MLS Eng Tokens: Pre-tokenized Audio Codec Tokens for TTS Training

somu9's mls_eng_tokens dataset provides pre-extracted audio codec tokens from the Multilingual LibriSpeech English corpus, tokenized using MOSS-Audio-Tokenizer. The dataset includes train, dev, and test splits and was last updated on 2026-05-17. Audio is processed at a 48,000 Hz sample rate and a 12.5 Hz frame rate.

Use Cases

• Train text-to-speech models based on pre-extracted audio codec tokens.
• Benchmark speech synthesis systems based on the standardized MLS English corpus.
• Experiment with residual vector quantization (RVQ) codebook structures based on the 16-codebook architecture.
• Fine-tune audio generation models using tokens with a defined 12.5 Hz frame rate.

Strengths

• Tokens are derived from the established Multilingual LibriSpeech (MLS) English dataset.
• Audio is processed with a specific codec configuration: 48,000 Hz sample rate, 12.5 Hz frame rate, and 16 RVQ codebooks.
• Dataset includes standard machine learning splits: train, dev, and test.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and license information are unknown, which may limit suitability assessment.

Provenance

Source

parler-tts/mls_eng (Multilingual LibriSpeech English dataset)

Collection Method

Pre-extracted and tokenized using MOSS-Audio-Tokenizer.

Time Range

null

Freshness

Last updated 2026-05-17 05:19:43; freshness should be verified.

Geography

null