GigaSpeechBench — Older Adults
SpeechColab
This benchmark provides Chinese and English older-adult speech with segment-level reference transcripts.
Mandarin, English2 test configurations
View detailsResearch resources
Frank curates this directory of publicly available speech and audio datasets. Access is free and requires no registration or application. Open a dataset’s details and select “Visit official source” for release information and access.
110 datasets
SpeechColab
This benchmark provides Chinese and English older-adult speech with segment-level reference transcripts.
Mandarin, English2 test configurations
View detailsBrigham Young University Matrix Lab and Cambodian Oral History Project
Khmer oral-history interviews pair natural long-form speech with transcripts collected by the Cambodian Oral History Project.
Khmer640 interviews, 356.6 hours of interview audio and 341.4 hours of aligned segments
View detailsMLLP research group at Universitat Politècnica de València
English talks from Large Hadron Collider Physics conferences pair recorded presentations with reference transcripts.
English13,755 segments across two test splits
View detailsHypaAI and AfroVoices
This compact evaluation set contains verbatim reference text for recordings across widely spoken and low-resource languages.
English, French, Spanish +8200 audio and transcript pairs
View detailsSpeechColab
This module benchmarks language and dialect coverage, including Southeast Asian languages and regional Arabic varieties.
Arabic, Indonesian, Malay +514 test configurations across language and regional varieties
View detailsSpeechColab
The Chinese and English medical subsets support recognition evaluation on domain-specific speech and terminology.
Mandarin, English2 medical-domain test configurations
View detailsQualiaLabs and ASLP Lab
DuplexConv provides large-scale Mandarin tutoring and casual conversation with model-assisted annotations of interaction events.
Mandarin2,000.21 hours across 93,709 files
View detailsHKUST Audio
This challenge corpus is designed for expressive synthesis that draws on surrounding dialogue and context annotations.
Mandarin, EnglishAbout 16,000 hours across 3 million segments
View detailsQualiaLabs and ASLP Lab
SmoothConv pairs naturally occurring Mandarin dialogue with expert annotations of turn events and conversational context.
Mandarin100.53 hours across 2,503 files
View detailsTrelis Research
This small synthetic benchmark pairs generated speech with labeled drug, condition and procedure terms.
English50 synthetic test utterances with labeled target terms
View detailsAalto University corpus, packaged by the GAIK project
This release packages Finnish Parliament recordings for recognition evaluation, including subsets with speakers seen and unseen during original training.
Finnish6,307 WAV and transcript pairs across five evaluation subsets
View detailsGoogle Research and WAXAL collection partners
WAXAL brings together natural speech recordings from African language communities. Its labeled ASR splits pair audio with same-language text.
Acholi, Akan, Amharic +16About 1,250 hours of transcribed speech across 19 languages
View detailsMeta FAIR
This Meta corpus contains transcribed spontaneous speech collected for languages that have limited representation in ASR research.
Multilingual3,350 hours across 348 languages
View detailsTaiwan Administration for Digital Industries
The Hakka release pairs speech files with reference sentences and duration metadata for each recording.
Hakka2 train archives and one test archive
View detailsTaiwan Administration for Digital Industries
Taiwanese Mandarin recordings are provided with Traditional Chinese references and locale metadata.
Mandarin24 train archives and one test archive
View detailsTaiwan Administration for Digital Industries
The Taiwanese Hokkien release provides recordings with corresponding written sentence references.
Taiwanese Hokkien7 train archives and one test archive
View detailsASLP Lab and Easy Turn contributors
Easy-Turn-Trainset combines real and simulated interactions for learning when conversational systems should speak or listen.
Mandarin, EnglishAbout 1,145 hours of training data
View detailsEIT-NLP and LLaSO authors
LLaSO-Align groups English audio and text pairs for the alignment phase of a speech-language model.
EnglishAbout 12 million alignment samples
View detailsEIT-NLP and LLaSO authors
LLaSO-Instruct organizes speech-language examples into instruction-following tasks for the LLaSO framework.
EnglishAbout 13.5 million instruction samples
View detailsEkaCare
EkaCare combines spoken medical terms, narrated sentences and demonstration conversations in an Indian healthcare context.
English, Hindi3,939 test clips (3,619 English, 320 Hindi)
View detailsContextASR-Bench research team
This benchmark pairs generated English and Mandarin speech with contextual cues for named entities across multiple domains.
English, Mandarin41,329 test entries across four language and task subsets
View detailsHiACC research team
Adult and child speakers switch between Hindi and English while answering questions, reading stories and describing pictures.
Hindi, English5.24 hours, 3,318 adult utterances and 1,858 child utterances
View detailsAISHELL and AISHELL-5 research collaborators
AISHELL-5 documents multi-speaker conversations in a car, with far-field door microphones and close-talk recordings.
MandarinOver 100 hours of real multi-channel vehicle recordings
View detailsIntron Health and research collaborators
This conversational benchmark pairs African-accented English audio with transcripts and diarization labels. Its healthcare scenarios sit alongside non-medical conversations.
EnglishAbout 6 hours across 11 accents
View detailsDailyTalk authors and Kyutai
Two speakers record scripted English conversations derived from DailyDialog. The packaged audio places the speakers on separate stereo channels.
English2,541 recorded source dialogues
View detailsBAAI and Nankai University
CS-Dialogue records spontaneous conversations that switch between Mandarin and English within the same interaction.
Mandarin, English104.02 hours
View detailsMicrosoft Research
Far-field recordings of English meetings capture several talkers around a table using single-channel and multichannel devices.
English315 recorded meetings in the paper’s benchmark
View detailsMultiMed research team
MultiMed pairs medical-domain audio with transcriptions across five languages and multiple formats including interviews, talks and podcasts.
Vietnamese, English, German +248,369 entries across published configuration splits
View detailsWestlake University and AISHELL
RealMAN provides real recordings made with a microphone array for speech enhancement and source localization.
Mandarin83.7 hours of speech signals and 144.5 hours of background noise
View detailsFondazione Bruno Kessler and Speech-MASSIVE contributors
Speakers record short voice-assistant requests with matching prompt text, intent and slot labels.
Arabic, German, Spanish +912 languages with 115 few-shot recordings per language
View detailsAdam Mickiewicz University CAI
nEMO collects acted Polish speech labeled with emotional categories.
PolishOver 3 hours of annotated speech
View detailsKhai Le-Duc and collaborators
VietMed combines labeled Vietnamese medical speech with unlabeled medical and general-domain audio.
Vietnamese16 labeled medical hours, 966 unlabeled medical hours and 1,204 unlabeled general-domain hours
View detailsEARS research team
EARS captures clean full-band speech in an anechoic chamber, including reading, freeform monologues and deliberate changes in delivery.
English100 hours from 107 speakers
View detailsNational Taiwan University research team
Recordings from a National Taiwan University machine learning course pair Mandarin and English lecture speech with text. The corpus includes spoken switches into English technical terms.
Mandarin, English35,692 segments across train, development and test
View detailsLibriheavy authors and k2-fsa contributors
Libriheavy aligns Libri-Light audiobook audio with English text while retaining punctuation, capitalization and surrounding context.
EnglishAbout 50,000 hours
View detailsRescueSpeech research team
German speech from simulated search-and-rescue exercises is available with matched noise-augmented versions.
German2,412 utterances, 1 h 36 min of reference speech
View detailsOnline Resources for African American Language · University of Oregon
CORAAL pairs conversational recordings from several African American English communities with time-aligned orthographic transcriptions.
EnglishOver 220 sociolinguistic interviews
View detailsLibriTTS-R applies learned audio restoration to the corresponding recordings from LibriTTS.
EnglishAbout 585 hours
View detailsUniversity of Edinburgh CSTR
EdAcc records English video-call conversations between people who know each other, with diverse accent and language backgrounds.
EnglishApproximately 40 hours of conversations from 120 speakers
View detailsMeta AI Research
Studio recordings cover expressive English reading styles including speech delivered softly, rapidly or with laughter.
English11,615 rows in the published read configuration
View detailsMLCommons
The People’s Speech assembles transcribed English from sources selected for reuse-friendly licensing.
EnglishOver 30,000 hours
View detailsVISAI and Data Wow
Older Thai speakers read sentences about health symptoms and smart-home tasks.
Thai17 h 11 min, 19,200 recordings from 32 speakers
View detailsFLEURS records translated sentences across many languages to support comparable multilingual speech evaluation.
Multilingual, English, Mandarin +8102 languages with about 12 hours per language
View detailsNordisk Språkteknologi and the National Library of Norway
The Norwegian dictation corpus includes controlled recordings made with close and distant microphones.
Norwegian9 training and 3 test shards per microphone configuration
View detailsBabylon Health research team
Clinicians and actors recorded simulated primary care visits with speaker-specific audio and manually annotated transcripts.
English57 consultations, 8 hours 38 minutes
View detailsTampere University
This DCASE development set records urban acoustic scenes with reference and mobile devices.
Language unconfirmed64 hours across 10 acoustic scenes
View detailsHKUST Centre for Artificial Intelligence Research
Bilingual speakers in Hong Kong hold unscripted conversations across multiple turns. The released segments include Mandarin speech, English speech and switches between the two.
Mandarin, English10.62 hours, 12,314 utterances from 23 bilingual speakers
View detailsRev.com research team
Earnings-22 pairs long-form corporate earnings calls from multiple countries with verbatim reference transcripts.
EnglishAbout 119 hours of earnings calls
View detailsMagic Data Technology
MagicData-RAMC pairs mobile recordings of Mandarin conversations with transcripts and topic labels.
Mandarin180 hours
View detailsNorwegian Language Bank, National Library of Norway
Recorded debates from the Norwegian Parliament pair public speech with time-aligned Bokmål and Nynorsk transcriptions.
Norwegian Bokmål, Norwegian Nynorsk125.7 hours excluding pauses, 140.3 hours including pauses
View detailsAlibaba DAMO Academy
AliMeeting combines meeting-room microphone arrays and headset recordings with transcripts for multi-party Mandarin speech.
Mandarin118.75 hours
View detailsATCO2 consortium
Airline pilots and air traffic controllers communicate over live radio using short turns and domain-specific phraseology.
EnglishAbout 1 hour, 871 segments in the data-card version
View detailsAI4Bharat
Contributors in India read text in 12 languages for a multilingual speech benchmark.
Bengali, Gujarati, Hindi +91,684 hours from 1,218 contributors
View detailsInstitute of Smart Systems and Artificial Intelligence, Nazarbayev University
KSC2 expands earlier Kazakh corpora with recordings from radio, television, parliamentary proceedings and podcasts.
KazakhAbout 1,200 hours and more than 600,000 utterances
View detailsReykjavik University Language and Voice Lab and Almannarómur
Children and teenagers read Icelandic prompts for a crowdsourced speech collection.
IcelandicMore than 137,000 validated recordings, ages 4–17
View detailsBeijing AISHELL Technology
AISHELL-4 records Mandarin meetings with an eight-channel circular microphone array.
MandarinAbout 120 hours
View detailsNVIDIA
Hi-Fi TTS selects high-quality English audiobook recordings from a small number of speakers for synthesis research.
English291.6 hours from 10 speakers
View detailsFacebook AI Research
The transcribed VoxPopuli release pairs European Parliament speech with transcripts across several languages.
Multilingual, English, German +41,791 transcribed hours in the main language table
View detailsRev
Public corporate earnings calls cover multiple industries and contain names, financial figures and dates.
English44 earnings calls, totaling 39 h 15 min
View detailsMeta Reality Labs
EasyCom captures group conversation around a table with wearable microphones, headset audio and synchronized egocentric video.
English6.6 hours total, including about 5 h 18 min of quality-screened data
View detailsSberDevices
Russian speech was collected through crowdsourcing and recordings for far-field device use.
RussianAbout 1,240 hours across crowd and far-field domains
View detailsNanosemantics Lab
MediaSpeech pairs short clips from online media with transcripts in Arabic, French, Spanish and Turkish.
Arabic, French, Spanish +140 hours across 4 languages
View detailsMUCS 2021 challenge organizers
The challenge corpora capture English terms and phrases embedded in Hindi and Bengali technical tutorials.
Hindi, Bengali, EnglishHindi–English 89.86 h train and 5.18 h test. Bengali–English 46.11 h train and 7.02 h test.
View detailsMultilingual TEDx research team
Multilingual TEDx aligns talk recordings with source-language transcripts and, where available, translated sentences.
Spanish, French, Portuguese +58 source languages, with language-specific archives
View detailsSpeechocean and corpus authors
Mandarin-first-language speakers read English sentences. Pronunciation judgments cover sounds, words and complete utterances.
English5,000 sentences, with child and adult speakers
View detailsNicolingua research team
The corpus samples Guinean radio broadcasts with local languages, background music and telephone contributions.
French, Guerze, Koniaka +717,090 thirty-second clips from 6 Guinean radio stations
View detailsMoussa Doumbouya, Lisa Einstein and Chris Piech
Speakers record short assistant commands, names and digits in four languages using laptops and phones.
French, Maninka, Pular +110,083 utterances from 49 speakers aged 5–76
View detailsYoloxóchitl Mixtec ASR research team
Field recordings of Yoloxóchitl Mixtec include natural speech and a transcription system that represents lexical tone.
Yoloxóchitl MixtecAbout 100 hours of expert transcripts, plus 8.36 hours in the trainee correction study
View detailsFacebook AI Research
Multilingual LibriSpeech builds parallel dataset structure across eight languages from LibriVox recordings.
English, German, Dutch +5About 50,500 hours
View detailsSLURP research team
SLURP links spoken assistant requests to text and semantic annotations for intents and entities.
English18 domains, with separate real and synthetic audio
View detailsBeijing AISHELL Technology
AISHELL-3 provides neutral, multi-speaker Mandarin recordings with character and pinyin transcriptions.
MandarinAbout 85 hours from 218 speakers
View detailsCHiME Challenge organizers
CHiME-6 provides unscripted home conversations with wearable binaural microphones and multiple room arrays. The release corrects array synchronization from CHiME-5.
English20 fully transcribed dinner parties
View detailsVolunteer speakers recorded Peruvian Spanish sentences, which are distributed with utterance-level text.
Spanish2 WAV archives totaling about 1.9 GB
View detailsFreesound and Universitat Pompeu Fabra research team
FSD50K provides human-labeled sound events from Freesound, with clips that may carry several event labels.
Language unconfirmed51,197 clips, 108.3 hours and 200 sound classes
View detailsLibriCSS research team
LibriCSS records overlapping playback of LibriSpeech utterances through a seven-channel array in a meeting room.
English10 hours across 10 sessions
View detailsFacebook AI Research and collaborators
Libri-Light provides English audiobook speech for recognition with little or no labeled training data.
EnglishOver 60,000 hours of unlabeled audio
View detailsUniversity of Edinburgh CSTR
VCTK records English speakers with varied accents reading common and speaker-specific material for speech technology research.
English110 speakers, about 400 sentences per speaker (v0.92)
View detailsTsinghua University CSLT
CN-Celeb provides Chinese speaker recordings drawn from multiple media genres and recording conditions.
MandarinAbout 274 hours (original CN-Celeb1)
View detailsUniversity of Tokyo research team
JVS supports multi-speaker Japanese synthesis with recordings in normal, whispered and falsetto styles.
JapaneseAbout 30 hours
View detailsLibriTTS restructures audiobook recordings and text from LibriSpeech for multi-speaker English synthesis.
EnglishAbout 585 hours
View detailsThis corpus provides Bengali speech recordings with transcripts and anonymized speaker identifiers.
BengaliApproximately 196,000 utterances
View detailsGoogle, Reykjavik University and Universitas Gadjah Mada
This crowdsourced speech collection provides short Javanese recordings with utterance and speaker identifiers.
JavaneseAbout 185,000 utterances
View detailsThis crowdsourced speech collection provides short Sundanese recordings with utterance and speaker identifiers.
SundaneseAbout 220,000 utterances
View detailsSMART Lab, Toronto Metropolitan University
Professional actors recorded controlled speech and song performances with labeled emotions and intensity levels.
English1,440 speech-only audio files, plus song and video archives
View detailsSpeech Commands contains short recordings of keywords and other words for small-footprint recognition.
English105,829 one-second recordings across 35 words
View detailsLIUM at Le Mans University
Recorded TED talks are distributed as English speech segments with aligned reference text.
English452 hours of training speech in Release 3
View detailsZeroth Project
The Zeroth Project releases transcribed Korean recordings together with a pronunciation lexicon and text tools.
Korean51.6 hours for training and 1.2 hours for testing
View detailsUniversity of Tokyo research team
A Japanese reading corpus designed around everyday character readings, with recordings by a single speaker.
JapaneseAbout 10 hours
View detailsBeijing AISHELL Technology
AISHELL-1 pairs read Mandarin utterances with transcriptions and includes speakers from multiple accent regions.
MandarinOver 170 hours from 400 speakers
View detailsUniversity of Edinburgh CSTR
The Edinburgh release provides paired noisy and clean utterances, plus text for speech enhancement and noise-robust speech workflows.
English28- and 56-speaker training configurations, with a separate test set
View detailsOpenSLR SLR28 release team
The release combines measured and simulated room impulse responses with directional and diffuse noise.
Language unconfirmed1.3 GB archive of room responses and noise
View detailsTsinghua University CSLT
A compact collection of Mandarin read speech with aligned transcripts and companion resources for noise testing.
MandarinAbout 30 hours
View detailsKarol J. Piczak
ESC-50 organizes short environmental audio into classes spanning animals, everyday activities and urban sounds.
Language unconfirmed2,000 five-second clips across 50 classes
View detailsLibriSpeech authors / OpenSLR
LibriSpeech provides aligned English audiobook recordings and transcripts with standard training, validation and test splits.
EnglishAbout 1,000 hours
View detailsDavid Snyder, Guoguo Chen and Daniel Povey
MUSAN collects music, multilingual speech and a range of technical and everyday noise recordings.
Multilingual11 GB archive with music, speech and noise
View detailsSeoul Corpus research team
The Seoul Corpus pairs spontaneous Korean speech with phonetic annotation.
Korean2.5 GB of FLAC audio plus separate TextGrid annotations
View detailsNew York University research team
UrbanSound8K collects labeled excerpts of city sounds such as sirens, street music and road equipment.
Language unconfirmed8,732 clips across 10 sound classes
View detailsUniversity of Edinburgh
MATCH records younger and older adults as they interact with a Wizard-of-Oz spoken dialogue system.
English24 older and 22 younger participants (project description)
View detailsEUROCONTROL Experimental Centre and Graz University of Technology
Controllers recorded through headsets during air traffic management simulations speak English in an aviation setting.
EnglishAbout 10 hours, 10,078 utterances from 10 speakers
View detailsUniversity of California, Santa Barbara
Recordings of everyday American English include face-to-face conversations, telephone calls and storytelling.
English60 recordings, typically about 20 minutes each
View detailsInternational Computer Science Institute
ICSI provides natural research meetings with synchronized audio and time-aligned orthographic transcripts.
EnglishAbout 72 hours of speech across 75 meetings
View detailsAMI Project Consortium
AMI pairs synchronized meeting audio and annotations with close-talking and distant microphone recordings.
EnglishAbout 100 hours of meetings
View detailsUniversidade Federal de Minas Gerais research team
CML-TTS prepares multilingual LibriVox audiobook recordings for text-to-speech research in seven languages.
Multilingual, Dutch, German +53,233.43 hours across seven languages
View detailsThe public release pairs Burmese recordings from female speakers with a separate text index.
Burmese948 MB audio archive (publisher listing)
View detailsKyubyong Park
KSS collects recordings from one Korean speaker for supervised text-to-speech research.
Korean12,853 clips, over 12 hours
View detailsMagic Data Technology
MagicData’s read-speech corpus contains scripted Mandarin recorded by speakers from mainland China.
Mandarin755 hours from 1,080 speakers
View detailsShanghai Primewords
Primewords Chinese Corpus Set 1 contains transcribed Mandarin utterances recorded with mobile devices.
MandarinAbout 100 hours
View detailsSurfingtech
ST-CMDS provides short Mandarin utterances collected from hundreds of speakers for speech recognition.
Mandarin102,600 utterances from 855 speakers
View detailsKeith Ito and Linda Johnson
The LJ Speech Dataset segments one narrator’s readings from public-domain books into English audio and text pairs.
English13,100 clips, about 24 hours
View detailsUniversity of Toronto and research collaborators
TORGO pairs acoustic and articulatory recordings of speakers with dysarthria and a comparison group.
English8 speakers with dysarthria and 7 controls
View details