metadata

dataset_info:
  features:
    - name: transcript_segment
      dtype: string
    - name: prior_context
      dtype: string
    - name: prior_speakers
      dtype: string
    - name: speaker
      dtype: string
    - name: change
      dtype: int64
  splits:
    - name: test
      num_bytes: 357571
      num_examples: 294
  download_size: 179841
  dataset_size: 357571
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
license: mit
task_categories:
  - text-generation
  - question-answering
language:
  - en
tags:
  - finance

Transcript Speaker Identification Dataset

Description

This dataset is designed to facilitate the development and evaluation of models that identify and assign speakers and speaker changes within event transcripts. It consists of segments from various transcripts where the primary task is to determine who the speaker is, based on the given textual context and a list of possible speakers. The dataset was assembled from three earnings events:

Q4 2023 Amazon.Com Inc Earnings Call
Q2 2024 Apple Inc Earnings Call
Q2 2024 Adobe Inc Earnings Call

Dataset Structure

Columns

transcript_segment: A specific segment of the transcript which requires speaker identification.
prior_context: The textual context preceding the transcript_segment, which can be instrumental in identifying the speaker.
possible_speakers: A list of names or identifiers representing individuals who could potentially be the speaker of the segment.
speaker: The actual speaker of the transcript_segment
change: 1 or 0 indicating whether the speaker has changed

Data Format

The dataset is presented in a tabular format, where each row corresponds to a data point that includes a transcript segment, its prior context, the possible speakers, the label indicating the actual speaker, and an indicator of whether the speaker has changed between the prior context and the transcript segment.

Use Cases

This dataset can be used for a variety of applications, including:

Training machine learning models for speaker identification in transcripts.
Enhancing speech recognition systems by improving their ability to attribute text to the correct speaker.
Developing tools for automated meeting summarization where speaker labels are essential.

Accessing the Dataset

You can access this dataset via the HuggingFace Datasets library using the following Python code:

from datasets import load_dataset

dataset = load_dataset("Aiera/aiera-speaker-assign")

A guide for evaluating using EleutherAI's lm-evaluation-harness is available on github.