datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
ContextASR-Bench
ContextASR-Bench: A Massive Contextual Speech Recognition Benchmark
Automatic Speech Recognition (ASR) has been extensively investigated, yet prior benchmarks have largely focused on assessing the acoustic robustness of ASR models, leaving evaluations of their linguistic capabilities relatively underexplored. This largely stems from the limited parameter sizes and training corpora of conventional ASR models, leaving them with insufficient world knowledge, which is crucial for… See the full description on the dataset page: https://huggingface.co/datasets/MrSupW/ContextASR-Bench.Agri_STT_Benchmarking_Dataset
Agri STT Benchmarking Dataset
10,808 farmer voice queries in Hindi, Telugu and Odia, with reference transcripts, for benchmarking automatic speech recognition in agricultural contexts. The audio is included in this repository.
Every recording is a smallholder farmer speaking a question to Farmer.Chat, an AI advisory service run by Digital Green. Reference transcripts were produced by human annotators. Nothing here is read from a script or recorded in a studio, so the audio… See the full description on the dataset page: https://huggingface.co/datasets/DigiGreen/Agri_STT_Benchmarking_Dataset.thai-aligner-bench
Thai Aligner Bench
🚧 Development in progress.
How accurately can a forced aligner place Thai token and word boundaries in
speech? This is a self-contained benchmark: one Python file
(aligner_bench.py) plus 1,572 clips of Thai speech with frame-exact timing
ground truth. No Thai NLP stack or other code is needed — just
numpy soundfile torch torchaudio transformers.
The ground truth is what makes the dataset useful: the audio was rendered by a
TTS model whose duration predictor… See the full description on the dataset page: https://huggingface.co/datasets/wayu-ai/thai-aligner-bench.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.amharic-asr-benchmark
Amharic ASR Benchmark
An evaluation of open speech recognition models for Amharic, on a test set with
certain labels and honest statistics.
16 models. 1,548 clips. 4.72 hours. Every hypothesis published.
Published by Dataset.ET.
Read this table first
Round 1 of this benchmark rested on a single clean claim: every model predated
our dataset, so none could have trained on it. That claim no longer holds.
Models trained on snapwre/amharic-speech now exist, and others… See the full description on the dataset page: https://huggingface.co/datasets/snapwre/amharic-asr-benchmark.a5sv2-asr-benchmark-dataset
A5Sv2 ASR Benchmark Dataset
Public references, saved predictions, scores, and provenance for the
A5Sv2 ASR benchmark. The benchmark evaluates
streaming English ASR on four fixed public corpora with approximately equal normalized reference
word counts.
Corpus
Fixed selection
Reference words
Audio in this repository
Mega-ASR / Voices-in-the-Wild-2M
1,250 utterances, 250 per acoustic condition
32,928
Yes
AMI
7 scenario-only unseen-evaluation meetings
32,928
Yes
DiPCo… See the full description on the dataset page: https://huggingface.co/datasets/AirCaps/a5sv2-asr-benchmark-dataset.quranic-asr-benchmark
Quranic ASR Benchmark - leakage-free, held-out
A small, leakage-free benchmark (600 clips) for evaluating Arabic ASR on Quranic recitation
(Hafs riwayah). Every clip is verified absent from our training data, so it measures
generalization, not memorization. Same clips + same scoring for every model.
📊 Live leaderboard: https://huggingface.co/spaces/Muno459/quranic-asr-leaderboard
The set (600 clips, 200 per source)
Source
n
What it is
everyayah_heldout… See the full description on the dataset page: https://huggingface.co/datasets/Quran-Lab/quranic-asr-benchmark.persian-accents-benchmark
Persian Accents Benchmark
Dataset Summary
A benchmark for Persian automatic speech recognition (ASR): 279 short
utterances of informal Persian (Farsi) dialect speech across 16 regional accents,
released as a fixed evaluation set. Total audio duration is approximately 4.4
hours. The primary label is the transcription; each utterance also carries an
accent label (usable for accent classification as a secondary task) and an
emotion label as auxiliary metadata.
This… See the full description on the dataset page: https://huggingface.co/datasets/MR3z4/persian-accents-benchmark.thai-contextasr-bench
Thai Contextual-Biasing ASR Benchmark
TL;DR
Does your Thai ASR system actually use the context you give it (e.g., a list of names, custom words from your own dictionary) — and does it hallucinate when the context is irrelevant?
Each utterance comes with a bias list: entity strings (brands, person names,
places) that may or may not be spoken in the audio, written the way a real Thai user
would write them — one list, mixed Thai and Latin script. A good system does… See the full description on the dataset page: https://huggingface.co/datasets/wayu-ai/thai-contextasr-bench.entity-transcription-benchmark
Entity Transcription Benchmark
Measures whether a speech recognition system transcribes named entities
correctly — as distinct from word error rate.
WER weights every token equally. The tokens that matter for redaction, lookup,
routing and search are proper nouns, and they are a small fraction of any
transcript. A system can improve WER while getting worse at exactly the words a
downstream consumer needs, and nothing in the standard evaluation will show it.
2,151 clips, 6.0… See the full description on the dataset page: https://huggingface.co/datasets/modulate/entity-transcription-benchmark.stt-unified-bench
am-pranav/stt-unified-bench
Private, language/locale-partitioned mini-benchmark for STT models.
Each subset is a dataset config (e.g., en, de, en_indian_accent, hi_in) with a single split val.
Audio is staged at 16 kHz and stored in-repo for reproducibility.
Schema
audio : Audio(sampling_rate=16000, decode=False)
text : reference transcription
lang : implied by dataset config name
source : upstream dataset tag
id : source-stable id
⚠️ For internal evaluation only.… See the full description on the dataset page: https://huggingface.co/datasets/am-pranav/stt-unified-bench.multilingual-tts-benchmark
Multilingual Speech Benchmark for Zero-Shot TTS
A voice-cloning and intelligibility benchmark for six language variants, built
from Common Voice 17.0 by coverage-driven selection rather than random sampling.
Every example pairs a reference clip of one speaker with a target text that
speaker never read, so a system is asked to clone a voice and produce new
speech, which is what zero-shot TTS is actually for.
Pipeline source code:… See the full description on the dataset page: https://huggingface.co/datasets/nineninesix/multilingual-tts-benchmark.TiCo-Bench
TiCo-Bench
Time-Controlled speech response benchmark. Given a spoken input and a
duration constraint stated in the instruction, the model must produce a spoken
response whose length matches the constraint.
Usage
from datasets import load_dataset
# config = task, split = duration regime
ds = load_dataset("WeiChihChen/TiCo-Bench", "qa", split="short")
ds = load_dataset("WeiChihChen/TiCo-Bench", "qa", split="long")
ds = load_dataset("WeiChihChen/TiCo-Bench", "reasoning"… See the full description on the dataset page: https://huggingface.co/datasets/WeiChihChen/TiCo-Bench.quran-alignment-benchmark
Quran Recitation Alignment Benchmark
Audio recordings of Quran recitation with a reviewed word-level ground truth: every recited word, in the order it was recited, with its start and end time, plus the reviewed segmentation and non-Quran regions. This is the corpus behind the Quran Recitation Alignment Benchmark; the task, scoring rules, leaderboard and submission format are documented there, not here.
16 recordings · 357 minutes · 18,421 recited words · Hafs ʿan ʿĀṣim ·… See the full description on the dataset page: https://huggingface.co/datasets/QUD-Technologies/quran-alignment-benchmark.commonvoice_benchmark_catalan_accentsA new presentation of the corpus Catalan Common Voice v17.0 - metadata annotated version with the splits redefined to benchmark ASR models with various Catalan accentPersian-ASR-BenchmarkThis dataset consists of 3 hours of 16kHz audio collected from diverse environments to better represent real-world scenarios. The recordings were sourced from audiobooks, YouTube, and other public sources, ensuring a wide variety of speech styles and acoustic conditions.
One key advantage of this dataset is that it was collected from recent sources within the last few months, ensuring no overlap with training data and fairness for evaluating other STT models.
To enable a robust and fair… See the full description on the dataset page: https://huggingface.co/datasets/C1Tech/Persian-ASR-Benchmark.pavo-bench
PAVO-Bench: 50K-Turn Benchmark for ASR-LLM-TTS Pipeline Routing
Code: github.com/vnmoorthy/pavo-bench · Paper: TMLR 2026 (accepted) · Authors: NarasingaMoorthy VeiluKanthaPerumal (UPenn), Mohammed Imthathullah (Google)
pip install git+https://github.com/vnmoorthy/pavo-bench.git
Headline results (vs fixed-cloud baseline, 50,000 voice turns)
Metric
Result
Significance
P95 end-to-end latency (H100, LibriSpeech)
−10.3% (−167 ms)
—
Median latency
−34%… See the full description on the dataset page: https://huggingface.co/datasets/vnmoorthy/pavo-bench.LCAR-Hallucination-Benchmark
LCAR Hallucination Benchmark
LCAR Hallucination Benchmark is a manually reviewed speech benchmark for
studying acoustic-grounding failures in LLM-based ASR. It contains two
500-utterance suites: controlled speech synthesized with IndexTTS2 and speech
derived from openly released corpora. The benchmark covers translation or
transliteration, spoken or text-prompt instruction execution, unsupported
repetition, and catastrophic deletion.
The benchmark is a targeted stress set. It is… See the full description on the dataset page: https://huggingface.co/datasets/aguangguang/LCAR-Hallucination-Benchmark.Voices-in-the-Wild-Bench
Voices-in-the-Wild-Bench
Voices-in-the-Wild-Bench is a bilingual benchmark for evaluating robust automatic speech recognition under real-world acoustic conditions. It contains 5,000 speech examples covering Chinese and English, synthetic and real-recorded speech, and eight acoustic scenarios.
Dataset Composition
Group
Count
Synthetic speech
3,500
Real-recorded speech
1,500
Chinese
2,500
English
2,500
Total
5,000
Splits
Splits are… See the full description on the dataset page: https://huggingface.co/datasets/zhifeixie/Voices-in-the-Wild-Bench.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.Vaani-Benchmark-V1.0
Vaani-Benchmark-V1.0
A curated ASR evaluation set drawn from the Vaani project. This benchmark contains 5,050 audio segments from 1,103 speakers across 104 Indian districts, each with three independent human transcriptions.
Evaluation Toolkit
A standalone toolkit implementing this benchmark's scoring methodology, plus
Latin-script normalization for code-switched predictions and one-command
publishing of results to a model's HF card, is available at… See the full description on the dataset page: https://huggingface.co/datasets/ARTPARK-IISc/Vaani-Benchmark-V1.0.an-be-kalan-bench
Bambara Educational Speech Dataset
This dataset is a collection of READ Bambara text based on educational children's books from RobotsMali's GAIFE project. It is designed to support the training and benchmarking of Automatic Speech Recognition (ASR) models, with a particular focus on child speech, regional acoustics, and repetitive text structures (inherent to the domain).
The dataset is structured into two separate subsets to support specialized training and evaluation… See the full description on the dataset page: https://huggingface.co/datasets/RobotsMali/an-be-kalan-bench.script-fidelity-benchmark
Script fidelity benchmark
Anonymous supplement for the paper "Script collapse in multilingual ASR:
A reference-free metric and 100-pair benchmark."
Script Fidelity Rate (SFR) measures the fraction of ASR hypothesis characters
that belong to the expected target script. WER measures word edits, while SFR
checks whether the output is written in the target orthography.
Related resources:
PyPI package: https://pypi.org/project/script-fidelity/
Hugging Face Evaluate metric:… See the full description on the dataset page: https://huggingface.co/datasets/themechanism/script-fidelity-benchmark.small-overlapping-speech-bench
Small Overlapping Speech Bench
A tiny, fully-reproducible benchmark for multilingual overlapping speech. Each of the
100 clips contains three people speaking at the same time, each in a different European
language, with ground-truth per-speaker timestamps, languages, and transcripts.
It is a deliberately hard "cocktail-party" stress test: how much of each simultaneous speaker can
an ASR (speech-to-text) model recover, and can a model tell how many people are talking?
100 clips… See the full description on the dataset page: https://huggingface.co/datasets/laion/small-overlapping-speech-bench.WESR-Bench
WESR-Bench
WESR-Bench is an expert-annotated natural speech dataset with word-level non-verbal vocal events, featuring both discrete events (standalone, denoted as [tag]) and continuous events (mixed with speech, denoted as <tag>...</tag>).
Supported Tags
Discrete events (15):
inhale, cough, laughs, laughing, crowd_laughter, chuckle, shout, sobbing, cry, giggle,exhale, sigh, clear_throat, roar, scream, breathing
Continuous events (6):
crying, laughing, panting… See the full description on the dataset page: https://huggingface.co/datasets/yfish/WESR-Bench.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.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.FinASR-Bench
FinASR-Bench
FinASR-Bench is a multilingual benchmark for evaluating and adapting automatic speech recognition (ASR) systems in the financial domain. It contains both real-world financial speech and synthetically generated financial speech in three languages:
English (EN)
Chinese (ZH)
Japanese (JA)
Unlike conventional ASR benchmarks that mainly evaluate transcription accuracy using WER or CER, FinASR-Bench is designed to support evaluation of financially critical recognition… See the full description on the dataset page: https://huggingface.co/datasets/wuxianchao/FinASR-Bench.product-bench
Product Bench
31-turn multi-turn speech-to-speech benchmark for evaluating voice AI models as a laptop comparison shopping assistant.
Part of Audio Arena, a suite of 6 benchmarks spanning 221 turns across different domains. Built by Arcada Labs.
Leaderboard | GitHub | All Benchmarks
Dataset Description
The model acts as a laptop comparison shopping assistant helping a customer evaluate, compare, and order laptops. The conversation features multi-intent turns… See the full description on the dataset page: https://huggingface.co/datasets/arcada-labs/product-bench.Fasih-TTS-Benchmark
Fasih-TTS-Benchmark
Evaluation audio and objective scores for the
Fasih-TTS-V1 Arabic (MSA / Fusha)
text-to-speech model. Every clip is the model's own output, paired with its reference text, its
Whisper-large-v3 transcription, and per-clip WER / CER.
Contents
Split (test_set)
Clips
Purpose
Mean CER
silma_msa
10
SILMA open-source Arabic TTS benchmark (MSA sentences)
2.0%
samples
3
General showcase (greeting, fiqh, reflection)
0.6%
consistency
8… See the full description on the dataset page: https://huggingface.co/datasets/NightPrince/Fasih-TTS-Benchmark.