license: apache-2.0
language:
- en
pretty_name: 'EarningsCallVoice: Core-100'
size_categories:
- n<1K
task_categories:
- automatic-speech-recognition
- text-to-speech
- audio-classification
tags:
- audio
- datasets
- earnings-calls
- paralinguistics
- voice-cloning
- question-answering
configs:
- config_name: default
default: true
data_files:
- split: core
path: hf_viewer/core.parquet
EarningsCallVoice: Core-100
EarningsCallVoice is a benchmark family for studying executive vocal delivery in earnings-call question answering. Core-100 contains 100 manually verified units. Each unit provides:
- an authentic reference clip from the executive's prepared remarks;
- the text of an analyst question;
- an authentic answer clip from the same executive in the Q&A;
- exact reference and answer transcripts;
- cryptographic hashes and technical metadata.
The question is text only. The release contains no cloned or otherwise synthetic speech. No executive names are supplied.
Why Core-100 exists
The benchmark is designed for controlled research on speaker preservation, voice cloning, delivery manipulation, and audio information beyond transcripts. Every admitted unit passed nine human gates covering complete boundaries, single-speaker purity, text-audio agreement, speaker identity agreement, question-answer coherence, and overall usability.
The full eligible pool contained 107 units. The public core was frozen without using delivery labels, model outputs, cloning scores, market outcomes, or downstream experimental results. It retains all 97 accepted original extractions and adds the three eligible boundary repairs with the highest frozen reference-answer Resemblyzer similarity. This tie-break rule was fixed before downstream modeling. The remaining 7 eligible repairs are retained in the provenance ledger as a reserve set.
Loading
from datasets import load_dataset
dataset = load_dataset("gmarti/EarningsCallVoice", split="core")
The Hub-native view embeds both audio columns for streaming and the Dataset
Viewer. Both are 16 kHz, mono, signed 16-bit PCM WAV. The original WAV files,
CSV, JSONL, and provenance ledger remain available under Files and
versions. The immutable, pre-viewer release payload is tagged v1.0.0.
As an independent release-level diagnostic, Whisper large-v3-turbo passed all 200 clips under the declared WER and first/last-edge thresholds. Mean normalized WER was 2.83%. This diagnostic does not replace the human judgments.
Data statement
- Domain: U.S. public-company earnings calls from 2019 to 2021.
- Language: English, including naturally occurring accents.
- Unit count: 100, one unit per call.
- Authentic reference duration: 12.73 seconds on average.
- Authentic answer duration: 17.56 seconds on average.
- Human validation: one expert reviewer, all nine gates required.
- Speaker identity: verified within each unit, not resolved across calls.
- Selection: source-quality-only and outcome-blind.
This is a small, high-precision benchmark. It is not a representative sample of all earnings calls, accents, genders, industries, or recording channels. It is not intended to train a general-purpose speech synthesizer.
Source and license
Core-100 is derived from FinCall-Surprise, released under Apache-2.0. Its ACL
paper states that the released components include call metadata, transcripts,
speaker audio alignment artifacts, slides, and processing code. See NOTICE
for attribution and LICENSE for the license text.
Responsible use
These are voices of real people speaking at public corporate events. See
RESPONSIBLE_USE.md. Do not use the clips for impersonation, authentication
bypass, fraud, harassment, or misleading attribution.
Reproducibility
provenance/selection_manifest.json records the frozen input hashes, inclusion
rule, selected repair IDs, reserve IDs, and selection scores. MANIFEST.sha256
binds every distributed file.
Citation
Please cite the forthcoming EarningsCallVoice paper or dataset record and the
upstream FinCall-Surprise paper. A provisional citation is supplied in
CITATION.cff.