README.md

---
license: cc-by-4.0
language:
- en
task_categories:
- automatic-speech-recognition
tags:
- speech
- asr
- benchmark
- finance
- earnings-calls
- long-form
- speaker-diarization
pretty_name: Earnings25
size_categories:
- n<1K
configs:
- config_name: segmented
  data_files:
  - split: test
    path: segmented/test-*.parquet
  default: true
- config_name: full
  data_files:
  - split: test
    path: full/test-*.parquet
---

# Earnings25

**A 500-hour speech benchmark for finance — S&P 500 earnings calls with reference
transcripts, industry labels, and named-speaker attribution.**

> **Citation**
>
> Earnings25 is introduced in our [Interspeech 2026 paper](https://arxiv.org/abs/2607.23813),
> which sets out the sampling design, the evaluation protocol, and reference
> baselines for Whisper and Parakeet-TDT. Start there for the full picture.
>
> Jiang, D., Zhou, H., Wadhawan, A., Fahy, B., Ramesh, V., Weisberg, D.,
> Derkachevskiy, D., Sheehan, H., Prasad, S., & Franceschini, M. (2026).
> Earnings25: A Comprehensive 500-Hour Speech Benchmark for Finance.
> *Proc. Interspeech 2026*, Sydney, Australia. arXiv:2607.23813.
> https://arxiv.org/abs/2607.23813
>
> ```bibtex
> @inproceedings{jiang2026earnings25,
>   title     = {Earnings25: A Comprehensive 500-Hour Speech Benchmark for Finance},
>   author    = {Jiang, Denglin and Zhou, Haoran and Wadhawan, Anshul and Fahy, Brendan and
>                Ramesh, Vinay and Weisberg, David and Derkachevskiy, Dmitriy and Sheehan, Helen and
>                Prasad, Srivas and Franceschini, Michele},
>   booktitle = {Proc. Interspeech 2026},
>   year      = {2026},
>   address   = {Sydney, Australia},
>   note      = {Dataset: \url{https://doi.org/10.5281/zenodo.18762167}}
> }
> ```

📄 [Paper (arXiv:2607.23813)](https://arxiv.org/abs/2607.23813) · 💾 [Zenodo DOI 10.5281/zenodo.18762167](https://doi.org/10.5281/zenodo.18762167)

**Denglin Jiang, Haoran Zhou, Anshul Wadhawan, Brendan Fahy, Vinay Ramesh,
David Weisberg, Dmitriy Derkachevskiy, Helen Sheehan, Srivas Prasad,
Michele Franceschini** — created at Bloomberg, Interspeech 2026 (Sydney).

## About the paper

Existing ASR benchmarks are dominated by read speech, broadcast news, and short
utterances. None of them stress what production financial speech systems
actually face: hour-long calls, telephony-grade audio, dense domain jargon, and
rapid turn-taking among operators, executives, and analysts.

Earnings25 was built to close that gap. It pairs **498 hours of complete S&P 500
earnings calls** from Q4 2025 with a **46-hour industry-balanced set** covering
290 distinct industries — one segment each — so error can be attributed to
domain rather than averaged away. Every item carries company, industry, country,
and market-capitalization metadata, and the long-form split adds speaker turns
with named attribution.

The paper reports reproducible baselines for Whisper (base / medium / large-v2)
and Parakeet-TDT under four scoring conditions, and shows that the ranking of
systems is stable across them while absolute error varies sharply by industry.
Parakeet-TDT leads at 10.8% WER on the long-form split; Whisper-large-v2 trails
it by roughly three points despite being the more widely deployed model.

## Quick start

```python
from datasets import load_dataset

# 46 h, one segment per industry, 16 kHz — start here
ds = load_dataset("florencejiang/earnings25", "segmented", split="test")

# 498 h of complete calls, ~58 min each — long-form, with speaker turns
full = load_dataset("florencejiang/earnings25", "full", split="test", streaming=True)

print(ds[0]["text"][:300])
print(ds[0]["industry"], ds[0]["company"], ds[0]["duration_s"])
```

## Why this benchmark

Earnings calls break general-purpose ASR in specific, measurable ways: telephony
compression and variable recording quality, dense financial jargon and acronyms,
rapid speaker transitions between operators, executives, and analysts, accented
English from international executives, and hour-long durations that expose
long-form decoding failures. Aggregate WER hides all of it. Earnings25 ships
industry labels on every item and speaker turns on the long-form split, so the
number can be broken apart.

## The two configs

| | `segmented` | `full` |
|---|---|---|
| Hours | 45.9 | ~498 |
| Items | 290 segments | 514 complete calls |
| Mean duration | 9.5 min | ~58 min |
| Audio | MP3, 16 kHz mono | MP3, original sample rate |
| Industries | 290 — exactly one segment each | 284 |
| Countries | US only | 12 (93.8% US) |
| Period | 2025 | 2025 Q4 |
| Speaker turns | **none** (`segments` is empty) | ~1,000 per call |
| Use it for | industry-stratified WER, fast iteration | long-form decoding, diarization |

Both are **evaluation-only**. There is no train split by design — treat any
training on this data as contamination of the benchmark.

## Fields

| Field | Type | Notes |
|---|---|---|
| `id` | string | item identifier; segmented ids encode the source call and offsets |
| `audio` | Audio | MP3 bytes, decoded on access |
| `text` | string | reference transcript for the whole item |
| `company` | string | issuer name |
| `industry` | string | industry label (290 distinct in `segmented`) |
| `country` | string | issuer domicile |
| `release_date` | string | ISO 8601 call timestamp |
| `market_cap` | float64 | issuer market capitalization at release |
| `duration_s` | float32 | |
| `sample_rate`, `channels` | int32 | as stored |
| `num_speakers` | int32 | count of attributed speakers |
| `speakers` | list | `{speaker_id, name}` — e.g. `{"1", "Operator"}`, `{"2", "Vicente Reynal"}` |
| `segments` | list | `{start, end, text, speaker_id}` — **populated in `full` only** |

### On speaker labels

Speakers are attributed by **name**, not by role. `speaker_id` in `segments`
joins to the `speakers` list, where `"Operator"` is identifiable by name but
executives and analysts are not distinguished from one another. Deriving a
role taxonomy requires joining the participant names against an external
source. Segment text is phrase-level (`"Thank you for standing by,"`), not
word-level.

## Baselines

Numbers below are from the paper. `WER-N` = text-normalized; `nc`/`np` =
case- and punctuation-insensitive. Score against the `text` field; for `full`,
decode long-form rather than truncating to the first 30 s.

**`full` (498 h)**

| Model | WER | WER-N | WER-nc-np | WER-N-nc-np |
|---|---|---|---|---|
| whisper-base | 0.1785 | 0.1707 | 0.1157 | 0.1104 |
| whisper-medium | 0.1440 | 0.1366 | 0.0859 | 0.0815 |
| whisper-large-v2 | 0.1404 | 0.1341 | 0.0842 | 0.0804 |
| parakeet-tdt-0.6b-v2 | **0.1084** | **0.1033** | **0.0641** | **0.0611** |

**`segmented` (46 h)**

| Model | WER | WER-N | WER-nc-np | WER-N-nc-np |
|---|---|---|---|---|
| whisper-base | 0.1937 | 0.1860 | 0.1202 | 0.1161 |
| whisper-medium | 0.1554 | 0.1493 | 0.0868 | 0.0837 |
| whisper-large-v2 | 0.1533 | 0.1472 | 0.0846 | 0.0817 |
| parakeet-tdt-0.6b-v2 | **0.1111** | **0.1050** | **0.0647** | **0.0606** |

**Submit a result:** open a [Discussion](https://huggingface.co/datasets/florencejiang/earnings25/discussions)
with your model, config, the four WER figures, and enough detail to reproduce
(decoding settings, normalization). Verified runs are added to the table above.

## Limitations

- English only; `segmented` is entirely US-domiciled and `full` is 93.8% US.
  Not a test of multilingual or non-US-market ASR.
- `segments` is empty for every item in `segmented`. Diarization and
  speaker-conditioned evaluation are possible on `full` only.
- Speaker turns are phrase-level. There are no word-level timestamps in this
  release.
- Speakers carry names, not roles — see [On speaker labels](#on-speaker-labels).
- `full` retains original recording quality, so audio conditions vary by call.
  That is deliberate, but it means cross-call WER variance is high.
- One segment per industry in `segmented` means per-industry WER rests on a
  single sample and is indicative, not significant.

## Relationship to the paper

This repository mirrors the [Zenodo release of record](https://doi.org/10.5281/zenodo.18762167)
(v1.0.0) and describes the files as shipped. Two details differ from the paper:
the paper lists `segmented` audio as 16 kHz WAV where the release ships 16 kHz
MP3, and describes word-level forced alignment that the released manifests do
not include. Cite the paper, not this mirror.

## License and provenance

Transcripts, annotations, and metadata are released under
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Audio is redistributed
from publicly disclosed corporate earnings calls. The dataset was created at
Bloomberg; attribution belongs to all ten authors.