README.md

metadata

license: apache-2.0
task_categories:
  - automatic-speech-recognition
language:
  - en
tags:
  - audio
  - speech
  - asr
  - robustness
  - conversational
  - multi-talker
  - overlap
  - noise
  - far-field
pretty_name: Mega-ASR Conversational Overlap
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: far_field
        path: data/far_field.parquet
      - split: far_field_noise
        path: data/far_field_noise.parquet
      - split: noise
        path: data/noise.parquet
      - split: obstructed_noise
        path: data/obstructed_noise.parquet
      - split: recording_noise
        path: data/recording_noise.parquet

Mega-ASR Conversational Overlap

Mega-ASR Conversational Overlap is a deterministic English ASR diagnostic set derived from AirCaps/mega-asr-noise-a5sv2, which in turn is sampled from the Mega-ASR training corpus zhifeixie/Voices-in-the-Wild-2M.

The existing AirCaps dataset evaluates single-utterance acoustic robustness. This companion dataset evaluates a different failure mode: two-turn conversational continuity with slight overlap and unequal turn loudness. It does not replace the single-utterance set, and results should identify which benchmark was used.

We used Mega-ASR-Train-derived material because, in our experiments, the standard Mega-ASR test set was not acoustically challenging enough to clearly distinguish robust ASR systems. This is a fixed diagnostic suite and should not be mixed into training when reporting results on it.

Contents

The dataset contains 625 conversations / 3.8351 hours. Each split contains 125 conversations made by pairing all 250 utterances from the corresponding single-utterance condition. Every source utterance is used exactly once within its condition; there are no repeats.

Split	Conversations	Hours
far_field	125	0.6938
far_field_noise	125	1.0413
noise	125	0.7703
obstructed_noise	125	0.6831
recording_noise	125	0.6467

All output audio is mono 16 kHz PCM16 WAV embedded in Parquet.

Construction

• Pairing occurs only within the same acoustic condition.
• Each conversation contains two distinct source utterances in A-then-B order.
• Turn gains are sampled independently from Uniform[-8, 0) dB.
• Leading and trailing padding are independently sampled from 250 to 1,000 ms.
• All conversations contain slight overlap:
  • 42 per split at 50-100 ms;
  • 42 per split at >100-200 ms;
  • 41 per split at >200-300 ms.
• No new noise is added; noise/reverberation/obstruction is inherited from the source condition.
• Scene-wide attenuation is applied only when required to remain below -1 dBFS.
• The reference is the original Mega-ASR text for turn A followed by the original Mega-ASR text for turn B. These are not silver labels.

The source manifests do not expose reliable speaker identities. Therefore each scene is guaranteed to use two distinct utterances, but is not guaranteed to use two distinct speakers.

Saved predictions and benchmark results

Raw predictions are provided for matched error analysis; they are not labels. All WER values below use corpus-level aggregation and the English normalizer documented in results/benchmark_summary.json.

System	Overall WER
A5S	21.02%
ElevenLabs Scribe v2 Realtime	23.52%
Kyutai STT 1B EN/FR	57.41%

Condition	N	A5S	ElevenLabs	Kyutai
far_field	125	6.06%	5.51%	20.70%
far_field_noise	125	32.61%	33.74%	86.24%
noise	125	22.14%	28.68%	54.46%
obstructed_noise	125	24.97%	29.53%	66.22%
recording_noise	125	9.53%	10.14%	37.54%

A5S used cache-aware greedy streaming at exactly 560 ms, BF16, and batch size 1. ElevenLabs used Scribe v2 Realtime with forced English, one independent WebSocket session per conversation, mono PCM16 at 16 kHz, and 100 ms chunks paced in wall-clock real time. Kyutai used the official causal Mimi-to-LMGen transcript path with 80 ms audio frames, a configured 0.5 s text delay, BF16, temperature 0, batch size 1, and fresh streaming state for every conversation. The pinned PyTorch checkpoint exposes no semantic-VAD extra heads, so no transcript-gating VAD was active. Kyutai returned a blank transcript for 24.48% of conversations; these count as deletions in WER.

Each row contains raw predictions and utterance-level S/D/I counts for all three systems. Aggregate results live under results/.

Loading

from datasets import load_dataset

dataset = load_dataset("AirCaps/mega-asr-conversational-overlap")
example = dataset["noise"][0]
print(example["text"])
print(example["a5s_prediction"])
print(example["kyutai_prediction"])

Recent versions of datasets may require torchcodec for decoded audio. Use cast_column("audio", Audio(decode=False)) if only the encoded bytes and metadata are needed.

Provenance and limitations

• Immediate source: AirCaps/mega-asr-noise-a5sv2
• Upstream source: zhifeixie/Voices-in-the-Wild-2M
• Pinned upstream revision: a8a35d3319737190d6fd3d39157b258eaab35980
• Construction date: 2026-08-24
• Source usage: 1,250 / 1,250, with zero within-condition repetitions
• Final audit: 625 decoded WAVs, 1,875 matched predictions, zero missing or corrupt rows

Because this is derived from a training corpus rather than a speaker-disjoint held-out benchmark, it should be treated as a fixed diagnostic suite. Do not train on it and then report evaluation results on the same rows.

License and attribution

The upstream dataset is released under Apache-2.0. Users should review both the upstream dataset card and the Mega-ASR paper for source provenance and limitations.

@misc{xie2026megaasrinthewild2speechrecognition,
  title={Mega-ASR: Towards In-the-wild^2 Speech Recognition via Scaling up Real-world Acoustic Simulation},
  author={Zhifei Xie and Kaiyu Pang and Haobin Zhang and Deheng Ye and Xiaobin Hu and Shuicheng Yan and Chunyan Miao},
  year={2026},
  eprint={2605.19833},
  archivePrefix={arXiv},
  primaryClass={cs.SD},
  url={https://arxiv.org/abs/2605.19833}
}