

The dataset exists to make one thing cheap to study: what happens to a
retrieval pipeline when its query arrives through ASR rather than as clean
text. Every row pairs a gold question (the clean-text oracle input) with
the transcription actually produced for one accent, so the degradation is
directly measurable without re-running any speech model.

**Code:** https://github.com/Continuum-AI-Corp/spoken-multihop-rag

## Contents

| Config | Questions | Rows | Question types |
|---|---|---|---|
| `hotpotqa` | 1,000 | 4,000 | bridge, comparison |
| `2wikimultihopqa` | 1,000 | 4,000 | + compositional, inference |
| `musique` | 1,000 | 4,000 | 2-, 3-, 4-hop compositions |

One row per (question, accent). Questions are sampled uniformly at random
from each benchmark's official validation set with seed `42`.

## Schema

| Field | Type | Description |
|---|---|---|
| `id` | string | Question id from the source benchmark |
| `benchmark` | string | `hotpotqa`, `2wikimultihopqa`, `musique` |
| `accent` | string | `us`, `in`, `ph`, `ng` |
| `voice` | string | Edge TTS voice used to synthesize this row |
| `question` | string | Gold question text, i.e. the clean-text oracle input |
| `answer` | string | Gold answer from the source benchmark |
| `transcription` | string | Whisper-large-v3 top-1 output. `null` when `tts_failed` |
| `wer` | float | Word error rate of `transcription` against `question`. `null` when `tts_failed` |
| `tts_failed` | bool | `true` for four rows whose speech synthesis produced an empty file |

`wer` uses unit-cost Levenshtein alignment over lowercased,
whitespace-tokenized words, with no punctuation stripping. This matches the
implementation used for every number in the paper
(`evaluation/core/metrics.py`).

## Mean WER

| Config | US | IN | PH | NG |
|---|---|---|---|---|
| `hotpotqa` | 9.4% | 11.0% | 10.6% | 14.5% |
| `2wikimultihopqa` | 13.6% | 14.1% | 13.8% | 17.1% |
| `musique` | 5.2% | 5.5% | 5.7% | 8.0% |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("orcarouter/spoken-multihop-rag", "2wikimultihopqa", split="test")

# clean-text vs ASR input for one accent
ng = ds.filter(lambda r: r["accent"] == "ng" and not r["tts_failed"])
print(ng[0]["question"])       # oracle input
print(ng[0]["transcription"])  # what the retriever actually receives
print(ng[0]["wer"])
```

## How it was built

**Speech synthesis.** Each question was synthesized with Microsoft Edge TTS
at 24 kHz, one voice per accent:

| Accent | Voice | Role |
|---|---|---|
| US | `en-US-JennyNeural` | High-resource baseline |
| IN | `en-IN-NeerjaNeural` | Non-native (Indian English) |
| PH | `en-PH-RosaNeural` | Non-native (Filipino English) |
| NG | `en-NG-EzinneNeural` | Low-resource (Nigerian English) |

**Transcription.** Whisper-large-v3, English forced, without timestamps.

## Limitations

Read these before using the dataset as an ASR benchmark.

- **No audio.** Only transcripts are released. The synthesized speech is
  Microsoft Edge TTS output and is not redistributed here. Audio can be
  regenerated with `data/build_dataset.py` in the code repository, though
  the service is not version-pinned, so regenerated audio and any
  transcripts derived from it will not match these byte for byte.
- **One voice per accent.** Accent is varied, speaker is not. Results
  should be read as a controlled probe of accent-induced ASR error, not as
  an estimate of population-level performance for any speaker group.
- **Synthetic speech understates real difficulty.** On a public corpus of
  real Nigerian-accented English, Whisper-large-v3 reaches 28.9% mean WER,
  against 17.1% / 14.5% / 7.9% for the synthesized Nigerian condition on
  2WikiMultiHopQA / HotpotQA / MuSiQue. The synthesized setting is easier
  by a factor of roughly 1.7x to 3.7x.
- **Four rows are unusable.** Speech synthesis returned an empty file for
  the four (question, accent) pairs listed below, so no transcription
  exists for them. They carry `tts_failed: true` with null transcription
  and WER, which keeps them out of aggregate statistics.

  | Config | Accent | id |
  |---|---|---|
  | `hotpotqa` | ng | `5a8603f655429960ec39b601` |
  | `hotpotqa` | ph | `5ade063d5542997dc7907129` |
  | `2wikimultihopqa` | ng | `69a429bb097911ebbdb0ac1f6bf848b6` |
  | `musique` | ng | `3hop2__326964_36852_7713` |

- **English only.** No claim is made about other languages.

## Licensing and attribution

This dataset is released under **CC BY-SA 4.0**, inherited from the most
restrictive upstream source. Each row's `question` and `answer` are
redistributed from one of three benchmarks:

| Source | License | Reference |
|---|---|---|
| HotpotQA | CC BY-SA 4.0 | [hotpotqa.github.io](https://hotpotqa.github.io/) |
| 2WikiMultiHopQA | Apache-2.0 | [Alab-NII/2wikimultihop](https://github.com/Alab-NII/2wikimultihop) |
| MuSiQue | CC BY 4.0 | [StonyBrookNLP/musique](https://github.com/StonyBrookNLP/musique) |

HotpotQA's ShareAlike term propagates to derivative works, which is why the
whole dataset carries CC BY-SA 4.0 rather than a per-config license. The
`transcription` field holds Whisper-large-v3 output; Whisper is
MIT-licensed. Accompanying code is Apache-2.0.

If you redistribute or build on this dataset, keep the attributions above
and license derivatives compatibly.

## Citation

```bibtex
@inproceedings{bao2026better,
  title     = {Better Retrieval, Worse Robustness: How Multi-hop {RAG} Amplifies Upstream {ASR} Errors},
  author    = {Bao, Zhenghua},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year      = {2026},
  note      = {To appear},
  url       = {https://arxiv.org/abs/2608.22872}
}
```

Please also cite the source benchmark you use, and Whisper
(Radford et al., 2023) for the transcriptions.