Dataset Card for YODAS-Granary

• Repository: NeMo-speech-data-processor: Granary
• Paper: Granary: Speech Recognition and Translation Dataset in 25 European Languages
• Shared by: ESPnet

Dataset Description

YODAS-Granary is a curated subset of the larger nvidia/Granary dataset, focusing on high-quality pseudo-labeled speech data for Automatic Speech Recognition (ASR) and Automatic Speech Translation (AST) across 23 European languages.

Overview

Derived from the espnet/yodas2 corpus, YODAS-Granary provides high-quality pseudo-labeled speech data, focusing on two core tasks: Automatic Speech Recognition (ASR): covers 23 European languages, with pseudo-labeled transcriptions generated using the Systran/faster-whisper-large-v3 model, post-processed to restore punctuation and capitalization using Qwen/Qwen2.5-7B-Instruct, and filtered for quality. Automatic Speech Translation (AST): covers 22 non-English languages and consists of high-quality translations into English, generated from ASR subset using the utter-project/EuroLLM-9B-Instruct model and filtered for quality.

Data Distribution

The following chart illustrates the distribution of data in the YODAS-Granary dataset across 23 European languages, measured in number of words (left) and total hours of audio (right), for both ASR and AST tasks.

🔍 AST data is always a filtered subset of ASR, which is why AST bars are never taller than their ASR counterparts.
🗣️ The English subset contains ASR data only.

The table below summarizes the storage footprint and sample counts per language in the YODAS-Granary dataset, broken down into corresponding splits:

• ast — the size and number of translated samples (X → English),
• asr_only — samples that exist only in the ASR subset and have no corresponding translation,

and combined size and number of all samples per language (total).

Language	Subsets	Samples [ast]	Size [ast]	Samples [asr_only]	Size [asr_only]	Total samples	Total size
Bulgarian	bg000	8844	1.9 GB	1533	208.8 MB	10377	2.1 GB
Czech	cs000	34360	7.5 GB	4185	442.1 MB	38545	8.0 GB
Danish	da000	9582	1.9 GB	656	65.3 MB	10238	2.0 GB
German	de000, de{100..102}	3335156	845.6 GB	415260	56.3 GB	3750416	901.9 GB
Greek	el000	4242	1.6 GB	514	113.5 MB	4756	1.7 GB
English	en00{0..7}, en{100..129}	❌	❌	40810517	11.3 TB	40810517	11.3 TB
Spanish	es000, es{100..108}	7923646	2.9 TB	951450	88.8 GB	8875096	3.0 TB
Estonian	et000	4437	901.8 MB	513	57.8 MB	4950	959.6 MB
Finnish	fi000	60729	17.4 GB	4637	419.3 MB	65366	17.8 GB
French	fr000, fr{100..103}	4766239	1.3 TB	558848	81.1 GB	5325087	1.4 TB
Croatian	hr000	5369	1.1 GB	261	27.9 MB	5630	1.1 GB
Hungarian	hu000	48263	16.4 GB	6530	962.7 MB	54793	17.4 GB
Italian	it000, it{100..101}	1226663	683.7 GB	86587	13.6 GB	1313250	697.3 GB
Lithuanian	lt000	2177	564.5 MB	390	71.1 MB	2567	635.6 MB
Latvian	lv000	272	66.5 MB	75	12.0 MB	347	78.5 MB
Dutch	nl000, nl100	865754	151.1 GB	71490	4.3 GB	937244	155.4 GB
Polish	pl000	264257	75.5 GB	28678	2.4 GB	292935	77.9 GB
Portuguese	pt000, pt{100..103}	5898764	1.5 TB	729138	30.1 GB	6627902	1.5 TB
Romanian	ro000	12276	3.7 GB	2303	663.4 MB	14579	4.4 GB
Russian	ru00{0..1}, ru{100..106}	7991038	2.1 TB	1876876	197.0 GB	9867914	2.3 TB
Slovak	sk000	3405	992.0 MB	287	51.9 MB	3692	1.0 GB
Swedish	sv000	54085	10.2 GB	3192	215.7 MB	57277	10.4 GB
Ukrainian	uk000, uk100	246373	68.3 GB	9479	1.2 GB	255852	69.4 GB

---

How to use

Standard Loading

You can load the dataset using the datasets library from Hugging Face:

from datasets import load_dataset

🔹 Load the entire dataset:

ds = load_dataset("espnet/yodas-granary")

🔹 Load a single language (e.g., Italian):

ds = load_dataset("espnet/yodas-granary", "Italian")

Streaming

Some language subsets are quite large and may not fit comfortably in memory. For efficient access and analysis without downloading the entire dataset, we recommend using streaming mode:

ds = load_dataset("espnet/yodas-granary", "English", streaming=True)

Using NeMo-speech-data-processor

You can use the NeMo-speech-data-processor to convert YODAS-Granary into a tarred WebDataset format suitable for training or fine-tuning NeMo ASR models.

Clone and install the processor:

git clone https://github.com/NVIDIA/NeMo-speech-data-processor.git
cd NeMo-speech-data-processor && pip install -e .

By specifying the desired source_lang, en_translation, num_shards, and buckets_num, the script will automatically download the required language subsets from Hugging Face and convert them into WebDataset format:

python main.py \
  --config-path=dataset_configs/multilingual/granary/ \
  --config-name=yodas2.yaml \
  params.source_lang="it" \  # target language
  params.en_translation=True \  # use AST or ASR subset
  params.convert_to_audio_tarred_dataset.num_shards=1024 \  # number of shards per bucket
  params.convert_to_audio_tarred_dataset.buckets_num=1  # number of output buckets

📘 For detailed setup instructions, see the NeMo-speech-data-processor: Granary.

---

Dataset Structure

Data Instance

Each utterance in the dataset includes the following fields: utt_id, audio, duration, lang, task, text, translation_en (null in asr_only), original_audio_id, and original_audio_offset.

Typical entry

from data/de101/translation/00000000.parquet

{
  "utt_id": "de101_00000000_Z0_gcPJVTqg_1004_62_1_74",
  "audio": {
    'path': 'de101_00000000_Z0_gcPJVTqg_1004_62_1_74.wav', 
    'bytes': ... 
  }
  "duration": 1.74,
  "lang": "<de>",
  "task": "<ast>",
  "text": "Ich muss mir das Zeug mal aus der Nähe ansehen.",
  "translation_en": "I have to take a closer look at this stuff.",
  "original_audio_id": "Z0_gcPJVTqg",
  "original_audio_offset": 1004.62
}

Data Fields

Field	Type	Description
utt_id¹	string	Unique identifier of the utterance, referencing the original segment.
audio	Audio (16 kHz)	Audio data of the utterance, stored as PCM waveform.
duration	float64	Duration of the utterance in seconds.
lang	string	Language of the utterance, in ISO 639-1 code (e.g., <de> for German).
task	string	Task type: either <asr> for transcription or <ast> for translation to English.
text	string	Transcription of the utterance in its original language.
translation_en	string	English translation of the utterance. null if split is asr_only.
original_audio_id	string	ID of the original audio file. This value corresponds to the audio_id field from the espnet/yodas2 dataset.
original_audio_offset	float64	Start time (in seconds) of the utterance within the original audio file.

¹ - utt_id is encoded as <subset>_<shard>_<wav_id>_<start_time_s>_<start_time_decimals>_<duration_s>_<duration_decimals>, where subset, shard², and wav_id match the utterance's location in the original espnet/yodas2 archive.

² - shard indices reflect those in espnet/yodas2, but some shards are missing due to filtering during data processing. In particular, bg000 is missing shard 00000011, en000 is missing shard 00000308, en003 is missing shard 00000221, en118 is missing shard 00000240.

Data Splits

The dataset is organized into language-specific subsets, each containing one or two splits, depending on the language:

• For non-English languages:
  • <ast> – samples with both high-quality transcriptions and translations into English.
  • <asr_only> – samples that passed transcription quality checks but do not include translations.
• For English:
  • <asr_only> split is available only, since English-to-English translation is not applicable.

Directory example

yodas_granary
└── data
    ├── da000   # subset
    │   ├── asr_only # corresponds to `asr_only` split
    │   │   ├── 00000000.parquet # shard
    │   │   ├── 00000001.parquet
    │   │   ├── 00000002.parquet
    │   │   └── ...
    │   └── ast # corresponds to `ast` split
    │       ├── 00000000.parquet # shard
    │       ├── 00000001.parquet
    │       ├── 00000002.parquet
    │       └── ...
    ├── cs000
    ├── bg000
    └── ...

---

Reference

@misc{koluguri2025granaryspeechrecognitiontranslation,
      title={Granary: Speech Recognition and Translation Dataset in 25 European Languages}, 
      author={Nithin Rao Koluguri and Monica Sekoyan and George Zelenfroynd and Sasha Meister and Shuoyang Ding and Sofia Kostandian and He Huang and Nikolay Karpov and Jagadeesh Balam and Vitaly Lavrukhin and Yifan Peng and Sara Papi and Marco Gaido and Alessio Brutti and Boris Ginsburg},
      year={2025},
      eprint={2505.13404},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.13404}, 
}