Dataset
EarningsCallTranscript
A curated collection of earnings call recordings broken into intelligent segments and paired with high-accuracy transcripts produced by Mistral's Voxtral model, intended for ASR benchmarking, transcription quality studies, and audio-to-text alignment research.
Data preview
Dataset card
EarningsCallTranscript is a Hugging Face dataset published by finosfoundation that pairs segmented earnings-call audio with machine-generated transcripts. It is distributed as parquet and is sized at fewer than 1,000 examples.
What this dataset contains
The dataset contains 153 rows in its train split, each representing an audio segment of an earnings call together with a corresponding transcript produced by Mistral's Voxtral model. Segments are created through voice activity detection and sentence boundary analysis, with target segment lengths between 3 and 30 seconds. Metadata also captures the source audio file and the transcription model used.
Structure and fields
Each row carries an audio object alongside string fields such as wav_filename, transcript, original_filename, transcription_model, transcription_timestamp, and segment_position. Numeric fields include wav_filesize, segment_id, start_time, end_time, duration_seconds, actual_duration, original_start_time, original_end_time, prompt_tokens, completion_tokens, total_tokens, transcript_length, transcript_word_count, and tokens_per_second. A boolean has_context_overlap field indicates whether the segment was extended to preserve surrounding context.
How analysts use it
The dataset is primarily intended for evaluating automatic speech recognition systems, benchmarking transcription quality, and aligning audio with text. In an investment workflow it can support building ASR pipelines or testing speech-to-text accuracy on financial speech, though the modest row count limits coverage of companies and time periods. Analyst-relevant text content sits in the transcript field, while timing columns such as start_time, end_time, and segment_position allow ordering of statements within a call.
Documentation & references 1
Related datasets
Browse moreA worldwide archive of insider trading activity that captures roughly three thousand reported transactions on a typical day.
Structured Form 144 filings used to track insider transactions and inform quantitative trading approaches.
Quarterly conference calls are processed to detect revisions in guidance and shifts in executive language.
A sharded Parquet archive of NSE equities and index prices from India spanning 2000 to 2026, covering more than 2,500 tickers and organized into roughly 1.5 GB files for efficient streaming.
A private collection of cleaned earnings-call transcript segments derived from a Motley Fool Kaggle release, split into 135,306 training rows tagged with ticker, exchange, date, and chunk text.
Curated analyst recommendations and updated price targets for Nasdaq 100 constituents, spanning roughly two decades.
Discussion
Questions or observations about this dataset? Join the conversation on X.
No comments yet. Start the conversation on X.