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.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.synthetic-speech-diarization-ru
synthetic-speech-diarization-ru
Synthetic speech diarization dataset in Parquet format.
Dataset Details
Number of tracks: 2000
Sampling rate: 16000 Hz
Audio format: Embedded in Parquet files (Audio feature compatible)
Storage: Parquet format for efficient loading
Dataset Structure
The dataset contains audio tracks with speaker diarization annotations, stored directly in Parquet format.
Features
audio: Audio waveform (Audio feature with array and… See the full description on the dataset page: https://huggingface.co/datasets/ivkond/synthetic-speech-diarization-ru.multilingual-speech
Multilingual Indian Conversational Speech
A dataset of naturalistic, spontaneous two-speaker conversations across
13 Indian languages, with segment-level transcripts, speaker profiles,
timestamps, and recording metadata. Designed for ASR, TTS, speaker
diarization, and conversational speech research.
Languages (13)
Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali,
Odia, Punjabi, Tamil, Telugu, Urdu.
Content
Conversations… See the full description on the dataset page: https://huggingface.co/datasets/eQOURSE/multilingual-speech.synthetic-speech-diarization-ru
synthetic-speech-diarization-ru
Synthetic speech diarization dataset in Parquet format.
Dataset Details
Number of tracks: 2000
Sampling rate: 16000 Hz
Audio format: Embedded in Parquet files (Audio feature compatible)
Storage: Parquet format for efficient loading
Dataset Structure
The dataset contains audio tracks with speaker diarization annotations, stored directly in Parquet format.
Features
audio: Audio waveform (Audio feature with array and… See the full description on the dataset page: https://huggingface.co/datasets/niobures/synthetic-speech-diarization-ru.pseudolabel-malaya-speech-stt-train-whisper-large-v3apple-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.audio-speech-realtime-voice-agents-2026
🎙️ Audio, Speech Foundation Models & Real-Time Voice Agents Dataset (2026 Edition)
A structured research dataset featuring 1,722 domain-verified research papers and 298 official code repositories focused on Full-Duplex Speech-to-Speech LLMs, Real-Time Voice Agents (<200ms Latency), Zero-Shot TTS, Voice Cloning, OpenAI Whisper-v3, Neural Audio Codecs (EnCodec/DAC/SNAC), and Generative Music (2023–2026).
Built with Universal Scientific Engine V18.1 Diamond, providing 48 schema… See the full description on the dataset page: https://huggingface.co/datasets/beatsprom/audio-speech-realtime-voice-agents-2026.asr-evaluationslost-in-speech
Lost in Speech
A trilingual benchmark for reference-free classification of synthetically introduced factual and contextual alterations in English, Russian, and Kazakh. It contains 12,013 samples derived from news articles, with text, synthesized speech, and ASR transcript representations used in the study.
Altered samples are LLM-generated rewrites with a controlled alteration type—contradiction, fabrication, or context inconsistency—and severity level—mild, moderate, or severe.… See the full description on the dataset page: https://huggingface.co/datasets/maristombayeva/lost-in-speech.lai-speech
lai-ching-te-speech
ASR dataset packaged from SRT-aligned audio clips.
Dataset info
Split
Samples
train
10
validation
0
test
0
Total audio: 0.0 hours
Usage
from datasets import load_dataset, Audio
ds = load_dataset("lai-ching-te-speech")
ds = ds.cast_column("audio", Audio(sampling_rate=16000))
# Access a sample
sample = ds["train"][0]
print(sample["text"]) # transcript
print(sample["audio"]) # {"array": ..., "sampling_rate":… See the full description on the dataset page: https://huggingface.co/datasets/lazy-worm/lai-speech.welsh-speech-dataset
Welsh Speech Dataset
A multimodal dataset of 33 speakers producing 10 Welsh phrases, captured using 3DMD technology with audio and dense facial landmark annotations.
Dataset Overview
Speakers: 33 participants
Phrases: 10 Welsh phrases per speaker
Sequences: ~330 (33 speakers x 10 phrases)
Modalities:
Audio recordings (.wav)
3D facial reconstructions (.obj meshes + texture maps)
68-point facial landmarks (ibug68 template)
Fluency Scores: Each phrase rated 0-5 (5 =… See the full description on the dataset page: https://huggingface.co/datasets/arvinsingh/welsh-speech-dataset.VNMed_Speech
VNMed Speech
A Vietnamese medical speech dataset with human-reviewed transcripts, derived from doof-ferb/VietMed_unlabeled.
Dataset Description
VNMed Speech contains segmented audio clips from Vietnamese medical lectures and consultations, paired with transcripts that have been manually reviewed and corrected by human annotators.
Language: Vietnamese
Domain: Medical (cardiology, clinical consultations, medical lectures)
Sample rate: 8,000 Hz (mono)
Total samples:… See the full description on the dataset page: https://huggingface.co/datasets/conglq/VNMed_Speech.droidnexus-arabic-editorial-speech-scorecard-mini
DroidNexus Arabic Editorial Speech Scorecard Mini
A public DroidNexus Labs scorecard dataset for Arabic speech workflows: representative editorial scenarios, latency targets, overlap pressure, and the metric stack that decides whether a transcript is usable.
Why this exists
This dataset is the first public speech artifact layer for DroidNexus Labs. It publishes representative editorial workloads and evaluation pressure before claiming a full source-audio benchmark.… See the full description on the dataset page: https://huggingface.co/datasets/driodnexus/droidnexus-arabic-editorial-speech-scorecard-mini.welsh-speech-dataset
Welsh Speech Dataset
A multimodal dataset of 33 speakers producing 10 Welsh phrases, captured using 3DMD technology with audio and dense facial landmark annotations.
Dataset Overview
Speakers: 33 participants
Phrases: 10 Welsh phrases per speaker
Sequences: ~330 (33 speakers x 10 phrases)
Modalities:
Audio recordings (.wav)
3D facial reconstructions (.obj meshes + texture maps)
68-point facial landmarks (ibug68 template)
Fluency Scores: Each phrase rated 0-5… See the full description on the dataset page: https://huggingface.co/datasets/Pedramebd/welsh-speech-dataset.speechocean762_fbankmiscellaneous_yt_chunked_speech_restorised_nfa_aligned
miscellaneous_yt_chunked_speech_restorised_nfa_aligned
Public, manually gated NFA-aligned Uzbek speech dataset derived from instinct-org/miscellaneous_yt_chunked_speech_restorised.
Contents
Parquet shards: 130
Rows: 528,187
Approx hours: 863.88
Audio column: audio with embedded FLAC bytes
Transcript column: transcription
Alignment columns: nfa_token_alignments, nfa_word_alignments, nfa_segment_alignments, nfa_character_alignments
Access And Use… See the full description on the dataset page: https://huggingface.co/datasets/instinct-org/miscellaneous_yt_chunked_speech_restorised_nfa_aligned.espeech_podcasts_chunked_speech_restorised_nfa_aligned
Forced-Aligned STT Dataset
Source dataset: instinct-org/espeech_podcasts_chunked_speech_restorised
Aligned dataset: instinct-org/espeech_podcasts_chunked_speech_restorised_nfa_aligned
Rows: 2467471 successfully aligned rows
This dataset adds NeMo Forced Aligner metadata for STT training and timestamp quality
control. Source rows that did not produce usable CTM alignments are excluded from the
published data.
Added columns:
nfa_token_alignments: NeMo token/subword CTM spans… See the full description on the dataset page: https://huggingface.co/datasets/instinct-org/espeech_podcasts_chunked_speech_restorised_nfa_aligned.yt_chunked_speech_restorised_nfa_aligned
Forced-Aligned STT Dataset
Source dataset: instinct-org/yt_chunked_speech_restorised
Aligned dataset: instinct-org/yt_chunked_speech_restorised_nfa_aligned
Rows: 416380 successfully aligned rows
This dataset adds NeMo Forced Aligner metadata for STT training and timestamp quality
control. Source rows that did not produce usable CTM alignments are excluded from the
published data.
Added columns:
nfa_token_alignments: NeMo token/subword CTM spans
nfa_word_alignments: word-level… See the full description on the dataset page: https://huggingface.co/datasets/instinct-org/yt_chunked_speech_restorised_nfa_aligned.yt1_chunked_speech_restorised_nfa_aligned
Forced-Aligned STT Dataset
Source dataset: instinct-org/yt1_chunked_speech_restorised
Aligned dataset: instinct-org/yt1_chunked_speech_restorised_nfa_aligned
Rows: 261565 successfully aligned rows
This dataset adds NeMo Forced Aligner metadata for STT training and timestamp quality
control. Source rows that did not produce usable CTM alignments are excluded from the
published data.
Added columns:
nfa_token_alignments: NeMo token/subword CTM spans
nfa_word_alignments: word-level… See the full description on the dataset page: https://huggingface.co/datasets/instinct-org/yt1_chunked_speech_restorised_nfa_aligned.yt2_chunked_speech_restorised_nfa_aligned
Forced-Aligned STT Dataset
Source dataset: instinct-org/yt2_chunked_speech_restorised
Aligned dataset: instinct-org/yt2_chunked_speech_restorised_nfa_aligned
Rows: 809612 successfully aligned rows
This dataset adds NeMo Forced Aligner metadata for STT training and timestamp quality
control. Source rows that did not produce usable CTM alignments are excluded from the
published data.
Added columns:
nfa_token_alignments: NeMo token/subword CTM spans
nfa_word_alignments: word-level… See the full description on the dataset page: https://huggingface.co/datasets/instinct-org/yt2_chunked_speech_restorised_nfa_aligned.yt3_chunked_speech_restorised_nfa_aligned
Forced-Aligned STT Dataset
Source dataset: instinct-org/yt3_chunked_speech_restorised
Aligned dataset: instinct-org/yt3_chunked_speech_restorised_nfa_aligned
Rows: 506444 successfully aligned rows
This dataset adds NeMo Forced Aligner metadata for STT training and timestamp quality
control. Source rows that did not produce usable CTM alignments are excluded from the
published data.
Added columns:
nfa_token_alignments: NeMo token/subword CTM spans
nfa_word_alignments: word-level… See the full description on the dataset page: https://huggingface.co/datasets/instinct-org/yt3_chunked_speech_restorised_nfa_aligned.yt4_chunked_speech_restorised_nfa_aligned
Forced-Aligned STT Dataset
Source dataset: instinct-org/yt4_chunked_speech_restorised
Aligned dataset: instinct-org/yt4_chunked_speech_restorised_nfa_aligned
Rows: 965192 successfully aligned rows
This dataset adds NeMo Forced Aligner metadata for STT training and timestamp quality
control. Source rows that did not produce usable CTM alignments are excluded from the
published data.
Added columns:
nfa_token_alignments: NeMo token/subword CTM spans
nfa_word_alignments: word-level… See the full description on the dataset page: https://huggingface.co/datasets/instinct-org/yt4_chunked_speech_restorised_nfa_aligned.