lonesamurai/emilia_clean_10k

EMILIA Clean 10k

A filtered subset of the amphion/Emilia-Dataset (English split), designed for single-speaker TTS training.

Dataset Statistics

• Total clips: 10,000
• Speakers: 200 (single-speaker English)
• Train / Val split: 8,000 / 2,000
• Duration per clip: 3–10 seconds
• Sample rate: 24 kHz (mono)
• Language: English (EN)

Filtering Pipeline

1. Candidate selection — Filtered EMILIA EN clips for duration (3–10s) and DNSMOS quality (≥3.2). Selected top 400 speakers with ≥60 clips each → 24,000 candidates.
2. Overlap detection — SpeechBrain ECAPA speaker embeddings with agglomerative clustering (threshold=0.30) on 1.5s windows. Kept clips with exactly 1 detected speaker.
3. Final selection — Top 200 speakers by clean-clip count, sampled 50 clips each. 80/20 train/val split per speaker.
4. Preprocessing — Extracted text tokens, prosody embeddings, F0/energy, RVQ codes (Mimi @ 16kHz), and speaker embeddings using Qwen3-TTS 1.7B base model.

Files

File	Size	Description
manifest.json	~5 MB	Train/val split definitions with clip metadata (id, speaker, text, duration, wav_path, split)
speakers.csv	~8 KB	Speaker metadata (ID, split, clip counts)
clean_audio.tar.gz	~2.7 GB	10,000 WAV files (24 kHz, mono) — raw audio for the final selected clips
preprocessed.tar.gz	~1.1 GB	10,000 PyTorch .pt files + stats.json — preprocessed training tensors

Archive Contents

clean_audio.tar.gz

clean_audio/
├── EN_B00000_S00012_W000020.wav
├── EN_B00000_S00012_W000023.wav
├── ... (10,000 files total)

Each file is named {BATCH}_{SPEAKER}_{CLIP}.wav.

preprocessed.tar.gz

preprocessed/
├── EN_B00000_S00012_W000020.pt
├── EN_B00000_S00012_W000023.pt
├── ... (10,000 .pt files)
└── stats.json

Each .pt file contains the following tensors:

Key	Shape	Description
input_ids	[1, T_text]	Tokenized text (Qwen3-TTS tokenizer)
prosody_embeds	[1, T_prosody, 512]	Prosody embeddings
f0_target	[T_prosody]	Log F0 target (pooled by factor 16)
energy_target	[T_prosody]	Log energy target
voiced_target	[T_prosody]	Voiced/unvoiced flag (0 or 1)
rvq_tokens	[T_rvq, 16]	Mimi RVQ codes at 12.5 Hz
spk_emb	[1, spk_dim]	Speaker embedding from ECAPA
voice_clone_spk_emb	[1, spk_dim]	Voice clone reference embedding
duration	scalar	Clip duration in seconds

stats.json contains global F0/energy normalization statistics computed from the train split:

{
  "f0_mean": 4.8196,
  "f0_std": 0.2878,
  "energy_mean": -3.3469,
  "energy_std": 1.5586,
  "n_train_samples": 8000
}

Usage

import json, torch

# Load manifest
with open("manifest.json") as f:
    manifest = json.load(f)

train_clips = manifest["train"]
val_clips = manifest["val"]

# Load a preprocessed sample
sample = torch.load("preprocessed/EN_B00000_S00012_W000020.pt")
print(sample.keys())

License

Derived from the EMILIA dataset. Please refer to the original dataset's license terms.