TTS-AGI/Emotion-Voice-Attribute-Reference-Snippets-DACVAE-Wave

Emotion and Voice Attribute Reference Snippets - DACVAE and Wave

Merged dataset combining TTS-AGI/enhanced-emo-snippets-balanced-DACVAE and TTS-AGI/emotion-attribute-conditioning-dacvae with decoded WAV audio.

Overview

• Total samples: 606,178
• Filtered out: 363,331 (samples with speech_quality < 1.8)
• Total tar files: 328
• Total size: 1.54 TB
• Audio format: WAV, 48kHz, PCM 16-bit mono
• Latents: DAC-VAE float16 [T, 128] at 25 frames/sec
• Dimensions: 57 (40 emotions + 15 voice attributes + 2 additional attributes)

File Structure

Each tar file is named {Dimension}_{bucket_range}.tar and contains WebDataset-formatted samples:

{key}.json          # Full metadata (scores, text, captions, etc.)
{key}.target.npy    # DACVAE latent for target speech [T, 128] float16
{key}.target.wav    # Decoded target audio (48kHz WAV)
{key}.ref.npy       # DACVAE latent for speaker reference [T, 128] float16 (if available)
{key}.ref.wav       # Decoded reference audio (48kHz WAV) (if available)

Samples prefixed with emo_ come from DS1 (enhanced-emo-snippets-balanced), samples prefixed with cond_ come from DS2 (emotion-attribute-conditioning). DS2 samples include speaker reference audio (.ref.npy / .ref.wav), while DS1 samples include speaker embeddings in the JSON metadata.

Dimensions

Emotions (40)

Dimension	Buckets	Tar Files
Affection	[0,1) to [4,5)	5
Amusement	[0,1) to [4,5)	5
Anger	[0,1) to [5,6)	6
Astonishment_Surprise	[0,1) to [4,5)	5
Awe	[0,1) to [4,5)	5
Bitterness	[0,1) to [4,5)	5
Concentration	[0,1) to [4,5)	5
Confusion	[0,1) to [4,5)	5
Contemplation	[0,1) to [3,4)	4
Contempt	[0,1) to [4,5)	5
Contentment	[0,1) to [3,4)	4
Disappointment	[0,1) to [4,5)	5
Disgust	[0,1) to [3,4)	4
Distress	[0,1) to [4,5)	5
Doubt	[0,1) to [4,5)	5
Elation	[0,1) to [5,6)	6
Embarrassment	[0,1) to [2,3)	3
Emotional_Numbness	[0,1) to [3,4)	4
Fatigue_Exhaustion	[1,2) to [4,5)	4
Fear	[0,1) to [3,4)	4
Helplessness	[0,1) to [3,4)	4
Hope_Enthusiasm_Optimism	[0,1) to [6,7)	7
Impatience_and_Irritability	[0,1) to [4,5)	5
Infatuation	[0,1) to [4,5)	5
Interest	[0,1) to [3,4)	4
Intoxication_Altered_States_of_Consciousness	[0,1) to [4,5)	5
Jealousy_and_Envy	[0,1) to [4,5)	5
Longing	[0,1) to [3,4)	4
Malevolence_Malice	[0,1) to [3,4)	4
Pain	[0,1) to [5,6)	6
Pleasure_Ecstasy	[0,1) to [3,4)	4
Pride	[0,1) to [4,5)	5
Relief	[0,1) to [5,6)	6
Sadness	[0,1) to [4,5)	5
Sexual_Lust	[0,1) to [4,5)	5
Shame	[0,1) to [5,6)	6
Sourness	[0,1) to [3,4)	4
Teasing	[0,1) to [3,4)	4
Thankfulness_Gratitude	[0,1) to [4,5)	5
Triumph	[0,1) to [4,5)	5

Voice Attributes (15 from DS1 + 2 from DS2)

Attributes from DS1 use integer bucket ranges. Attributes from DS2 use float-valued bucket ranges derived from the conditioning pipeline.

Dimension	Bucket Type	Tar Files
Age	Integer [0,6) + Float [0.00, 5.14)	12
Arousal	Integer [0,6) + Float [0.00, 4.00)	13
Authenticity	Integer [1,5)	4
Background_Noise	Integer [0,3)	3
Confident_vs._Hesitant	Integer [0,5) + Float [0.00, 4.00)	12
Gender	Integer [0,3) + Float [0.29, 2.00)	6
High-Pitched_vs._Low-Pitched	Integer [0,5) + Float [0.00, 3.43)	11
Monotone_vs._Expressive	Integer [0,5) + Float [0.00, 4.00)	12
Recording_Quality	Integer [0,5)	5
Serious_vs._Humorous	Integer [0,6) + Float [0.00, 4.00)	13
Soft_vs._Harsh	Integer [0,2) + Float [0.29, 2.00)	5
Submissive_vs._Dominant	Integer [0,3) + Float [0.43, 3.00)	6
Valence	Integer [0,4) + Float [0.43, 3.00)	7
Vulnerable_vs._Emotionally_Detached	Integer [0,5)	5
Warm_vs._Cold	Integer [0,3) + Float [0.29, 2.00)	6
duration	Float [1.00, 30.00)	7
talking_speed	Float [5.00, 25.00)	7

Metadata Fields

Each sample's .json contains:

From DS1 (enhanced-emo-snippets-balanced):

• transcription — Speech transcript
• caption, detailed_caption, bude_whisper_caption — Natural language audio descriptions
• empathic_insight_scores — 59 float scores (40 emotions + 15 attributes + 4 quality)
• speaker_embedding — 128-dim speaker embedding vector
• emotion_vector — Encoded emotion vector
• enhancement_model — Speech enhancement model used (MossFormer2_SE_48K)
• duration — Audio duration in seconds

From DS2 (emotion-attribute-conditioning):

• text — Speech transcript
• caption — Natural language audio description
• annotation_scores — 59 float scores (same dimensions as DS1)
• target_duration, context_duration — Target and reference durations
• speaker, language — Speaker ID and language code

Added by merge pipeline:

• _source_dataset — "enhanced-emo-snippets-balanced" or "emotion-attribute-conditioning"
• _dimension — The emotion/attribute dimension name
• _bucket — The bucket label
• has_reference — Whether reference audio is available

Quality Scores

All samples include 59 annotation scores from Empathic Insight Voice Plus:

• 40 emotion scores: Amusement, Anger, Fear, Sadness, etc.
• 15 attribute scores: Valence, Arousal, Age, Gender, etc.
• 4 quality scores: score_overall_quality, score_speech_quality, score_content_enjoyment, score_background_quality

Only samples with score_speech_quality >= 1.8 are included in this dataset.

Sources

• DS1: TTS-AGI/enhanced-emo-snippets-balanced-DACVAE — Quality-ranked emotion/attribute snippets with speech enhancement
• DS2: TTS-AGI/emotion-attribute-conditioning-dacvae — Emotion/attribute conditioning pairs with speaker references
• DACVAE: mrfakename/dacvae-watermarked — DAC-VAE model for audio codec

Usage

import webdataset as wds
import numpy as np
import json, io, soundfile as sf

url = "https://huggingface.co/datasets/TTS-AGI/Emotion-Voice-Attribute-Reference-Snippets-DACVAE-Wave/resolve/main/data/Anger_4to5.tar"
ds = wds.WebDataset(url).decode()

for sample in ds:
    meta = json.loads(sample["json"])
    target_wav = sample["target.wav"]       # decoded 48kHz audio
    target_latent = np.load(io.BytesIO(sample["target.npy"]))  # [T, 128] float16

    if "ref.wav" in sample:
        ref_wav = sample["ref.wav"]         # speaker reference audio
        ref_latent = np.load(io.BytesIO(sample["ref.npy"]))    # [T, 128] float16

    # Access emotion scores
    scores = meta.get("empathic_insight_scores") or meta.get("annotation_scores", {})
    speech_quality = scores.get("score_speech_quality", 0)
    anger_score = scores.get("Anger", 0)

DACVAE Encode/Decode

Audio was decoded from DAC-VAE latents at 48kHz, 25 latent frames/sec:

import torch
from dacvae import DACVAE
from huggingface_hub import hf_hub_download

model = DACVAE.load(hf_hub_download("mrfakename/dacvae-watermarked", "weights.pth")).cuda().eval()

# Decode: latent -> audio
z = torch.from_numpy(latent.T).unsqueeze(0).float().cuda()  # [1, 128, T_latent]
audio_48k = model.decode(z).squeeze().cpu()                  # [T_audio] at 48kHz

# Encode: audio -> latent
audio = torch.from_numpy(wav).unsqueeze(0).unsqueeze(0).float().cuda()  # [1, 1, T_audio]
z_encoded = model.encode(audio)                                          # [1, 128, T_latent]
latent = z_encoded.squeeze(0).T.cpu().half().numpy()                     # [T_latent, 128] float16