brthor/libriheavy-tts-3

LibriHeavy TTS (WIP, not released yet)

An improved version of LibriHeavy designed for TTS training quality. It is built on top of mythicinfinity/libriheavy and focuses on better audio/text supervision quality.

Libriheavy: a 50,000 hours ASR corpus with punctuation casing and context. Libriheavy is a labeled version of Librilight.

Audio files are encoded using the Opus 68kbps codec to retain quality and reduce size.

Why This Dataset

• Utilizes higher-fidelity LibriVox source audio.
• Adds corrected training text (text_corrected) to reduce text/audio mismatch noise.
• Filters rows that have truncated speech at the end of the audio.
• Provides VAD-based trim metadata for cleaner supervision.
• Keeps full untrimmed audio so consumers can apply their own trimming policy.

Usage (Trim Transform)

import math
from datasets import load_dataset

ds = load_dataset('mythicinfinity/libriheavy-tts', 'dev')

def trim_audio_transform(batch):
    audios = batch['audio']
    starts = batch['audio_trim_start_s']
    ends = batch['audio_trim_end_s']
    out = []
    for audio, start_s, end_s in zip(audios, starts, ends):
        start = float(start_s)
        end = float(end_s)
        sr = int(audio['sampling_rate'])
        
        arr = audio['array']
        
        start_idx = max(0, int(math.floor(start * sr)))
        end_idx = max(0, int(math.ceil(end * sr)))

        out.append({'array': arr[start_idx:end_idx], 'sampling_rate': sr})
    
    batch['audio'] = out
    return batch
    
    
ds = ds.with_transform(trim_audio_transform)

Which Text Column Should I Use?

• Use text_corrected by default for TTS training targets.
• Use text_original when you want the original reference text from the base dataset.
• Use text_transcription when you prefer transcription-style text.

Column Descriptions

• audio: Full untrimmed audio waveform for each utterance.
• text_corrected: Corrected text intended as the primary training text.
• text_original: Original text from the base LibriHeavy dataset.
• text_transcription: Transcription text from the base LibriHeavy dataset.
• audio_trim_start_s: Suggested start trim boundary (seconds).
• audio_trim_end_s: Suggested end trim boundary (seconds).
• id: Utterance identifier.
• audio_duration: Duration in seconds from the base dataset, when present.
• speaker_id: Speaker identifier from the base dataset, when present.
• librivox_book_id: LibriVox book identifier from the base dataset, when present.

Improvement Details

1. Source Quality

• Motivation: TTS quality is sensitive to source fidelity and compression artifacts.
• Method: download higher quality LibriVox source files and extract audio segments. Also we retain the original higher sampling rate. This usually improves source audio from a 64kbps mp3 to a 128kbps mp3.
• Expected impact: cleaner acoustic detail for synthesis model learning.

2. Transcript Correction

• Motivation: text mismatches can introduce noisy supervision. We observe frequent and sometimes large mismatches.
• Method: we train a finetuned 8B LLM to match the content of text_transcription to the format of text_original. This retains the punctuation and specific word spellings the transcription model does not provide while ensuring the given text more accurately reflects what is spoken.
• Models: brthor/transcript-correction-loras.
• Expected impact: lower effective text/audio mismatch for training.

3. Truncation Detection

• Motivation: truncated endings can teach undesirable truncation bias. We observe some truncation.
• Method: truncation detection is applied and truncated samples are filtered out.
• Model: mythicinfinity/speech-truncation-detection-12M.
• Expected impact: reduced truncation-related artifacts in downstream models.

4. VAD-Based Trimming

• Motivation: excess leading/trailing silence is usually undesirable for TTS supervision.
• Method: VAD-derived trim boundaries are provided as metadata columns.
• Expected impact: easier dataset cleanup while preserving full original audio.

Configs

Each dataset config exposes a single split named train.

These are the config details of the original Libriheavy dataset.

Approximately 3% of rows have been filtered out by truncation detection.

• small (train): 509 hours of speech. 417 speakers averaging 1.22 hours per speaker.
• medium (train): 5042 hours of speech. 1531 speakers averaging 3.29 hours per speaker.
• large (train): 50794 hours of speech. 6736 speakers averaging 7.54 hours per speaker.
• dev (train): 22.3 hours of speech. 141 speakers averaging 0.16 hours per speaker.
• test_clean (train): 10.5 hours of speech. 70 speakers averaging 0.15 hours per speaker.
• test_other (train): 11.5 hours of speech. 72 speakers averaging 0.16 hours per speaker.
• test_clean_large (train): 107.5 hours of speech. 72 speakers averaging 1.49 hours per speaker.
• test_other_large (train): 100.3 hours of speech. 73 speakers averaging 1.37 hours per speaker.

Usage

Load a Single Config

from datasets import load_dataset

small = load_dataset("mythicinfinity/libriheavy-tts", "small", split="train")

Targeting a specific config only downloads files declared for that config, which is a good way to control disk usage.

Load the Full Dataset (All Configs)

from datasets import concatenate_datasets, load_dataset

ALL_CONFIGS = [
    "small",
    "medium",
    "large",
    "dev",
    "test_clean",
    "test_clean_large",
    "test_other",
    "test_other_large",
]


def load_libriheavy_tts_all_train(configs: list[str] | None = None):
    cfgs = configs or ALL_CONFIGS
    parts = [load_dataset("mythicinfinity/libriheavy-tts", cfg, split="train") for cfg in cfgs]
    return concatenate_datasets(parts)


full = load_libriheavy_tts_all_train()

Dataset Intended Use Cases

• Text-to-Speech (TTS)
• Automatic Speech Recognition (ASR)

Provenance, License, and Citation

• Base dataset: mythicinfinity/libriheavy
• Base project homepage: https://github.com/k2-fsa/libriheavy
• LibriHeavy paper: Libriheavy: a 50,000 hours ASR corpus with punctuation casing and context
• License: apache-2.0

Citation

@misc{kang2023libriheavy,
      title={Libriheavy: a 50,000 hours ASR corpus with punctuation casing and context},
      author={Wei Kang and Xiaoyu Yang and Zengwei Yao and Fangjun Kuang and Yifan Yang and Liyong Guo and Long Lin and Daniel Povey},
      year={2023},
      eprint={2309.08105},
      archivePrefix={arXiv},
      primaryClass={eess.AS}
}