Google FLEURS is a multilingual speech dataset, not a sentence-pair corpus. Its Hindi subset is useful for automatic speech recognition (ASR), speech evaluation, and experimentation with Indian-language voice systems. This guide shows where to download it on Hugging Face, how to load it with Python, and what to check before using it in a 2026 project.
Where to download Google FLEURS Hindi
The canonical Hugging Face dataset repository is google/fleurs. Open the page and select the Hindi configuration, commonly shown as hi_in or a closely related Hindi language configuration in the available configuration list. Do not rely on a third-party re-upload when the official repository is available: the original entry provides the dataset card, configuration names, split information, and licensing details.
The direct dataset URL is:
If you are searching manually, use FLEURS, not “Flurs”. FLEURS stands for Few-shot Learning Evaluation of Universal Representations of Speech. The dataset contains spoken utterances paired with transcriptions across many languages, including Hindi.
For broader discovery, the Open Source AI Datasets for India: A Builder’s Guide is useful when you need to compare FLEURS with Indic speech, text, and multimodal resources.
Download the Hindi configuration with Python
Install the current Hugging Face Datasets library and load the Hindi configuration:
pip install -U datasets[audio] soundfilefrom datasets import load_dataset
fleurs_hi = load_dataset(
"google/fleurs",
"hi_in",
trust_remote_code=False,
)
print(fleurs_hi)
print(fleurs_hi["train"][0])The exact configuration label can change in the dataset interface. If hi_in is rejected, inspect the available configurations from the dataset page or run:
from datasets import get_dataset_config_names
print(get_dataset_config_names("google/fleurs"))Then substitute the Hindi configuration returned by that command. This is safer than copying an old code snippet because dataset libraries and repository metadata can change.
To download only one split, specify it directly:
train_hi = load_dataset(
"google/fleurs",
"hi_in",
split="train",
)Hugging Face caches downloaded files locally. For controlled experiments, set a dedicated cache directory or use HF_HOME so large audio files do not fill your system disk unexpectedly.
What is included in the Hindi data?
Each record generally includes an audio field and a transcription, along with metadata such as the language identifier and sample information. Inspect the first example rather than assuming a fixed schema:
example = train_hi[0]
print(example.keys())
print(example["transcription"])
print(example["audio"]["sampling_rate"])The audio may be decoded automatically by datasets. If you need the raw path or want to control decoding, consult the repository’s dataset card and the installed datasets version. For ASR training, normalise text consistently, preserve Devanagari characters, and avoid silently removing punctuation unless your evaluation protocol requires it.
FLEURS is primarily an evaluation and supervised speech resource. It should not be described as a Hindi dialect corpus, a translation dataset, or a general sentiment-analysis dataset. It is better suited to:
- Hindi ASR fine-tuning and evaluation
- Comparing multilingual speech encoders
- Testing language identification or speech representation models
- Building reproducible baselines for Indian-language speech recognition
- Measuring transcription quality before adding domain-specific audio
For model selection and evaluation design, see Indian Language LLM Benchmark Datasets: A 2026 Evaluation Guide. Speech benchmarks require different metrics and data controls from text-only LLM benchmarks.
Train, validation, and test splits
Start by checking the available splits:
print(train_hi)
print(fleurs_hi.keys())Use the supplied training, validation, and test splits as intended when reproducing published or community baselines. Avoid repeatedly tuning on the test set. If you merge splits for a small experiment, clearly document the change because the resulting word error rate (WER) will no longer be comparable with standard results.
A minimal inspection workflow should record:
- Number of examples in every split
- Audio sampling rate and duration range
- Empty or unusually short transcriptions
- Duplicate audio or transcription records
- Unicode normalisation and punctuation policy
- Model tokenizer and vocabulary coverage for Hindi
For Hindi ASR, report both WER and character error rate (CER) where possible. WER can be sensitive to tokenisation and whitespace conventions in Indic scripts; CER often gives additional insight into Devanagari transcription quality. If you are specifically investigating recognition quality, the topic on Hindi ASR low WER provides useful context for metric interpretation.
Common problems when downloading FLEURS
Configuration error: The Hindi name may not match an old tutorial. Query the available configurations and use the current repository metadata.
Authentication or access error: Confirm that you are using the official public dataset URL, have an up-to-date huggingface_hub installation, and are not behind a proxy blocking Hugging Face storage.
Audio decoding failure: Install the audio extras and system dependencies required by your environment. In notebooks, restart the runtime after upgrading packages.
Insufficient storage: Load a single split, use streaming for inspection, or relocate the Hugging Face cache. Streaming is useful for checking records, but full training typically benefits from a stable local cache.
Unexpected Hindi text: Check the language configuration and inspect the metadata. Do not assume that a multilingual dataset call selected Hindi merely because the repository contains Hindi data.
Licensing, attribution, and responsible use
Read the FLEURS dataset card before redistribution or commercial deployment. Record the dataset version, configuration, library versions, preprocessing steps, and model checkpoint used in every experiment. Dataset availability does not automatically grant unrestricted rights to every downstream use.
FLEURS should be treated as a benchmark or starting point, not a substitute for representative production data. Hindi speech varies by region, age, accent, recording device, code-switching patterns, and speaking context. Validate a model on consented, task-relevant data before deploying it in education, public services, customer support, or voice assistants. If you are building a Hindi voice interface, compare the dataset with the resources discussed in Open-Source Hindi Voice Assistant Libraries.
Recommended workflow
1. Open the official google/fleurs repository.
2. Confirm the current Hindi configuration name.
3. Load one split with load_dataset.
4. Inspect audio, transcription, duration, and Unicode formatting.
5. Establish a reproducible baseline with WER and CER.
6. Keep validation and test data isolated.
7. Document licensing, versions, and preprocessing.
8. Test on representative Indian speech before making deployment claims.
This approach gives you a reliable Hindi FLEURS download and a defensible foundation for ASR experiments, rather than simply copying audio files into a training directory.