datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
VaaniVAANI is an India-representative multi-modal multi-lingual dataset.
The current version (phase 1- 80 districts, phase 2- 85 districts) contains ~31278 hours of spontaenous,image-prompted speech by 156K speakers across 165 districts, talking about 288K images covering 105 languages.
From this audio data, 2,122 hours of transcribed data(text) is available, spanning almost evenly across the 165 districts.
Project Vaani, by IISc, Bangalore and ARTPARK, is capturing the true diversity of India’s… See the full description on the dataset page: https://huggingface.co/datasets/ARTPARK-IISc/Vaani.ghana-speech-ipa
Ghana Speech — Audio with IPA Transcripts
Speech with both transcript forms: the original orthography and the IPA phoneme
sequence read off the audio by ASR. Each language is a subset, with real
train/validation splits.
from datasets import load_dataset
ds = load_dataset("ghanaopendata/ghana-speech-ipa", "Akuapem_Twi_twi", split="train")
ds[0]["audio"] # decoded waveform, 16 kHz
ds[0]["text"] # original orthography
ds[0]["ipa"] # IPA phonemes
369,347 clips · ~747 h… See the full description on the dataset page: https://huggingface.co/datasets/ghanaopenai/ghana-speech-ipa.IndicSynth
IndicSynth: Indian Multilingual Audio Deepfake Detection & Anti-Spoofing Dataset
A Large-Scale Multilingual Synthetic Speech Dataset for Low-Resource Indian Languages to facilitate audio deepfake detection and anti-spoofing research
🏆 Outstanding Paper Award, ACL 2025
🧠 Overview
IndicSynth is a novel multilingual synthetic speech dataset designed to advance multilingual audio deepfake detection (ADD) and anti-spoofing research. It covers 12 low-resource Indian… See the full description on the dataset page: https://huggingface.co/datasets/vdivyasharma/IndicSynth.indic-dialect-asr
Indic Dialect ASR Dataset
A multilingual ASR dataset covering 30 Indic dialect/languages with 2.8M+ samples.
Usage
from datasets import load_dataset
# Load a specific language
ds = load_dataset("grushaaaaa/indic-dialect-asr", "assamese", split="train")
Features
audio: 16kHz WAV audio
sentence: Transcription text
language: Language name
source: Source dataset
IndicSynth
IndicSynth: Indian Multilingual Audio Deepfake Detection & Anti-Spoofing Dataset
A Large-Scale Multilingual Synthetic Speech Dataset for Low-Resource Indian Languages to facilitate audio deepfake detection and anti-spoofing research
🏆 Outstanding Paper Award, ACL 2025
🧠 Overview
IndicSynth is a novel multilingual synthetic speech dataset designed to advance multilingual audio deepfake detection (ADD) and anti-spoofing research. It covers 12 low-resource Indian… See the full description on the dataset page: https://huggingface.co/datasets/ksmashhero/IndicSynth.ISSAI_KSC_335RS_v_1_1
Dataset Card for "ISSAI_KSC_335RS_v_1_1"
Kazakh Speech Corpus (KSC)
Identifier: SLR102
Summary: A crowdsourced open-source Kazakh speech corpus developed by ISSAI (330 hours)
Category: Speech
License: Attribution 4.0 International (CC BY 4.0)
Downloads (use a mirror closer to you):
ISSAI_KSC_335RS_v1.1_flac.tar.gz [19G] (speech, transcripts and metadata ) Mirrors: [US] [EU] [CN]
About this resource:
A crowdsourced open-source speech corpus for the Kazakh language. The KSC… See the full description on the dataset page: https://huggingface.co/datasets/Shirali/ISSAI_KSC_335RS_v_1_1.dsb_audio_corpus
Acknowledgements
Thanks to all speakers that contributed to this dataset!
Thanks to "Ludowe Nakładnistwo Domowina" and "Rěčny Centrum WITAJ" for donation of their recordings!
Vaani-transcription-partThis dataset is part of the Vaani dataset and consists of only transcribed speech data. It has a total duration of 2041.54 hours, covering 59 languages.
This table represents the audio and transcription duration data for various languages.
Language
Angami
Angika
Ao
Assamese
Awadhi
Bajjika
Bearybashe
Bengali
Bhili
Bhojpuri
Bundeli
Chakhesang
Chakma
Chhattisgarhi
English
Garhwali
Garo
Gondi
Gujarati
Halbi
Haryanvi
Hindi
IduMishmi
Kannada
Kashmiri
Karbi
Khariboli
Khortha
Kokborok
Konkani… See the full description on the dataset page: https://huggingface.co/datasets/ARTPARK-IISc/Vaani-transcription-part.new-twi-tts-aligned-ipa
new-twi-tts-aligned + IPA phonemes
ghanaopendata/new-twi-tts-aligned with a machine-generated IPA phoneme
transcription for every clip, produced with
ghananlpcommunity/ghana-speech-phoneme-asr.
Audio included — this is self-contained, no join with the source dataset needed.
Contents
split
clips
hours
phoneme units
mean units/clip
test
16,140
17.24
663,140
41.1
train
145,258
155.21
5,945,389
40.9
Columns
column
type
meaning… See the full description on the dataset page: https://huggingface.co/datasets/ghanaopenai/new-twi-tts-aligned-ipa.indic-diarbench
Indic DiarBench
A multilingual joint diarization and ASR benchmark for Indian languages, spanning all 22 scheduled languages of India with approximately 108 hours of natural multi-speaker audio.
Paper: Indic DiarBench: A Multilingual Joint Diarization and ASR Benchmark for Indian Languages (Interspeech 2026)
Dataset Summary
Indic DiarBench is a conversational speech benchmark designed to evaluate speaker-attributed ASR in realistic multi-speaker settings for… See the full description on the dataset page: https://huggingface.co/datasets/sarvamai/indic-diarbench.biggest-ru-bookA bigger version of its5Q/bigger-ru-book, the smaller set being a subset of this one. Almost 1000 hours of high-quality audio.
risale-i-nur-sohbet
Risale-i Nur Sohbet
Prof. Dr. Şener Dilek’ten izin alındı.
Türkçe
Risale-i Nur sohbetlerini ses, ham ASR metni ve zaman hizalı segmentler hâlinde
birlikte sunan bağımsız bir veri kümesidir. İlk sürüm izinli ve doğrulanmış
sohbetleri içerir; kitap metni, grounded, çok dilli veya kitap seslendirme veri
kümelerine karıştırılmaz.
Kapsam
2095 sohbet, 954.66 saat 16 kHz mono FLAC ses
Aynı derslerin ölçülmüş 48 kHz kalite katmanı; 786 derste
seçici… See the full description on the dataset page: https://huggingface.co/datasets/risaleinur/risale-i-nur-sohbet.ghana-speech-ipa
Ghana Speech — Audio with IPA Transcripts
Speech with both transcript forms: the original orthography and the IPA phoneme
sequence read off the audio by ASR. Each language is a subset, with real
train/validation splits.
from datasets import load_dataset
ds = load_dataset("ghanaopendata/ghana-speech-ipa", "Akuapem_Twi_twi", split="train")
ds[0]["audio"] # decoded waveform, 16 kHz
ds[0]["text"] # original orthography
ds[0]["ipa"] # IPA phonemes
369,347 clips · ~747 h… See the full description on the dataset page: https://huggingface.co/datasets/ghananlpcommunity/ghana-speech-ipa.kupe-asr-en-data
kupe-asr-en-mini-150m — data
Two loadable configs, packed into ~20-25 bunch_*.parquet files each (Hub-quota friendly):
raw — 24 kHz mono English audio (flac bytes) + text. The encode stage reads this.
mimi — Mimi c0..c7 codes (12.5 Hz) + text. Training reads this.
Ledgers under ledger/ (data.json, mimi.json) track collected/encoded hours and resume state.
from datasets import load_dataset
ds = load_dataset("anuj-inavlabs/kupe-asr-en-data", "mimi", split="train")
IndicSynth
IndicSynth: Indian Multilingual Audio Deepfake Detection & Anti-Spoofing Dataset
A Large-Scale Multilingual Synthetic Speech Dataset for Low-Resource Indian Languages to facilitate audio deepfake detection and anti-spoofing research
🏆 Outstanding Paper Award, ACL 2025
🧠 Overview
IndicSynth is a novel multilingual synthetic speech dataset designed to advance multilingual audio deepfake detection (ADD) and anti-spoofing research. It covers 12 low-resource Indian… See the full description on the dataset page: https://huggingface.co/datasets/mrunmai18/IndicSynth.infore2_audiobooks
unofficial mirror of InfoRe Technology public dataset №2
official announcement: https://www.facebook.com/groups/j2team.community/permalink/1010834009248719/
415h, 315k samples, vietnamese audiobooks of chinese wǔxiá 武俠 & xiānxiá 仙俠
bộ dữ liệu bóc ra từ YouTube đọc truyện võ hiệp & tiên hiệp, áp dụng kĩ thuật đối chiếu văn bản để dán nhãn tự động
official download:… See the full description on the dataset page: https://huggingface.co/datasets/doof-ferb/infore2_audiobooks.russian_librispeech
Russian LibriSpeech (RuLS)
Identifier: SLR96 from openslr.org
Summary: This dataset is based on LibriVox audiobooks
Category: Speech
License: The dataset is Public Domain in the USA.
About this resource:
Russian LibriSpeech (RuLS) dataset is based on LibriVox's public domain audio books (see BOOKS.TXT for the list of included books) and contains about 98 hours of audio data.
MonsoonASR-Open-ASR-leaderboard-en-IN
Voice Arena Monsoon en-IN (public test)
Part of the Open ASR Leaderboard, in the main board's default column set, so it contributes to the headline Average WER for every model listed.
A conversational Indian English ASR test set that records who was speaking, not only what
was said. Every clip carries twelve speaker attributes — gender, age, native district
and state, education, occupation, income band, handset — so a difference between two
systems can be traced to a group of… See the full description on the dataset page: https://huggingface.co/datasets/VoiceArena/MonsoonASR-Open-ASR-leaderboard-en-IN.seamless-interaction-jefferson-annotations
Seamless Interaction Jefferson-Style Annotations
An automatic, turn-oriented annotation layer for the
Meta Seamless Interaction Dataset.
It compares the dataset's traditional transcript with an ASR-derived
Jefferson-style condition and supplies speech-act, communicative-purpose,
interactional-signal, alignment, and quality fields.
This is a derived noncommercial research dataset. It does not redistribute
the source audio. Every record retains the original interaction ID, split… See the full description on the dataset page: https://huggingface.co/datasets/kennethli319/seamless-interaction-jefferson-annotations.african-speech-ipa
African Speech IPA
AfriSpeech audio
paired with IPA phoneme transcriptions, for 141 languages.
Each row has the audio, the source transcript, and ipa — the transcript converted to
space-separated IPA phoneme units with
africa-g2p 0.2.0. Units are kept whole, so
t͡ʃ, k͡p and kʰ are one token each rather than two or three characters, and
punctuation is preserved as its own unit so alignment and TTS keep their phrasing.
from datasets import load_dataset
ds =… See the full description on the dataset page: https://huggingface.co/datasets/AfriSpeech/african-speech-ipa.fleurs
FLEURS
Fleurs is the speech version of the FLoRes machine translation benchmark.
We use 2009 n-way parallel sentences from the FLoRes dev and devtest publicly available sets, in 102 languages.
Training sets have around 10 hours of supervision. Speakers of the train sets are different than speakers from the dev/test sets. Multilingual fine-tuning is
used and ”unit error rate” (characters, signs) of all languages is averaged. Languages and results are also grouped into seven… See the full description on the dataset page: https://huggingface.co/datasets/ItzmeNishh/fleurs.YouTube-Cantonese-Emilia
YouTube Cantonese — Emilia
2,064,679 speaker-homogeneous Cantonese speech segments — 5,312.6 hours — produced by
running alvanlii/cantonese-youtube
through the Emilia
speech-data pipeline (source separation → diarization → VAD segmentation → ASR → MOS filtering).
Each row is one clean, single-speaker segment of 3–30 s with a transcript, a speaker turn
label and a DNSMOS quality score. Audio is shipped separately as MP3s inside zip parts, in
both an original and a silence-trimmed… See the full description on the dataset page: https://huggingface.co/datasets/Scicom-intl/YouTube-Cantonese-Emilia.liepa3
LIEPA-3 Lithuanian Speech Corpus
This repository repackages the original LIEPA-3 release into Hugging Face
Parquet shards with embedded FLAC audio bytes. The original transcriptions are
kept as released: normalized lowercase Lithuanian text without punctuation,
digits, capitalization, or other symbols.
Recommended use:
read: cleanest subset and the default starting point for TTS or ASR.
spon: spontaneous/broadcast/media speech; useful for ASR, not a clean TTS default.
dial:… See the full description on the dataset page: https://huggingface.co/datasets/i4tech/liepa3.ghana-english-speech-ipa
Ghanaian English Speech — Audio with IPA Transcripts
Speech with both transcript forms: the original orthography and the IPA phoneme
sequence read off the audio by ASR. Each language is a subset, with real
train/validation splits.
from datasets import load_dataset
ds = load_dataset("ghanaopendata/ghana-english-speech-ipa", "English_eng", split="train")
ds[0]["audio"] # decoded waveform, 16 kHz
ds[0]["text"] # original orthography
ds[0]["ipa"] # IPA phonemes
52,855… See the full description on the dataset page: https://huggingface.co/datasets/ghanaopenai/ghana-english-speech-ipa.yodas-ja000
YODAS Japanese (ja000)
Japanese manual caption subset of the YODAS dataset, repackaged for easier use.
Source
Original dataset: espnet/yodas (ja000 config)
Paper: YODAS: YouTube-Oriented Dataset for Audio and Speech
License: CC BY 3.0
Citation
If you use this dataset, please cite the original YODAS paper:
ivirits-audio-v2-30s
ivrit.ai audio-v2 — 2–30 s segments
ivrit-ai/audio-v2 (>20k hours of Hebrew
audio) cut into 2–30 second speech segments with machine transcripts, ready for ASR
fine-tuning.
How it was built
VAD — Silero VAD (ONNX) over each episode decoded to 16 kHz mono. Speech regions
longer than 30 s are split at the quietest sufficiently-long pause inside the window,
so cuts land in silence rather than mid-word. Regions shorter than 2 s are dropped.
Transcription —… See the full description on the dataset page: https://huggingface.co/datasets/notmax123/ivirits-audio-v2-30s.Medical_InterviewThe dataset was re-organized and used in the following paper. Please cite if you adopted the corpus in your work.
@inproceedings{liu2024post,
title={Post-decoder Biasing for End-to-End Speech Recognition of Multi-turn Medical Interview},
author={Liu, Heyang and Wang, Yanfeng and Wang, Yu},
booktitle={Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
pages={12917--12926},
year={2024}
}
Whisper-Hallucination
Whisper Hallucination and Repetition Probes
This is a benchmark. Every evaluation config is test — do not fine-tune on it.
lexicon_synth is the exception: synthetic training material with its own train/test
split, and not one of the eight benchmark arms.
To build training data, exclude the items in
benchmark/exclusions.json
(546 FMA tracks, 1,168 FSD50K ids, 2,620 LibriSpeech utterances, the Malay stems). The
benchmark draws FSD50K eval and FMA shards 0–1, so training can use… See the full description on the dataset page: https://huggingface.co/datasets/Scicom-intl/Whisper-Hallucination.icelandic_asr
Icelandic ASR Collection
This repository collects six Icelandic speech corpora in directly loadable
Parquet form. Audio is embedded as 16 kHz mono FLAC bytes. The repository is a
convenience repackaging: the linked CLARIN-IS records and original dataset
repositories remain the canonical sources and should be cited when using the
data.
No configuration is selected by default. Choose a corpus configuration and,
for this large collection, normally choose a split explicitly.… See the full description on the dataset page: https://huggingface.co/datasets/Aalto-Speech-Synthesis/icelandic_asr.AfriSwitch
AfriSwitch: In-the-Wild African Code-Switched Speech Benchmark
Dataset Description
AfriSwitch is a 54.41-hour, human-transcribed test benchmark of in-the-wild, conversational
code-switched speech spanning 14 African languages: Amharic, Pidgin, Kinyarwanda, Yoruba, Hausa,
Oromo, Igbo, Zulu, French, Shona, Swahili, Tswana, Luganda, Afrikaans, each switching with English.
This release is distributed as a single test split (evaluation-only benchmark).
License: CC BY… See the full description on the dataset page: https://huggingface.co/datasets/intronhealth/AfriSwitch.