datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
yodas-granary
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… See the full description on the dataset page: https://huggingface.co/datasets/espnet/yodas-granary.yodas2_sidon
YODAS2-Sidon
Overview
This dataset is a cleansed version of YODAS-2 with Sidon speech restoration mode for Speech Synthesis and Spoken Language Modeling.
YODAS-2 is a massive, multilingual YouTube-derived dataset. We have applied the Sidon restoration model to remove background noise and enhance audio quality, making it suitable for high-quality generation tasks.
We resampled original sidon output to 24kHz due to a storage constraints.
The dataset is provided in… See the full description on the dataset page: https://huggingface.co/datasets/sarulab-speech/yodas2_sidon.YO-CPT-ru
YO-CPT-ru
YouTube-Oriented dataset for Continual Pre-Training (Russian). A large, heavily
quality-filtered corpus of Russian speech mined from YouTube (via YODAS2)
and processed into clean, single-speaker, TTS-grade utterances. Every utterance ships with an
ensemble-verified transcription, a punctuated/denormalized and stress-marked text variant, word-level
forced alignment, within- and cross-video speaker identities, an audio-quality (MOS) score, and a
speaker persona built… See the full description on the dataset page: https://huggingface.co/datasets/NCSpeech/YO-CPT-ru.tat_youtubeopen-yap-1k
Open Yap 1K: 1,000 hours of full-duplex natural conversation, free for commercial use
Today we're releasing Open Yap 1K: 1,000 hours of dual-channel English conversation, capturing how people speak together naturally in real-world environments recorded in 48kHz. The dataset ships free for both commercial and research use.
The sample on the Hugging Face Hub - 8.9 hours, 16 conversations, CC-BY-4.0, listenable in the dataset viewer.
The full corpus - 1,000 hours, 1,602… See the full description on the dataset page: https://huggingface.co/datasets/TheAgenticDataCompany/open-yap-1k.yodas_owsmv4🏆 News: Our OWSM v4 paper won the Best Student Paper Award at INTERSPEECH 2025!
Dataset Card for YODAS_OWSMv4
Paper: OWSM v4: Improving Open Whisper-Style Speech Models via Data Scaling and Cleaning (Best Student Paper at INTERSPEECH 2025)
Authors: Yifan Peng, Muhammad Shakeel, Yui Sudo, William Chen, Jinchuan Tian, Chyi-Jiunn Lin, Shinji Watanabe
Data Cleaning Scripts: ESPnet
Model Demo: Gradio
Dataset Description
Open Whisper-style Speech Model (OWSM)is the first… See the full description on the dataset page: https://huggingface.co/datasets/espnet/yodas_owsmv4.emilia-yodasA mirror of the Emilia-YODAS dataset. Only includes the YODAS subset from the original dataset.
https://huggingface.co/datasets/amphion/Emilia-Dataset
yt-danish-public-v2ytseg
YTSeg: A Benchmark for Audio Chaptering and Video Transcript Segmentation
We present YTSeg, a topically and structurally diverse benchmark for the audio chaptering and transcript segmentation task based on YouTube videos. The dataset comprises 19,299 videos from 393 channels, amounting to 6,533 content hours. The topics are wide-ranging, covering domains such as science, lifestyle, politics, health, economy, and technology. The videos are from various types of content formats… See the full description on the dataset page: https://huggingface.co/datasets/retkowski/ytseg.Multi-Talker-SD
Dataset Card for Multi-Talker-SD
Dataset Description
Multi-Talker-SD is a large-scale bilingual (English–Mandarin) multi-speaker meeting dataset designed to support research on speaker diarization and meeting transcription.
Size: 1,000 simulated meetings
Participants per meeting: 10–30 speakers
Average duration: ~20 minutes per meeting, up to one hour
Languages: English, Mandarin (code-switching possible)
Audio characteristics: realistic speaker overlap… See the full description on the dataset page: https://huggingface.co/datasets/yihao005/Multi-Talker-SD.YO-CPT-kk
YO-CPT-kk
YouTube-Oriented dataset for Continual Pre-Training (Kazakh). A heavily
quality-filtered corpus of Kazakh speech mined from YouTube and processed into clean, single-speaker,
TTS-grade utterances. Every utterance ships with an ensemble-verified transcription, a
punctuated/denormalized and stress-marked text variant, word-level forced alignment, within- and
cross-video speaker identities, an audio-quality (MOS) score, and a speaker persona built from the
voice and, where… See the full description on the dataset page: https://huggingface.co/datasets/NCSpeech/YO-CPT-kk.pseudolabel-malaysian-youtube-whisper-large-v3
Pseudolabel Malaysian Youtube videos using Whisper Large V3
Original dataset at https://huggingface.co/datasets/malaysia-ai/crawl-youtube, distributed pseudolabelled using 4x A100s
script at https://github.com/mesolitica/malaysian-dataset/tree/master/speech-to-text-semisupervised/pseudolabel-whisper
Each audio is 30 seconds.
Each audio saved in 16k sample rate.
audiobooks170 hours of aligned audiobooks taken from tatkniga.ru. There are 4 speakers with 17+ hours of audio and 20 speakers in total. All the books are in free access and most of them in public domain.
cantonese-youtube-tts
Cantonese Audio TTS Dataset
This dataset contains alvanlii/cantonese-radio, alvanlii/cantonese-youtube, plus a dataset of equal size. It is catered towards TTS (text-to-speech) use cases, more than the 2 previously published datasets, as there is more extensive filtering and audio enhancement. For speaker labelling, you can use speaker embedding models like Nvidia's TitaNet
Filtered out:
Overlapped voices, detected using pyannote/speaker-diarization-3.1
Music, detected using a… See the full description on the dataset page: https://huggingface.co/datasets/alvanlii/cantonese-youtube-tts.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.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:
audiobooks-xxlfleurs
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/yohannabelay/fleurs.YouTube-Commons-nl-audio
YouTube Commons NL Audio
This dataset has audio files for the Dutch-language videos in Rijgersberg/YouTube-Commons-nl-transcriptions,
all under a CC BY 4.0 license.
It contains 11,669 files for a total runtime of 2493h 43m 5s, coming in at approximately 130 GB.
Source
The original source of the dataset (minus the titles, descriptions and audio files) is YouTube Commons:
YouTube-Commons is a collection of audio transcripts of 2,063,066 videos shared on YouTube… See the full description on the dataset page: https://huggingface.co/datasets/Rijgersberg/YouTube-Commons-nl-audio.WaxalNLP
Waxal Datasets
The WAXAL dataset is a large-scale multilingual speech corpus for African languages, introduced in the paper WAXAL: A Large-Scale Multilingual African Language Speech Corpus.
Dataset Description
The Waxal project provides datasets for both Automated Speech Recognition (ASR)
and Text-to-Speech (TTS) for African languages. The goal of this dataset's
creation and release is to facilitate research that improves the accuracy and
fluency of speech and… See the full description on the dataset page: https://huggingface.co/datasets/youvoi/WaxalNLP.cantonese-youtube
Cantonese Youtube Pseudo-Transcription Dataset
Contains approximately 10k hours of audio sourced from YouTube
Videos are chosen at random, and scraped on a channel basis
Includes news, vlogs, entertainment, stories, health
Columns
transcript_whisper: Transcribed using Scrya/whisper-large-v2-cantonese with alvanlii/whisper-small-cantonese for speculative decoding
transcript_sensevoice: Transcribed using FunAudioLLM/SenseVoiceSmall
used OpenCC to convert to traditional chinese… See the full description on the dataset page: https://huggingface.co/datasets/alvanlii/cantonese-youtube.news_youtube_uzbek_speech_dataset
News Youtube Uzbek Speech Dataset
Dataset Description
This dataset contains audio clips and their corresponding transcriptions in the Uzbek language with differenent dialects. The data was collected from publicly available news videos on YouTube. It is designed for training and evaluating Automatic Speech Recognition (ASR) models.
Most of the content comes from the Kunuz, Qalampir YouTube channels. The data was transcribed using Gemini 2.5 Pro and was intelligently… See the full description on the dataset page: https://huggingface.co/datasets/islomov/news_youtube_uzbek_speech_dataset.youtube-persian-asrThis dataset consists of over 385 hours of audio extracted from various YouTube videos in the Persian language.
Note: This dataset contains raw, unvalidated transcriptions. Users are advised to:
1. Perform their own quality assessment
2. Create their own train/validation/test splits based on their specific needs
3. Validate a subset of the data if needed for their use caseChildTalk
ChildTalk: A Multi-Dialect Chinese Child Speech Corpus with Full-Length Child–Caregiver Conversations for Speech Recognition
📖 Overview
ChildTalk is a large-scale, publicly available multi-dialect Chinese child speech dialogue dataset. This dataset solves key problems found in existing Chinese child ASR corpora — mainly their small size, lack of natural conversations, limited dialectal coverage, and missing full-length dialogue recordings. It provides a solid… See the full description on the dataset page: https://huggingface.co/datasets/yujie-ovo/ChildTalk.yoruba-speech-text-parallel
Yoruba Speech-Text Parallel Dataset
Dataset Description
This dataset contains 1647022 parallel speech-text pairs for Yoruba, a language spoken primarily in Nigeria and other West African countries. The dataset consists of audio recordings paired with their corresponding text transcriptions, making it suitable for automatic speech recognition (ASR) and text-to-speech (TTS) tasks.
Dataset Summary
Language: Yoruba - yo
Task: Speech Recognition, Text-to-Speech… See the full description on the dataset page: https://huggingface.co/datasets/michsethowusu/yoruba-speech-text-parallel.youtube_transcriptions
Dataset Description
A speech dataset of Uzbek language audio clips sourced from YouTube videos. Audio segments were extracted, separated by speaker using vocal isolation, and transcribed using Google's Gemini 2.0 Flash model. Speaker identities were clustered using ECAPA-TDNN embeddings.
Use Cases
Automatic Speech Recognition (ASR) for Uzbek
Text-to-Speech (TTS) synthesis for Uzbek
Fine-tuning speech models on Uzbek language data (e.g., Qwen3-TTS)
Speaker-conditioned TTS… See the full description on the dataset page: https://huggingface.co/datasets/openbank-uz/youtube_transcriptions.yodas3
YODAS v3
Paper
YODAS v3 is a large web-crawled dataset containing over 1.1 million hours of audio that were originally released under a CC-BY-3.0 license. The dataset contains audio in over 100 languages. YODAS v3 can be used for a variety of multi-modal tasks, including Automatic Speech Recognition, Text-to-Speech, and Audio Representation Learning. We crawl a distinct set of videos from the v1 and v2 versions of YODAS, to guarantee that there are no overlaps in the data.
For… See the full description on the dataset page: https://huggingface.co/datasets/espnet/yodas3.YouTube-Cantonese
Cantonese Audio Dataset from YouTube
This dataset contains Cantonese audio segments and creator uploaded transcripts (likely higher quality) extracted from various YouTube channels, along with corresponding transcript metadata. The data is intended for training automatic speech recognition (ASR) models.
Data Source and Processing
The data was obtained through the following process:
Download: Audio (.m4a) and available Cantonese subtitles (.srt for zh-TW, zh-HK, zh-Hant)… See the full description on the dataset page: https://huggingface.co/datasets/OrcinusOrca/YouTube-Cantonese.vi-asr-youtube-1582h
vi-asr-youtube-1693h
Bo du lieu ASR tieng Viet cat tu audio YouTube. Nhan sinh boi MAI-Transcribe-2
(Azure Speech) voi timestamp muc tu, loc bang mot model Zipformer doc lap.
So doan
694,936
Tong thoi luong
1,693.4 gio
Dinh dang
MP3 64 kbps, 16 kHz, mono
Video nguon
7,683
Kenh
13
Loai cat
So doan
Gio
Do dai TB
Muc dich
long
493,045
1,439.9
10.5 s
Doc dai lien tuc
short
201,891
253.6
4.5 s
Dictation, cau ngan
Cau truc… See the full description on the dataset page: https://huggingface.co/datasets/anhtunguyen98/vi-asr-youtube-1582h.OCW
MIT OpenCourseWare dataset
MIT OpenCourseWare dataset which consists of speech and its corresponding transcript.