somu9/mls_eng_tokens

MLS English - Pre-tokenized Audio Codec Tokens

Pre-extracted audio codec tokens from the Multilingual LibriSpeech (MLS) English dataset, tokenized using MOSS-Audio-Tokenizer for text-to-speech training.

Source

Property	Value
Source dataset	parler-tts/mls_eng
Audio codec	MOSS-Audio-Tokenizer
Language	English
Codec sample rate	48,000 Hz (stereo)
Frame rate	12.5 Hz (1 frame = 80ms)
Codebooks	16 (RVQ, 1024 vocab each)
Splits included	train, dev, test

Dataset Details

The MLS English dataset contains ~44,000 hours of read English audiobook speech derived from LibriVox recordings. This tokenized version provides pre-extracted codec tokens ready for autoregressive TTS model training, eliminating the need to run the audio codec encoder during training.

Each audio sample has been:

1. Decoded from the source parquet files
2. Resampled to 48kHz stereo (codec input format)
3. Encoded using MOSS-Audio-Tokenizer into 16-codebook RVQ tokens at 12.5 Hz
4. Stored as JSONL chunks with transcripts and speaker IDs

Format

The dataset is stored as JSONL chunk files in the chunks/ directory. Each chunk corresponds to one source parquet file. Each line in a chunk file is a JSON object:

{
  "text": "The transcribed text",
  "audio_codes": [[cb0, cb1, ..., cb15], ...],
  "n_frames": 125,
  "n_codebooks": 16,
  "speaker_id": "mls_eng_12345"
}

Fields

Field	Type	Description
text	string	Transcript of the utterance
audio_codes	list[list[int]]	[n_frames, 16] — RVQ token IDs (0-1023) per frame, ordered coarse to fine
n_frames	int	Number of audio frames (duration = n_frames / 12.5 seconds)
n_codebooks	int	Always 16
speaker_id	string	Speaker identifier, prefixed with mls_eng_

Usage

Reading the tokens

import json
from pathlib import Path

chunks_dir = Path("chunks")
for chunk_file in sorted(chunks_dir.glob("chunk_*.jsonl")):
    with open(chunk_file) as f:
        for line in f:
            sample = json.loads(line)
            text = sample["text"]
            codes = sample["audio_codes"]  # [n_frames, 16]
            speaker = sample["speaker_id"]
            duration = len(codes) / 12.5
            print(f"{speaker} | {duration:.1f}s | {text[:80]}")

Decoding tokens back to audio

from transformers import AutoModel
import torch

codec = AutoModel.from_pretrained(
    "OpenMOSS-Team/MOSS-Audio-Tokenizer",
    trust_remote_code=True,
).eval()

codes = torch.tensor(sample["audio_codes"])  # [T, 16]
codes = codes.T.contiguous()                 # [16, T]
decoded = codec.batch_decode([codes], num_quantizers=16, chunk_duration=None)
waveform = decoded.audio[0]                  # [2, samples] at 48kHz stereo

Loading all chunks into a single dataset

import json
from pathlib import Path

samples = []
for chunk_file in sorted(Path("chunks").glob("chunk_*.jsonl")):
    with open(chunk_file) as f:
        for line in f:
            samples.append(json.loads(line))
print(f"Loaded {len(samples)} samples")

Tokenization Details

• Duration filter: Samples shorter than 0.5s or longer than 30s are excluded
• Audio preparation: All audio is resampled to 48kHz and converted to stereo before encoding
• Quantization: Residual Vector Quantization (RVQ) with 16 codebooks, each with a vocabulary of 1024
• Speaker IDs: Preserved from the source dataset with mls_eng_ prefix for global uniqueness across multi-dataset training

Credits & Acknowledgements

Source Dataset

This dataset is derived from Multilingual LibriSpeech (MLS):

Pratap, V., Xu, Q., Sriram, A., Synnaeve, G., & Collobert, R. (2020). MLS: A Large-Scale Multilingual Dataset for Speech Research. Proc. Interspeech 2020, 2757-2761.

The MLS dataset is based on read audiobooks from LibriVox. The English subset hosted by Parler-TTS was used as the source.

Audio Codec

Audio tokenization was performed using MOSS-Audio-Tokenizer by the OpenMOSS Team at Shanghai Jiao Tong University:

Jiang, Z., Liu, J., Chen, J., et al. (2025). MOSS-TTS: Modular Open Speech Synthesis System for Text-to-Speech. arXiv:2603.18090.

Tokenization Pipeline

Tokenization was performed using the MintTTS pipeline.

License

This dataset inherits the CC BY 4.0 license from the source MLS dataset. The MOSS-Audio-Tokenizer is used under its respective license. Please refer to the original dataset and model cards for full licensing terms.