--- dataset_info: - config_name: Bulgarian features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Czech features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Danish features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: German features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Greek features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: English features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Spanish features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Estonian features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Finnish features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: French features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Croatian features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Hungarian features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Italian features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Lithuanian features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Latvian features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Dutch features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Polish features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Portuguese features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Romanian features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Slovak features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Slovenian features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: Swedish features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli - config_name: All features: - name: utt_id dtype: string - name: duration dtype: float64 - name: sampling_rate dtype: int32 - name: bandwidth dtype: int32 - name: lang dtype: string - name: text dtype: string - name: num_speakers dtype: int16 - name: single_speaker_id dtype: string - name: diarization_coverage_ratio dtype: float32 - name: speech_event_coverage_ratio dtype: float32 - name: speech_event_evidence dtype: float32 - name: clean_speech_event_score dtype: float32 - name: mos dtype: float32 - name: ipa dtype: string - name: early_cut_off dtype: bool - name: context_utt_id dtype: string - name: context_audio_speaker_similarity dtype: float32 - name: context_audio_text_overlap dtype: float32 - name: original_audio_id dtype: string - name: original_audio_offset dtype: float64 - name: HiFi dtype: bool splits: - name: VoxPopuli configs: - config_name: Bulgarian data_files: - split: VoxPopuli path: data/bg/VoxPopuli/*.parquet - config_name: Czech data_files: - split: VoxPopuli path: data/cs/VoxPopuli/*.parquet - config_name: Danish data_files: - split: VoxPopuli path: data/da/VoxPopuli/*.parquet - config_name: German data_files: - split: VoxPopuli path: data/de/VoxPopuli/*.parquet - config_name: Greek data_files: - split: VoxPopuli path: data/el/VoxPopuli/*.parquet - config_name: English data_files: - split: VoxPopuli path: data/en/VoxPopuli/*.parquet - config_name: Spanish data_files: - split: VoxPopuli path: data/es/VoxPopuli/*.parquet - config_name: Estonian data_files: - split: VoxPopuli path: data/et/VoxPopuli/*.parquet - config_name: Finnish data_files: - split: VoxPopuli path: data/fi/VoxPopuli/*.parquet - config_name: French data_files: - split: VoxPopuli path: data/fr/VoxPopuli/*.parquet - config_name: Croatian data_files: - split: VoxPopuli path: data/hr/VoxPopuli/*.parquet - config_name: Hungarian data_files: - split: VoxPopuli path: data/hu/VoxPopuli/*.parquet - config_name: Italian data_files: - split: VoxPopuli path: data/it/VoxPopuli/*.parquet - config_name: Lithuanian data_files: - split: VoxPopuli path: data/lt/VoxPopuli/*.parquet - config_name: Latvian data_files: - split: VoxPopuli path: data/lv/VoxPopuli/*.parquet - config_name: Dutch data_files: - split: VoxPopuli path: data/nl/VoxPopuli/*.parquet - config_name: Polish data_files: - split: VoxPopuli path: data/pl/VoxPopuli/*.parquet - config_name: Portuguese data_files: - split: VoxPopuli path: data/pt/VoxPopuli/*.parquet - config_name: Romanian data_files: - split: VoxPopuli path: data/ro/VoxPopuli/*.parquet - config_name: Slovak data_files: - split: VoxPopuli path: data/sk/VoxPopuli/*.parquet - config_name: Slovenian data_files: - split: VoxPopuli path: data/sl/VoxPopuli/*.parquet - config_name: Swedish data_files: - split: VoxPopuli path: data/sv/VoxPopuli/*.parquet - config_name: All default: true data_files: - split: VoxPopuli path: data/*/VoxPopuli/*.parquet license: openmdw-1.1 task_categories: - automatic-speech-recognition language: - bg - cs - da - de - el - en - es - et - fi - fr - hr - hu - it - lt - lv - nl - pl - pt - ro - sk - sl - sv pretty_name: Granary-HiFi tags: - granary - multilingual - speech - metadata --- # 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](https://huggingface.co/datasets/nvidia/Granary-v2) enriched metadata for the [VoxPopuli](https://huggingface.co/datasets/facebook/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. ![Retained and filtered-out hours by language](granaryhifi_retention_chart.png) #### Filtered dataset statistics
Language VoxPopuli
Samples Speakers Hours Retention rate
Bulgarian 668011346374960.3 30.4%
Croatian 389344195082984.6 38.5%
Czech 561828400524131.0 24.0%
Danish 406567338602854.9 21.6%
Dutch 891239476016465.8 34.9%
English 136135214903610544.4 45.7%
Estonian 236378242161664.0 17.1%
Finnish 337660315492435.6 17.8%
French 1213618753829077.6 40.4%
German 1009003773927540.5 33.0%
Greek 608222385694428.7 28.1%
Hungarian 812785409546101.9 37.1%
Italian 15292185859211862.0 55.1%
Latvian 361104299432563.4 20.9%
Lithuanian 464933324863307.6 24.9%
Polish 613470467264500.6 24.1%
Portuguese 867272431586307.6 36.9%
Romanian 515009397613762.8 22.0%
Slovak 275033273301999.5 17.7%
Slovenian 354422247532541.1 26.1%
Spanish 969596560007132.3 34.0%
Swedish 548861380094004.1 25.9%
Total 149949251009514111170.3 31.4%
--- ## How to use ### Standard Loading Load one language from the full annotated VoxPopuli dataset: ```python from datasets import load_dataset ds = load_dataset( "nvidia/Granary-HiFi", "English", split="VoxPopuli", ) ``` Load its high-fidelity subset from the `HiFi` revision: ```python ds = load_dataset( "nvidia/Granary-HiFi", "English", split="VoxPopuli", revision="HiFi", ) ``` ### Streaming To stream all 22 languages: ```python ds = load_dataset( "nvidia/Granary-HiFi", "All", split="VoxPopuli", revision="HiFi", streaming=True, ) ``` ## Dataset Structure ### 📊 Data Sample Structure ```python { # 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. ```text data/ └── en/ └── VoxPopuli/ ├── en_2009.parquet ├── en_2010.parquet └── ... ``` --- ## Reference * **Code:** [NVIDIA-NeMo / Curator (Granary-HiFi)](https://github.com/NVIDIA-NeMo/Curator/pull/2442) * **Source corpora:** [VoxPopuli](https://github.com/facebookresearch/voxpopuli), [nvidia/Granary](https://huggingface.co/datasets/nvidia/Granary), [nvidia/Granary-v2](https://huggingface.co/datasets/nvidia/Granary-v2) ```bibtex @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](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).