# FLORAS: long-form speech recognition, translation and summarization benchmark

## Facts

- Provider: ESPnet / CMU WAVLab
- License: CC-BY-3.0 (https://arxiv.org/pdf/2406.00899)
- Size: 7,138 files
- Format: Parquet (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/espnet/floras
- Access: public, readable without an account
- Category: Speech & Audio
- Source: https://huggingface.co/datasets/espnet/floras
- Page: https://scigantic.com/archive/3353ba0a-080f-5afe-bce7-26c2700c15e4
- Keywords: Speech & Audio, speech recognition, benchmark, Parquet, Hugging Face

## About

FLORAS is a 50-language benchmark For LOng-form Recognition And Summarization of spoken language. Unlike pre-segmented, single-speaker read-speech benchmarks such as LibriSpeech and FLEURS, it tests models on raw long-form conversational audio with one or many speakers. It provides one- to three-way parallel data for long-form automatic speech recognition, X-to-English speech translation and speech summarization, so some samples have only speech and transcripts while others add translations and summaries; in total it contains roughly 32,000 hours of raw audio. FLORAS is derived from YODAS, the Creative Commons YouTube crawl, filtered with CTC alignment scores to remove noisy data; translations and summaries were produced by pseudo-labelling with a large language model, as the card describes. This repository holds two configurations, monolingual and multilingual, as Parquet shards (922 files in monolingual; the multilingual folder lists at least 1,000) with 16 kHz audio and text fields (measured 2026-11-14; the Hub reports about 3 TB). Load with datasets.load_dataset('espnet/floras', 'monolingual', streaming=True). It inherits the CC BY 3.0 licence of its YODAS source, as stated in the YODAS paper and on the card.

## Load the data

```python
from datasets import load_dataset
ds = load_dataset("espnet/floras")
print(ds)
```

## Data structure

File types in sampled files: Parquet

Example file: `monolingual/dev-00000-of-00008.parquet`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/3353ba0a-080f-5afe-bce7-26c2700c15e4
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/3353ba0a-080f-5afe-bce7-26c2700c15e4
- Over MCP: `get_schema_card` with id `3353ba0a-080f-5afe-bce7-26c2700c15e4` at https://api.scigantic.com/mcp

## Related topics

- [Parquet datasets](https://scigantic.com/topics/parquet)
- [Benchmark datasets](https://scigantic.com/topics/benchmark)
- [Speech & Audio datasets](https://scigantic.com/topics/speech-and-audio)
- [Speech recognition datasets](https://scigantic.com/topics/speech-recognition)
