Earnings22-Cleaned-AA-chunked

Quick links: AA Streaming Speech to Text Leaderboard | Speech to Text methodology

Earnings22-Cleaned-AA-chunked is a chunked version of Earnings22-Cleaned-AA, the cleaned Earnings-22 subset used by Artificial Analysis for streaming Speech to Text evaluation.

The original Earnings-22 data comes from esb/datasets, a corpus of corporate earnings calls. Artificial Analysis manually reviewed and corrected the reference transcripts in the cleaned subset to reduce ground-truth errors that can unfairly affect word error rate (WER).

Overview

Methodology

Some streaming STT systems and hosted APIs work best, or only work, on shorter audio segments. To support these systems while preserving comparability with non-streaming AA-WER figures as much as possible, we split the cleaned Earnings22 audio into shorter chunks.

We used Silero VAD to identify speech start and end timestamps in each parent call. Adjacent timestamped segments were combined into chunks up to, but not exceeding, 30 seconds.

Each row contains the parent sample ID, chunk index, start time, end time, duration, and audio file name. This chunked dataset does not include per-chunk reference transcripts; reference text for scoring is derived from the cleaned parent transcript and aligned during the Artificial Analysis evaluation pipeline.

Usage

from datasets import load_dataset

dataset = load_dataset("ArtificialAnalysis/Earnings22-Cleaned-AA-chunked", split="test")

Audio files are included in this repository under audio/. For each row, the audio path is audio/{file_name}.

from huggingface_hub import hf_hub_download

audio_path = hf_hub_download(
    repo_id="ArtificialAnalysis/Earnings22-Cleaned-AA-chunked",
    repo_type="dataset",
    filename=f"audio/{dataset[0]['file_name']}",
)

Resources

• AA Streaming Speech to Text Leaderboard
• Benchmarking methodology
• Earnings22-Cleaned-AA
• Original ESB datasets

Versioning

Current version: 1.0
Dataset file: earnings22_cleaned_aa_chunked_v1.jsonl

License

This dataset is released under Apache-2.0. For upstream terms, see esb/datasets.