consulted-graphs/rhea-tts-vocoder

Unified Vocoder Training Dataset

This dataset combines LJSpeech, VCTK, and LibriTTS for training neural vocoders. Each dataset is available as a separate configuration, with a default 'all' configuration that combines all datasets. You can train on individual datasets or use the combined dataset as needed.

Dataset Statistics

• Total files: 43,691
• Total duration: 71.2 hours
• Total speakers: 326

Per-dataset breakdown:

• libritts:
  • Files: 43,691
  • Duration: 71.2 hours
  • Speakers: 326

Audio Processing Parameters

All audio files have been processed with the following parameters:

• Sample rate: 22,050 Hz
• Mel-spectrogram bins: 80
• FFT size: 1024
• Hop length: 256
• Window length: 1024
• Frequency range: 0.0-8000.0 Hz
• Normalization: [-4, 4]

Usage

The dataset is organized by source dataset (ljspeech, vctk, libritts) with each stored as a separate configuration. The default 'all' configuration automatically combines all datasets.

Loading Combined Dataset (Default)

from datasets import load_dataset

# Load all datasets combined (default configuration)
all_data = load_dataset("consulted-graphs/rhea-tts-vocoder", split="train")
all_data_val = load_dataset("consulted-graphs/rhea-tts-vocoder", split="validation")

# Or explicitly specify the 'all' configuration
all_data = load_dataset("consulted-graphs/rhea-tts-vocoder", "all", split="train")

# Access a sample
sample = all_data[0]
audio = sample['audio']['array']  # Audio waveform
mel_spectrogram = sample['mel_spectrogram']  # Mel-spectrogram
speaker_id = sample['speaker_id']  # Global speaker ID
dataset_name = sample['dataset']  # Source dataset (ljspeech, vctk, or libritts)

Loading Individual Datasets

from datasets import load_dataset

# Load individual datasets - HuggingFace will automatically load all parquet files
ljspeech = load_dataset("consulted-graphs/rhea-tts-vocoder", "ljspeech", split="train")
vctk = load_dataset("consulted-graphs/rhea-tts-vocoder", "vctk", split="train")
libritts = load_dataset("consulted-graphs/rhea-tts-vocoder", "libritts", split="train")

# Each dataset has its own train/validation splits where applicable
libritts_val = load_dataset("consulted-graphs/rhea-tts-vocoder", "libritts", split="validation")

# Access a sample
sample = ljspeech[0]
audio = sample['audio']['array']  # Audio waveform
mel_spectrogram = sample['mel_spectrogram']  # Mel-spectrogram
speaker_id = sample['speaker_id']  # Global speaker ID

Manual Dataset Combination

from datasets import load_dataset, concatenate_datasets

# Method 1: Load and concatenate manually
all_train_datasets = []
for ds_name in ['ljspeech', 'vctk', 'libritts']:
    try:
        ds = load_dataset("consulted-graphs/rhea-tts-vocoder", ds_name, split="train")
        all_train_datasets.append(ds)
    except:
        print(f"Skipping {ds_name} (not found)")

# Combine all datasets
combined_train = concatenate_datasets(all_train_datasets)
print(f"Total training samples: {len(combined_train)}")

# Method 2: Load with specific speaker filtering
def load_with_speaker_filter(max_speakers_per_dataset=10):
    filtered_datasets = []
    
    for ds_name in ['ljspeech', 'vctk', 'libritts']:
        ds = load_dataset("consulted-graphs/rhea-tts-vocoder", ds_name, split="train")
        
        # Get unique speakers
        unique_speakers = ds.unique('speaker_id')[:max_speakers_per_dataset]
        
        # Filter dataset
        filtered_ds = ds.filter(lambda x: x['speaker_id'] in unique_speakers)
        filtered_datasets.append(filtered_ds)
    
    return concatenate_datasets(filtered_datasets)

# Load with limited speakers for faster experimentation
small_train = load_with_speaker_filter(max_speakers_per_dataset=5)

Streaming Mode

For large-scale training without downloading the entire dataset:

from datasets import load_dataset, interleave_datasets

# Stream all datasets combined (default configuration)
all_stream = load_dataset("consulted-graphs/rhea-tts-vocoder", split="train", streaming=True)

# Or stream individual datasets
ljspeech_stream = load_dataset("consulted-graphs/rhea-tts-vocoder", "ljspeech", split="train", streaming=True)
vctk_stream = load_dataset("consulted-graphs/rhea-tts-vocoder", "vctk", split="train", streaming=True)
libritts_stream = load_dataset("consulted-graphs/rhea-tts-vocoder", "libritts", split="train", streaming=True)

# Method 1: Interleave streams with custom probabilities
combined_stream = interleave_datasets(
    [ljspeech_stream, vctk_stream, libritts_stream],
    probabilities=[0.2, 0.4, 0.4]  # More weight on multi-speaker datasets
)

# Method 2: Stream specific datasets only
multi_speaker_stream = interleave_datasets(
    [vctk_stream, libritts_stream],
    probabilities=[0.3, 0.7]  # Focus on LibriTTS
)

# Use in training loop
for batch in combined_stream.batch(batch_size=16):
    # batch['audio'] - audio arrays
    # batch['mel_spectrogram'] - mel spectrograms
    # batch['speaker_id'] - speaker IDs
    # Process batch...

Training Examples

Single Dataset Training

import torch
from torch.utils.data import DataLoader
from datasets import load_dataset

# Load a single dataset for focused training
ljspeech = load_dataset("consulted-graphs/rhea-tts-vocoder", "ljspeech", split="train")

# Create PyTorch dataset wrapper
class VocoderDataset(torch.utils.data.Dataset):
    def __init__(self, hf_dataset):
        self.dataset = hf_dataset
    
    def __len__(self):
        return len(self.dataset)
    
    def __getitem__(self, idx):
        sample = self.dataset[idx]
        return {
            'audio': torch.tensor(sample['audio']['array']),
            'mel': torch.tensor(sample['mel_spectrogram']).T,
            'speaker_id': sample['speaker_id']
        }

# Single speaker training (LJSpeech)
train_dataset = VocoderDataset(ljspeech)
train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True)

Multi-Dataset Training

from datasets import load_dataset, concatenate_datasets

# Load all datasets
datasets = []
for config in ['ljspeech', 'vctk', 'libritts']:
    try:
        ds = load_dataset("consulted-graphs/rhea-tts-vocoder", config, split="train")
        datasets.append(ds)
        print(f"Loaded {config}: {len(ds)} samples")
    except:
        print(f"Skipping {config}")

# Combine for multi-speaker training
combined_dataset = concatenate_datasets(datasets)
print(f"Total samples: {len(combined_dataset)}")

# Create balanced sampler for equal speaker representation
from torch.utils.data import WeightedRandomSampler

def create_balanced_sampler(dataset):
    speaker_counts = {}
    for i in range(len(dataset)):
        spk_id = dataset[i]['speaker_id']
        speaker_counts[spk_id] = speaker_counts.get(spk_id, 0) + 1
    
    # Create weights inverse to speaker frequency
    weights = []
    for i in range(len(dataset)):
        spk_id = dataset[i]['speaker_id']
        weights.append(1.0 / speaker_counts[spk_id])
    
    return WeightedRandomSampler(weights, len(weights))

# Multi-speaker training with balanced sampling
train_dataset = VocoderDataset(combined_dataset)
sampler = create_balanced_sampler(combined_dataset)
train_loader = DataLoader(train_dataset, batch_size=16, sampler=sampler)

Citation

Please cite the original datasets:

• LJSpeech: https://keithito.com/LJ-Speech-Dataset/
• VCTK: https://datashare.ed.ac.uk/handle/10283/3443
• LibriTTS: https://openslr.org/60/

License

Please refer to the individual dataset licenses.