This dataset contains synthetic speech. You agree not to use it to impersonate real people or to create deceptive content, and to use it for non-commercial research only.

ITTS Speaker Diversity

Dataset summary

Generated English speech for studying speaker diversity, demographic/affective conditioning, gender–age intersections, and sampling sensitivity. Audio is losslessly encoded as FLAC and embedded in Parquet, at each system's original sample rate. This release contains synthetic speech only, not real Emilia recordings; speaker embeddings of the real Emilia prompts are provided as a reference (see Speaker embeddings).

Configurations and release coverage

Config	Paper scope	Intended rows	Available released rows
overall	§4.1: neutral, three partitions	210,000	210,000
attribute	§5.2–5.5: 16 single attribute values	336,000	336,000
gender_x_age	§5.6: eight gender–age combinations	336,000	336,000
sampling_sweep	§4.2: five settings and three nominal seeds	450,000	450,000

Overall is complete: all seven systems (moss_voicegen, parler_large, parler_mini, capspeech_nar, qwen3_tts, voxinstruct, and voxcpm2) contain 30,000 rows each, with 10,000 rows in each of the three original partitions. The five previously missing splits were recovered from their original WAV outputs; no other experimental audio was substituted. Every recovered FLAC was decoded and checked for sample-exact equality with its source WAV. See metadata/overall_completion.json for the validation summary and metadata/overall_completion/*.jsonl for per-record source, text, and hash manifests.

All seven systems are included in attribute and gender_x_age; sampling_sweep contains only moss_voicegen, voxinstruct, and voxcpm2.

Single attributes: gender = female/male; age = child/young_adult/middle_aged/elderly; accent = north_american/british_isles/oceanian/south_asian/south_african/germanic; emotion = angry/happy/neutral/sad. Each single value has 3,000 rows per system. Each gender–age combination has 6,000 rows per system.

Fields

All configs share the same schema; inapplicable fields are null.

Field	Type	Description
id	string	Globally unique release ID; qualified sweep IDs include parameter value and nominal seed (owner-approved exception).
system	string	System split name.
instruction	string	Source style prompt; original neutral default for overall/sweep.
text	string	Text to be spoken.
emilia_id	string	Emilia utterance ID; no real source recording is distributed.
attribute	string	none, gender, age, accent, emotion, or gender_x_age.
value	string	Requested value, crossed combination, or none.
gender	string	Requested gender where applicable.
age	string	Requested age group where applicable.
template_id	string	Source s-prefix plus template/take; crossed rows also retain frame and word order.
partition	int32	Overall partition 1, 2, or 3; null in other configs.
sweep_param	string	temperature or cfg, sweep only.
sweep_value	float32	0.5, 1.0, 1.5, 2.0, or 2.5, sweep only.
seed	int32	Nominal experiment-group seed, sweep only; not a verified per-file recovery RNG seed.
decoding_params	string	Known parameter overrides as JSON; null when unrecorded. Full historical completeness is not asserted.
duration_s	float32	Positive decoded audio duration in seconds.
original_filename	string	Unchanged original WAV basename.
audio	Audio	Embedded lossless FLAC; no resampling, trimming, downmixing, or loudness normalization.

Loading

from datasets import load_dataset

dataset = load_dataset("dlion168/itts-speaker-diversity", "attribute", split="qwen3_tts")
print(dataset[0])

Each system is a separate split. Access requires accepting this repository's gate and authenticating with an authorized Hugging Face account.

Speaker embeddings

Precomputed speaker embeddings of every row of attribute, gender_x_age and sampling_sweep, of all seven splits of overall, and of the real Emilia prompts (below) are provided for 8 speaker-verification encoders, as one Parquet file per config, split and encoder:

embeddings/<config>/<split>/<encoder>.parquet    columns: id (string), embedding (float32[dim])

id is identical to the audio row's id, and rows are in the same order as the audio shards, so the two can be joined on id. Embeddings were extracted from the original WAVs (resampled to 16 kHz where needed) and are stored unnormalized; apply L2 normalization before computing cosine similarities. The three ModelScope encoders resample internally; all generated-speech embeddings here used torchaudio.transforms.Resample for that step. The overall embeddings of moss_voicegen, parler_large, parler_mini, qwen3_tts and voxinstruct were added on 2026-10-22, extracted from this release's audio; for these five splits the paper's original ModelScope embeddings had partly used sox resampling instead, so ModelScope figures computed from them differ from the paper's first version (up to 12% in Vendi score). metadata/embeddings.json lists the encoders, dimensions and per-file row counts.

Encoder folder	Model	Checkpoint	Dim
`speechbrain_ecapa`	ECAPA-TDNN (VoxCeleb)	HF `speechbrain/spkrec-ecapa-voxceleb`	192
`speechbrain_resnet`	ResNet (VoxCeleb)	HF `speechbrain/spkrec-resnet-voxceleb`	256
`wespeaker`	ResNet221-LM (VoxCeleb)	WeSpeaker `voxceleb_resnet221_LM`	256
`wespeaker_voxblink2_resnet152_16k`	SimAM-ResNet34 (VoxBlink2)	WeSpeaker `voxblink2_samresnet34`	256
`nemo_titanet`	TitaNet-Large	HF `nvidia/speakerverification_en_titanet_large`	192
`modelscope_eres2net`	ERes2Net	ModelScope `iic/speech_eres2net_sv_zh-cn_16k-common`	192
`modelscope_eres2netv2`	ERes2NetV2	ModelScope `iic/speech_eres2netv2_sv_zh-cn_16k-common`	192
`modelscope_campplus`	CAM++	ModelScope `iic/speech_campplus_sv_zh-cn_16k-common`	192

Real Emilia reference embeddings

embeddings/real_emilia/partition_<k>/<encoder>.parquet holds the same 8 embeddings for the 10,000 real Emilia prompt recordings of partition k (30,000 rows in total), as a human-speech reference. Columns are id (Emilia utterance ID, matching emilia_id in the audio configs and metadata/emilia_ids_overall_partition_<k>.txt), speaker (Emilia's own speaker ID, from its diarization) and embedding. Partition k supplies the text of overall partition k; sampling_sweep uses partition 1. Only embeddings are released, not the recordings; Emilia speaker IDs are per source recording, so one person may appear under several IDs, and several clips share a speaker ID (about 3 clips per ID on average).

The folder name wespeaker_voxblink2_resnet152_16k is historical: it holds embeddings from the VoxBlink2 SimAM-ResNet34 model, not a ResNet152. Encoder weights are subject to their own licenses.

import numpy as np
from huggingface_hub import hf_hub_download
import pyarrow.parquet as pq

path = hf_hub_download("dlion168/itts-speaker-diversity",
                       "embeddings/attribute/qwen3_tts/nemo_titanet.parquet", repo_type="dataset")
table = pq.read_table(path)
ids = table["id"].to_pylist()
X = np.stack(table["embedding"].to_numpy(zero_copy_only=False))

Provenance and known limitations

The original neutral instruction is: Speak like a person — natural, neutral tone, moderate pace, conversational and clear delivery. It is retained, not replaced by the specification example Speak in a natural voice..

An exact source-prompt ID join supplies attribute text/instructions. Emilia IDs use the owner-approved s-prefix/one-based-row mapping into the corresponding Emilia input partition, with exact text agreement checked for every selected row.

Sweep IDs use sampling_sweep/<split>/<temperature_or_cfg>_<value>/seed_<seed>/<original_stem> because the originally requested shorter ID collides across sweep groups. Other configs retain <config>/<split>/<original_stem>.

Historical model revisions and complete nondefault decoding parameters were not consistently recorded. Unknown revisions are explicitly null in metadata/systems.json; missing complete parameters are not fabricated. Sweep decoding JSON contains only the known swept parameter, not an assertion that all other overrides are known. Recovery jobs may have used a different RNG seed from the nominal sweep-group seed. Per-record provenance gaps, including original filenames and source-relative paths, are listed in metadata/provenance_gaps.jsonl. The release owner requested following the original discrepancy-reporting workflow rather than silently correcting or inventing metadata.

Two pairs of rows have byte-identical source audio, and therefore identical embeddings for most encoders: the real Emilia clips EN_B00032_S05646_W000000 and EN_B00032_S05666_W000000 (one segment assigned to two Emilia speaker IDs), and sampling_sweep/voxinstruct/cfg_2.0/seed_42/EN_B00053_S03356_W000000 and sampling_sweep/voxinstruct/cfg_2.5/seed_42/EN_B00053_S03356_W000000 (generated in separate runs). Both are kept as is.

A few overall clips are empty or shorter than 0.1 s, so some encoders return no embedding or a NaN one: overall/voxinstruct/EN_B00079_S01244_W000008 (empty audio) has no row for six encoders and a NaN row for nemo_titanet; overall/parler_large/EN_B00059_S04550_W000002, EN_B00032_S08946_W000001 and EN_B00059_S05490_W000000 have NaN rows for some encoders. Drop non-finite rows before computing metrics.

metadata/checksums.sha256 records original WAV SHA-256 hashes. Earlier entries use paths relative to the original release data root; recovered overall entries use paths relative to the completion source root recorded in metadata/provenance.json and metadata/overall_completion.json. These hashes are not hashes of the converted FLAC bytes. metadata/shards.json contains Parquet SHA-256 hashes. The stored template_id retains the source s-prefix and template/take suffix; s-prefix also identifies the source text partition and must not be interpreted as a newly invented template.

Excluded material

Real Emilia recordings are excluded for licensing and speaker privacy; only their speaker embeddings are included. The release excludes non_binary, l2_accent, native_accent, MOSS-TTS-v1.5, Parler-v1.1, PromptTTS2, InstructTTS, unseeded outputs/param_sweep, embeddings of excluded material, model weights, generation logs, environment files, and credentials.

License

CC BY-NC 4.0. Emilia source text is subject to CC BY-NC 4.0 and CapSpeech-NAR is restricted to noncommercial use; the combined dataset is therefore released for noncommercial research under CC BY-NC 4.0. This dataset license does not override upstream model terms or confer rights to impersonate real speakers.

Intended use

Noncommercial research on synthetic speaker diversity, conditioning behavior, representation analysis, and reproducibility. Do not use this dataset to impersonate real people, create deceptive content, or support commercial uses prohibited by its license. Attribute prompts describe requested generation conditions, not verified demographic identities of real people.

Citation

Downloads last month: 352

Total file size:

363 GB