datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
emolia-thinking
Emolia-Thinking — a VoiceNet-annotated, balanced subset of Emolia
Emolia-Thinking is a richly annotated speech dataset created for the VoiceNet project. It takes a balanced subset of the Emolia corpus — balanced across speaker-embedding clusters and emotion-embedding clusters so that speakers, voices and emotional states are evenly represented rather than dominated by the most common cases — and annotates every clip along the full VoiceNet Extended voice-performance taxonomy… See the full description on the dataset page: https://huggingface.co/datasets/VoiceNet/emolia-thinking.Emolia
Dataset Card for Emolia
Dataset Description
This dataset is an enhanced version of the Emilia dataset, enriched with detailed emotion annotations. The annotations were generated using models from the EmoNet suite to provide deeper insight into the emotional content of speech. This work is based on the research and models described in the blog post "Do They See What We See?".
The annotations include 54 scores for each sample, covering a wide range of emotional and… See the full description on the dataset page: https://huggingface.co/datasets/laion/Emolia.emolia-thinking-balanced-buckets
Emolia-Thinking — Balanced Per-Dimension Bucket Subset
A balanced, per-dimension bucket subset of
VoiceNet/emolia-thinking,
derived from that dataset's zero-shot VoiceNet-dimension labels.
For every VoiceNet voice/prosody/timbre/style dimension, this subset draws a
roughly equal number of clips from each ordinal bucket (0–6), so that
downstream training / probing sees a balanced distribution along each axis
instead of the strongly skewed natural distribution.
How… See the full description on the dataset page: https://huggingface.co/datasets/laion/emolia-thinking-balanced-buckets.emolia
emolia-balanced-5M-subset · flac 48 kHz · WebDataset (paired)
This is the emolia-balanced-5M-subset corpus repackaged for high-quality
audio–text contrastive training. Audio is re-encoded as mono FLAC at 48 kHz
(PCM 16-bit) and stored as a WebDataset of paired <key>.flac + <key>.json
samples.
The JSON sidecar carries the full annotation stack:
Original metadata (id, text, duration, speaker, language, dnsmos).
A free-text emotion_caption derived from the emotion-annotation scalars.… See the full description on the dataset page: https://huggingface.co/datasets/VoiceNet/emolia.emolia_filtered_nano_codec_21_dataset
Emolia · Filtered · NanoCodec (FSQ) Tokens
A cleaned, pre-tokenized version of laion/Emolia
prepared for text-to-speech (TTS) training.
The pipeline is two steps:
Quality filtering with the open-source
audio_filter tool — this
removes the dirtiest recordings (noise, clipping, band-limiting, robotic artifacts,
overlapping speakers), which matters a lot for TTS quality.
Discrete audio tokenization with NVIDIA
nvidia/nemo-nano-codec-22khz-1.89kbps-21.5fps
(an FSQ neural audio… See the full description on the dataset page: https://huggingface.co/datasets/nineninesix/emolia_filtered_nano_codec_21_dataset.emolia-hq
Emolia-HQ
Emolia-HQ is a high-quality, speaker-paired subset of the LAION Emolia dataset. Each sample includes a target utterance and a reference utterance from the same speaker, enabling speaker-conditioned tasks such as voice conversion, expressive TTS, and speaker-aware emotion recognition.
Source
Derived from laion/Emolia by:
Quality filtering: Only samples with dnsmos >= 3.0 are retained.
Speaker pairing: Each target sample is matched with a reference audio from the… See the full description on the dataset page: https://huggingface.co/datasets/TTS-AGI/emolia-hq.emolia-3k-speaker-clusters
Emolia 3K Speaker Clusters
A curated set of 3,000 diverse speaker clusters derived from the TTS-AGI/emolia-hq dataset, with up to 20 representative audio samples per cluster.
Overview
The original emolia-hq dataset contains hundreds of thousands of speech samples with 128-dimensional WavLM speaker timbre embeddings. These were first clustered into 10,000 centroids, then intelligently pruned to 3,000 using density-aware farthest-point sampling to ensure:
Outlier… See the full description on the dataset page: https://huggingface.co/datasets/laion/emolia-3k-speaker-clusters.emolia-3k-speaker-clusters-DACVAE
Emolia 3K Speaker Clusters
A curated set of 3,000 diverse speaker clusters derived from the TTS-AGI/emolia-hq dataset, with up to 20 representative audio samples per cluster.
Overview
The original emolia-hq dataset contains hundreds of thousands of speech samples with 128-dimensional WavLM speaker timbre embeddings. These were first clustered into 10,000 centroids, then intelligently pruned to 3,000 using density-aware farthest-point sampling to ensure:
Outlier… See the full description on the dataset page: https://huggingface.co/datasets/TTS-AGI/emolia-3k-speaker-clusters-DACVAE.emolia-voicenet-gemini-annotations
Emolia VoiceNet Gemini Annotations
468,180 dimension-level annotations over 236,613 Emolia speech clips,
each scored 0-6 (0-2 for the content-safety dimension) on one of 57 perceptual
voice / speech dimensions - arousal, valence, brightness, resonance placement, speaking
styles, genuineness, recording quality, and more - by Gemini 3.5 Flash (non-thinking,
temperature 0). This repository ships the annotations, audio provenance, per-dimension
statistics, and the full scoring… See the full description on the dataset page: https://huggingface.co/datasets/laion/emolia-voicenet-gemini-annotations.emolia_filtered_v1
Emolia Filtered v1 (103,521 samples)
Subset of laion/Emolia processed through
the audio_filter pipeline.
All samples preserved (good + bad + uncertain), with filter results as additional columns.
Pipeline
Stage
Model
Purpose
1. Quality
Dual LogisticRegression (V1 SR<=24kHz / V3 SR>24kHz) on DSP metrics
Detect noise, clipping, robotic, bandwidth-limited audio
2. Speaker
Pyannote ONNX segmentation-3.0
Detect overlapping speakers
Speaker filter runs only on… See the full description on the dataset page: https://huggingface.co/datasets/KaniTTS-research-team/emolia_filtered_v1.emolia-hq
Emolia-HQ
Emolia-HQ is a high-quality, speaker-paired subset of the LAION Emolia dataset. Each sample includes a target utterance and a reference utterance from the same speaker, enabling speaker-conditioned tasks such as voice conversion, expressive TTS, and speaker-aware emotion recognition.
Source
Derived from laion/Emolia by:
Quality filtering: Only samples with dnsmos >= 3.0 are retained.
Speaker pairing: Each target sample is matched with a reference audio… See the full description on the dataset page: https://huggingface.co/datasets/laion/emolia-hq.moss-emolia-elise-hq-captioned
MOSS · Emolia + Elise + Inline-Bursts — HQ, captioned
A high-quality, richly captioned slice of the MOSS-local voice-acting corpus: expressive speech
clips scored by a panel of acoustic detectors, filtered to the top by a composite reward, and captioned
in the voice-acting format (a "how the voice sounds / how to perform it" description plus the script
with inline vocal-burst tags). Audio is shipped both as flac (WebDataset tars) and as pre-computed
MOSS-Audio-Tokenizer codes… See the full description on the dataset page: https://huggingface.co/datasets/TTS-AGI/moss-emolia-elise-hq-captioned.emolia-balanced-5M-subset
emolia-balanced-5M-subset
A balanced ~5.26M-sample subset of laion/Emolia (80.5M speech samples), packaged as WebDataset-compatible tar shards for direct use in training pipelines.
How this subset was filtered
Samples were selected if they met either of two criteria:
1. Emotion thresholds
Each sample carries 40 emotion annotation scores (from the Emonet taxonomy) in its metadata. A sample qualifies for an emotion bucket if its score for that emotion meets or… See the full description on the dataset page: https://huggingface.co/datasets/laion/emolia-balanced-5M-subset.emolia_top_1000_subsets
emolia_top_1000_subsets
This dataset contains tar archives with top-1000 subsets derived from the LAION EMOLIA audio emotion dataset: for each EMOLIA annotation category we selected the 1000 clips with the highest scores and re-annotated them with Gemini 2.5 Flash, producing captions that focus strongly on foreground versus background sounds, environmental and background noise, speaker emotion, and stable speaker attributes, so these subsets provide compact, high-quality material… See the full description on the dataset page: https://huggingface.co/datasets/laion/emolia_top_1000_subsets.emolia-hq-tokenizedemolia-extreme-buckets-hq-captionedemolia-emotion-subsetsemolia-emonet-voice-100-subsetsemolia_filtered_v1_bb_featuresemolia-hq-normalizedemolia_filter_sr_32
emolia_filter_sr_32
60 аудиосэмплов для прослушивания и калибровки порогов фильтра качества.
Отсортированы от лучшего к худшему по bad_prob. SR: 24kHz.
Колонки
Колонка
Тип
Описание
audio
Audio
Аудиофайл (24kHz)
category
string
best_good / boundary_good / boundary_bad / worst_bad
bad_prob
float64
Вероятность плохого качества (0.0–1.0)
sample_id
string
ID сэмпла
speaker_status
string
GOOD / UNCERTAIN / BAD (speaker overlap)
speaker_overlap
float64… See the full description on the dataset page: https://huggingface.co/datasets/KaniTTS-research-team/emolia_filter_sr_32.emolia_filtered_features