Dataset Card for Granary-HiFi

Dataset Description

Granary-HiFi is a corpus of multilingual speech data to which complementary filtering approaches have been applied to identify high-fidelity samples. This release presents Granary-v2 enriched metadata for the VoxPopuli subset; audio bytes are not distributed in this repository.

The processing pipeline includes:

1. Sound event detection to identify speech coverage and filter samples contaminated by non-speech sounds.
2. Speaker labelling to identify single-speaker segments within each source recording and assign consistent speaker labels.
3. UTMOS estimation to measure predicted perceptual speech quality and reject low-quality audio.
4. CTC/MFA alignment cascade to detect early cut-offs where the speech audio ends before the transcript.
5. Context retrieval to select a nearby, speaker-consistent reference segment with limited text overlap, providing useful acoustic and speaker context without duplicating the target content.

Overview

• Languages: 22
• Annotated VoxPopuli: 354,389.9 hours across 46,055,576 samples
• Granary-HiFi: 111,170.3 hours across 14,994,925 samples (31.4% retained)
• Speaker labels: 1,356,748 overall; 1,009,514 in Granary-HiFi

The main revision contains all annotated VoxPopuli samples and the HiFi membership flag. The HiFi revision contains only samples with HiFi=true; it omits the num_speakers and HiFi columns.

Data Distribution

The chart below shows Granary-HiFi and filtered-out VoxPopuli hours by language on a logarithmic scale. Retention rate is HiFi hours divided by all annotated hours.

Filtered dataset statistics

Language	VoxPopuli
	Samples	Speakers	Hours	Retention rate
Bulgarian	668011	34637	4960.3	30.4%
Croatian	389344	19508	2984.6	38.5%
Czech	561828	40052	4131.0	24.0%
Danish	406567	33860	2854.9	21.6%
Dutch	891239	47601	6465.8	34.9%
English	1361352	149036	10544.4	45.7%
Estonian	236378	24216	1664.0	17.1%
Finnish	337660	31549	2435.6	17.8%
French	1213618	75382	9077.6	40.4%
German	1009003	77392	7540.5	33.0%
Greek	608222	38569	4428.7	28.1%
Hungarian	812785	40954	6101.9	37.1%
Italian	1529218	58592	11862.0	55.1%
Latvian	361104	29943	2563.4	20.9%
Lithuanian	464933	32486	3307.6	24.9%
Polish	613470	46726	4500.6	24.1%
Portuguese	867272	43158	6307.6	36.9%
Romanian	515009	39761	3762.8	22.0%
Slovak	275033	27330	1999.5	17.7%
Slovenian	354422	24753	2541.1	26.1%
Spanish	969596	56000	7132.3	34.0%
Swedish	548861	38009	4004.1	25.9%
Total	14994925	1009514	111170.3	31.4%

---

How to use

Standard Loading

Load one language from the full annotated VoxPopuli dataset:

from datasets import load_dataset

ds = load_dataset(
    "nvidia/Granary-HiFi",
    "English",
    split="VoxPopuli",
)

Load its high-fidelity subset from the HiFi revision:

ds = load_dataset(
    "nvidia/Granary-HiFi",
    "English",
    split="VoxPopuli",
    revision="HiFi",
)

Streaming

To stream all 22 languages:

ds = load_dataset(
    "nvidia/Granary-HiFi",
    "All",
    split="VoxPopuli",
    revision="HiFi",
    streaming=True,
)

Dataset Structure

📊 Data Sample Structure

{
    # Core utterance metadata
    "utt_id": str,                                  # Unique utterance identifier
    "duration": float,                              # Duration in seconds
    "sampling_rate": int | None,                    # Sampling rate in Hz
    "bandwidth": int,                               # Estimated audio bandwidth in Hz
    "lang": str,                                    # ISO 639-1 language code
    "text": str,                                    # Normalized transcript

    # Speaker identity and diarization
    "num_speakers": int,                            # Estimated number of speakers
    "single_speaker_id": str | None,                # Speaker ID when exactly one speaker is present
    "diarization_coverage_ratio": float,             # Fraction covered by diarized speech

    # Speech-event detection
    "speech_event_coverage_ratio": float,            # Fraction covered by speech events
    "speech_event_evidence": float,                  # Speech-event evidence from the SED model
    "clean_speech_event_score": float,               # Classifier score for clean speech

    # Perceptual speech quality
    "mos": float,                                    # Predicted mean opinion score

    # Transcript and boundary quality
    "ipa": str,                                      # Transcript in the International Phonetic Alphabet
    "early_cut_off": bool,                           # Whether the utterance ends prematurely

    # Context selection
    "context_utt_id": str | None,                    # Selected context utterance; null when HiFi=false
    "context_audio_speaker_similarity": float | None,# Target/context speaker similarity; null when HiFi=false
    "context_audio_text_overlap": float | None,      # Target/context text overlap; null when HiFi=false

    # Source provenance
    "original_audio_id": str,                        # Identifier of the original source audio
    "original_audio_offset": float | None,           # Start offset in the original audio, in seconds

    # Granary-HiFi membership
    "HiFi": bool,                                    # Membership in the filtered Granary-HiFi subset
}

The HiFi revision contains the same fields except num_speakers and HiFi.

Data Splits and Files

Each language configuration exposes one VoxPopuli split. The All configuration exposes the same split across all 22 languages. The file hierarchy preserves the source-dataset name for convenient provenance tracking.

data/
└── en/
    └── VoxPopuli/
        ├── en_2009.parquet
        ├── en_2010.parquet
        └── ...

---

Reference

• Code: NVIDIA-NeMo / Curator (Granary-HiFi)
• Source corpora: VoxPopuli, nvidia/Granary, nvidia/Granary-v2

@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},
}

The Granary-HiFi paper is in preparation. The BibTeX entry above contains the Granary-v1 citation.

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here.