datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
common-voice-scripted-speech-26
Common Voice Scripted Speech
A row-normalized multilingual ASR dataset built from Mozilla Data Collective
Common Voice Scripted Speech. Each upstream archive is converted to appendable
parquet shards under data/<upstream_split>/, one shard per source archive and
split, with audio bytes embedded in an audio struct column.
Status
Manifest languages: 60
Languages uploaded: 18
Columns
audio (bytes, path)
sentence, locale, language, upstream_split… See the full description on the dataset page: https://huggingface.co/datasets/Peacockery/common-voice-scripted-speech-26.UniST
UniST
This dataset contains UniST codec-token training data exported from local metadata and codec results.
We train UniSS with UniST data.
Schema
id: sample identifier
transcription: source transcription from metadata text
translation: qwen_trans, falling back to trans_text
source_glm, target_glm: GLM token lists
source_bicodec, target_bicodec: bicodec semantic token lists
bicodec_global: source bicodec global token list
dataset_name, src_lang, tgt_lang, split:… See the full description on the dataset page: https://huggingface.co/datasets/cmots/UniST.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.voice-code-bench
VoiceCodeBench
VoiceCodeBench is a test-only benchmark for evaluating whether automatic
speech recognition (ASR) systems preserve exact structured values in English
workplace speech.
Paper: VoiceCodeBench: Evaluating Exact Structured-Token Recovery in Automatic Speech Recognition
The benchmark targets cases where a transcript is software input: callback
numbers, email addresses, command-line flags, file paths, URLs, account
identifiers, dates, measurements, and similar values… See the full description on the dataset page: https://huggingface.co/datasets/besimple-ai/voice-code-bench.Earnings22-Cleaned-AA-chunked
Earnings22-Cleaned-AA-chunked
Quick links: AA Streaming Speech to Text Leaderboard | Speech to Text methodology
Earnings22-Cleaned-AA-chunked is a chunked version of Earnings22-Cleaned-AA, the cleaned Earnings-22 subset used by Artificial Analysis for streaming Speech to Text evaluation.
The original Earnings-22 data comes from esb/datasets, a corpus of corporate earnings calls. Artificial Analysis manually reviewed and corrected the reference transcripts in the cleaned subset… See the full description on the dataset page: https://huggingface.co/datasets/ArtificialAnalysis/Earnings22-Cleaned-AA-chunked.darija-asr-corpus
Darija ASR Corpus (dataset-core)
Arabizi (Latin-script) transcriptions of Moroccan Darija speech, produced for a
Whisper fine-tuning pipeline (paper not yet published -- citation forthcoming).
This repo contains four source subsets: DODa, DVoice, Wiki, and
YouTube. Each subset carries its own upstream license/terms -- see below --
because they are drawn from four different original projects.
Subsets
Config
Rows
Audio bundled?
Upstream license
Upstream source… See the full description on the dataset page: https://huggingface.co/datasets/abnajlae/darija-asr-corpus.Inkling-Small-Multimodal-Calibration
Inkling-Small Multimodal Calibration
The exact 1,663 samples used for BF16 routed-expert importance collection
for Inkling-Small Mixed Quant GGUF.
This is calibration material, not a held-out evaluation benchmark.
The primary balanced pass is:
Category
Samples
Valid decoder tokens
Share
Text / reasoning
462
471,858
44.976%
Code / tool-oriented source text
205
209,715
19.989%
Real image / document
486
262,476
25.018%
Real speech audio
309
105,080
10.016%
Total… See the full description on the dataset page: https://huggingface.co/datasets/Baekpica/Inkling-Small-Multimodal-Calibration.Indonesian-ASR-11-Class-Dataset
Indonesian ASR 11-Class Dataset
Public Hugging Face repository for an Indonesian ASR corpus and its paper-supporting benchmark artifacts.
Dataset summary
Audio files: 104,500 WAV files
Real/human recordings: 104,368
Synthetic repair files: 132
Sentence classes: 11 Indonesian sentence categories
Canonical balanced sentence slots: 209 (11 categories × 19 retained slots)
Public speaker labels: M1..M12, F1..F8, plus synthetic labels Ms*/Fs*
Audio format: 16 kHz… See the full description on the dataset page: https://huggingface.co/datasets/Atika88/Indonesian-ASR-11-Class-Dataset.Neapolitan-Spoken-Corpus
Neapolitan Spoken Corpus (NSC)
A corpus of read Neapolitan speech for ASR evaluation, with a validated
Neapolitan–Italian lexicon, LOSO fine-tuning splits, trained LoRA adapters,
metric implementations, per-clip results, and error annotations.
This release supersedes the earlier 141-clip single-speaker version of this
repository. The earlier release corresponds to Speaker S1 of the present
corpus; the old audioData/ and transcripts.csv are replaced by
data/audio/ and… See the full description on the dataset page: https://huggingface.co/datasets/anonymous-nsc-author/Neapolitan-Spoken-Corpus.librivox-tracks-vad
librivox-tracks-vad
This dataset is produced from pykeio/librivox-tracks with single-reader filtering and Silero VAD segmentation.
data/train-*.parquet: all utterances (split is always train), collected until a global total audio budget is reached (see run manifest / script args: (2442/5994)*3600 * multiplier seconds by default).
Each row stores source metadata plus a mono WAV payload (audio_bytes) and sampling_rate.
chinese-lips-speech-slide-probe
Chinese-LiPS Speech + Slide Probe
A self-contained probe set for testing whether visual slide context helps
simultaneous speech translation — with the input as audio, not transcripts.
Why audio matters: feeding a transcript to a text LLM deletes the acoustic
ambiguity (homophones, polysemy) that slide context is meant to resolve; the
transcript already commits to one reading. Any honest test of "does vision help
streaming ST" must consume speech.
Contents… See the full description on the dataset page: https://huggingface.co/datasets/gavinlaw/chinese-lips-speech-slide-probe.MM-ContextASR-Bench
MM-ContextASR Bench
Metadata and evaluation splits for Multimodal Conversational Context for
LLM-Based ASR: Data Construction, Training, and Benchmark.
Dataset summary
Config
Examples
Audio
Context
Primary metric
mm_contextasr
1,250 (250 current utterances × 5 histories)
1,439 WAV files included
Controlled user-assistant dialogue
entity Recall
kespeech
19,212
Source ID only
Same-speaker speech and transcript
CER, SER, entity Recall
cv_yue
3,525… See the full description on the dataset page: https://huggingface.co/datasets/lilonghao/MM-ContextASR-Bench.common_voice_26_0_de
Mozilla Common Voice 26.0 - German (IPA & Clean Validated Subset)
Repacking version of Common Voice 26.0 German officialy published by Mozilla Data Collective, following Hugging Face Parquet Shards standard, with feature for listening to audio directly on the Web Hub, and the addition of a data column for the IPA transcription of each sentence.
📊 Dataset parameters
Origin: Mozilla Common Voice 26.0 (version 18/06/2026).
Data amount (Validated): 950,877 MP3 audio… See the full description on the dataset page: https://huggingface.co/datasets/q1805/common_voice_26_0_de.transcription-corpus
UN Transcription Corpus
Two splits of UN meeting audio paired with official verbatim records.
Splits
sessions — Whole meeting sessions (SC + GA plenary)
One row per meeting. Audio from UN Web TV, verbatim records from documents.un.org.
Column
Description
symbol
UN document symbol, e.g. S/PV.9826
webtv_url
URL on UN Web TV
duration_ms
Session duration in milliseconds
num_speakers
Number of speaker turns in the verbatim record
audio_floor
Floor… See the full description on the dataset page: https://huggingface.co/datasets/united-nations/transcription-corpus.cv-corpus-25.0-ja
Mozilla Common Voice 25.0 - Japanese Test Set (Complete)
Dataset Description
Complete Japanese test set from Mozilla Common Voice Corpus 25.0. This dataset contains all 9,019 validated test samples, compared to the partial 2,334-sample version previously available on HuggingFace.
Key Features
Size: 9,019 validated test utterances
Coverage: 100% of official Common Voice 25.0 Japanese test split
Multi-speaker: Diverse set of speakers with demographic metadata… See the full description on the dataset page: https://huggingface.co/datasets/FluidInference/cv-corpus-25.0-ja.clipquill-asr-benchmark
Measuring whisper-tiny vs whisper-base in a browser tab
Word error rate, wall-clock timing, transfer size and peak memory for two
quantised Whisper tiers running entirely client-side in a real Chrome window,
with the scripts that produced every number.
If you are building an in-browser transcription page, the two results worth
knowing before you pick a model tier:
On clean synthetic audio the two tiers tie. If that is all you test, you
will conclude the tier does not matter… See the full description on the dataset page: https://huggingface.co/datasets/sophia8888/clipquill-asr-benchmark.voice-code-bench
VoiceCodeBench
VoiceCodeBench is a test-only benchmark for evaluating whether automatic
speech recognition (ASR) systems preserve exact structured values in English
workplace speech.
Paper: VoiceCodeBench: Evaluating Exact Structured-Token Recovery in Automatic Speech Recognition
The benchmark targets cases where a transcript is software input: callback
numbers, email addresses, command-line flags, file paths, URLs, account
identifiers, dates, measurements, and similar values… See the full description on the dataset page: https://huggingface.co/datasets/yunqi1766/voice-code-bench.preprocessed-whisper-btb-cv-cvad-wlga-ca-2607
Dataset Card
Preprocessed Dataset: DewiBrynJones/preprocessed-whisper-btb-cv-cvad-wlga-ca-2607
Revision: main
Dataset Statistics
Train Split Statistics
Dataset
Revision
Split
Duration (HH:MM:SS)
Clips
Words
Words/Clip
%
DewiBrynJones/banc-trawsgrifiadau-bangor-2605
5bfe2d098c8486d97fac8be76d86ec9146435245
train
56:46:32
50,557
589,095
11.7
31.9
techiaith/corpws-clllc-wlga
5d00294c31c78b1d7937bb2c2bc6cc70bc18d410
clips
48:20:49
27,579… See the full description on the dataset page: https://huggingface.co/datasets/DewiBrynJones/preprocessed-whisper-btb-cv-cvad-wlga-ca-2607.candor-turntaking-annotations
CANDOR - Turn-Taking Annotations
Speech transcription and turn-taking annotation dataset built from the CANDOR corpus using NVIDIA Canary-Qwen2.5B ASR.
Dataset Description
This dataset contains 172,591 transcribed speech segments from the CANDOR conversational speech corpus (1,656 conversations). Each segment is a per-speaker utterance with Canary ASR transcript, designed for turn-taking prediction research.
Source
Audio corpus: CANDOR (English conversational… See the full description on the dataset page: https://huggingface.co/datasets/hiraki/candor-turntaking-annotations.pinga-fogo-chico-xavier
🎙️ Pinga-Fogo com Chico Xavier — TV Tupi, 1971
As duas entrevistas históricas do médium Chico Xavier, transmitidas ao vivo pela
TV Tupi em 1971, transcritas e estruturadas em turnos de fala com timestamp.
345 turnos (115 deles respostas do próprio Chico Xavier), a partir de
6 horas de áudio — o registro mais extenso do médium falando de improviso,
sem edição, diante de um painel de jornalistas.
Arquivos
Arquivo
Programa
Turnos
Respostas do Chico… See the full description on the dataset page: https://huggingface.co/datasets/ia-espirita/pinga-fogo-chico-xavier.ljspeech-mimi-codes
LJSpeech — Mimi Codes
Pre-extracted Kyutai Mimi neural-codec tokens for the
LJSpeech corpus — 13,100 English utterances
from a single female speaker reading public-domain audiobook passages (~24 hours).
This dataset contains codes only, not audio. For waveforms, go to the original LJSpeech
release; these codes are designed to be loaded alongside it for training Mimi-based speech
models without paying the ~1 hour of GPU extraction cost.
Schema
One row per utterance:… See the full description on the dataset page: https://huggingface.co/datasets/shangeth/ljspeech-mimi-codes.chinese-lips-longform-debug
Chinese-LiPS Long-Form (zh long streaming speech)
Reconstructed continuous long-speech streams from
BAAI/Chinese-LiPS, for
slide-aware / streaming speech-translation development and evaluation. Each
source video (one speaker, one scripted lecture with slides) was released as
pre-segmented clips; here they are re-joined into the full talk.
Two variants of the same 3 talks (~97 min speech total):
config
how segments are placed
use
orig_timeline
at their original session… See the full description on the dataset page: https://huggingface.co/datasets/gavinlaw/chinese-lips-longform-debug.Saudilang-Code-Switch-Corpus
SCC - Saudilang Code-Switch Corpus
The National Center for Artificial Intelligence at the Saudi Data and Artificial Intelligence Authority (SDAIA), published the "SCC" dataset, which stands for "Saudilang Code-Switch Corpus”.
This dataset contains a transcription of general conversations taken from a YouTube podcast "Thmanyah" that has been transcribed by the National Center for Artificial Intelligence in SDAIA. The data features three episodes covering different domains: investment… See the full description on the dataset page: https://huggingface.co/datasets/SDAIANCAI/Saudilang-Code-Switch-Corpus.apple-speechanalyzer-vs-whisper-cpp-mac
Apple SpeechAnalyzer vs whisper.cpp on Mac
Four complete speech-recognition benchmark runs over the same deterministic
40-speaker LibriSpeech test-clean snapshot:
Engine
Model path
WER
CER
Repeated median post-speech latency
Repeated p95
Apple SpeechAnalyzer
progressiveTranscription on macOS 26.5
1.98%
1.02%
125–132 ms
194–201 ms
whisper.cpp server
1.8.4 · ggml-small.en
4.28%
1.79%
122–125 ms
152–161 ms
Every run completed 40/40 clips with no failures. Accuracy… See the full description on the dataset page: https://huggingface.co/datasets/researchaudio/apple-speechanalyzer-vs-whisper-cpp-mac.librispeech-mimi-codes
LibriSpeech — Mimi Codes
Pre-extracted Kyutai Mimi neural-codec tokens for the
LibriSpeech corpus — multi-speaker English audiobook readings
from the LibriVox project.
This dataset contains codes only, not audio. For waveforms, use any of the LibriSpeech
mirrors (e.g. openslr/librispeech_asr);
these codes let you skip the ~hours of GPU extraction needed to train Mimi-based speech models.
Schema
One row per utterance:
Column
Type
Notes
id
string… See the full description on the dataset page: https://huggingface.co/datasets/shangeth/librispeech-mimi-codes.cm.trial
Dataset Card for Common Voice Corpus 11.0
Dataset Summary
The Common Voice dataset consists of a unique MP3 and corresponding text file.
Many of the 24210 recorded hours in the dataset also include demographic metadata like age, sex, and accent
that can help improve the accuracy of speech recognition engines.
The dataset currently consists of 16413 validated hours in 100 languages, but more voices and languages are always added.
Take a look at the Languages page to… See the full description on the dataset page: https://huggingface.co/datasets/taqwa92/cm.trial.mls-mimi-codes
Multilingual LibriSpeech (MLS) — Mimi Codes
Pre-extracted Kyutai Mimi neural-codec tokens
for Multilingual LibriSpeech —
LibriVox audiobooks in 7 non-English languages.
English is intentionally excluded. For English Mimi codes, use:
shangeth/librispeech-mimi-codes — LibriSpeech (~280k rows, 7 splits)
shangeth/libritts-r-mimi-codes — LibriTTS-R (~360k rows, 7 splits, 24 kHz native)
shangeth/vctk-mimi-codes — VCTK (~44k rows, 110 speakers w/ accents)
shangeth/jenny-mimi-codes — Jenny… See the full description on the dataset page: https://huggingface.co/datasets/shangeth/mls-mimi-codes.samromur_children_test
Dataset Card for samromur_children
Dataset Summary
The Samrómur Children Corpus consists of audio recordings and metadata files containing prompts read by the participants. It contains more than 137000 validated speech-recordings uttered by Icelandic children.
The corpus is a result of the crowd-sourcing effort run by the Language and Voice Lab (LVL) at the Reykjavik University, in cooperation with Almannarómur, Center for Language Technology. The recording process has… See the full description on the dataset page: https://huggingface.co/datasets/Ericwang/samromur_children_test.crowd-recital-yi
About
This dataset was created by crowd-sourced recording sessions in Yiddish as part of the ivrit.ai Crowd Recital project.
Volunteers read on normal desktop or mobile setting Wikipedia articles while time-stamping every sentence read.
Later this data is normalized by aligning the gathered captions with the audio using Stable Whisper (See Below).
The recording project is an ongoing effort and new data will be appended to this dataset periodically as it is being generated.… See the full description on the dataset page: https://huggingface.co/datasets/ivrit-ai/crowd-recital-yi.spite-CV16-TP9B
Spite Dataset
Pseudolabeled speech translation data with quality annotations from multiple metrics. This version uses transcripts from Common Voice 16.1 and translations from Tower-Plus-9B.
Configs
en_de
en_es
en_fr
en_it
en_ko
en_nl
en_pt
en_ru
en_zh
Usage
from datasets import load_dataset
ds = load_dataset("bpop/spite-CV16-Euro9B", "en_pt")