---
license: cc-by-4.0
language: [en]
size_categories: [100K<n<1M]
tags: [audio, speech, libritts, codec, tokens]
---

# libritts-mimi-tokens

> To learn about Trelis Enterprise Voice Services, see [Trelis.com/voice-ai-services](https://Trelis.com/voice-ai-services).

LibriTTS-R encoded with [kyutai/mimi](https://huggingface.co/kyutai/mimi) (RVQ codec, 8 codebooks x 2,048 entries, 12.5 frames/sec). Two streams per row:

- `codes_semantic` (list[uint32], 12.5 fps, vocab 2,048) — codebook 0 only (WavLM-distilled, content-aligned).
- `codes_all_flat` (list[uint32], 100 fps, **offset vocab 16,384**) — all 8 codebooks interleaved per frame, with codebook k mapped to `[k*2048, (k+1)*2048)` so a flat LM can disambiguate.

## Splits

Mirrors the source LibriTTS-R splits (filtered by parler-tts; total ≈ 538 h):

| split | utterances | hours |
|---|---:|---:|
| `train.clean.100` | ~32 k | ~53 h |
| `train.clean.360` | ~112 k | ~218 h |
| `train.other.500` | ~250 k | ~258 h |
| `dev.clean` | ~5.6 k | ~9 h |

Source: [`parler-tts/libritts_r_filtered`](https://huggingface.co/datasets/parler-tts/libritts_r_filtered).
Disjointness between splits is structural (HF split definition — no speaker
overlap between train.* and dev.*).

## Schema (one row per utterance)

| column | type | description |
|---|---|---|
| `id` | string | Source utterance id (LibriTTS `speaker_chapter_segment`) |
| `speaker` | string | LibriTTS speaker id |
| `duration` | float32 | Audio duration in seconds |
| `text` | string | `text_normalized` from source |
| `codes_semantic` | list[uint32] | Mimi codebook 0 only, 12.5 fps |
| `codes_all_flat` | list[uint32] | All 8 codebooks interleaved with offset vocab, 100 fps |

## Loading

```python
from datasets import load_dataset

# Load the whole dataset (all 4 splits)
ds = load_dataset("Trelis/libritts-mimi-tokens")

# Just one split
clean_360 = load_dataset("Trelis/libritts-mimi-tokens", split="train.clean.360")

# Combine all train splits
all_train = load_dataset("Trelis/libritts-mimi-tokens",
                         split="train.clean.100+train.clean.360+train.other.500")

print(all_train[0])
```

## Companion datasets (same audio content, different tokenization)
- `Trelis/libritts-bpe-tokens`
- `Trelis/libritts-snac-tokens`
- `Trelis/libritts-neucodec-tokens`

## Reproducing

Encoder code: [TrelisResearch/audio-bits](https://github.com/TrelisResearch/audio-bits)
(see `scripts/encode_codec.py` and `scripts/tokenize_text.py`).

**Audio prep**: source LibriTTS-R is 24 kHz. NeuCodec (16 kHz) is downsampled
with `torchaudio.functional.resample(method="sinc_interp_kaiser")` (polyphase,
anti-aliased). Mimi and SNAC consume 24 kHz directly. Utterances > 20 s are
truncated to 20 s of audio (text is kept full).

**Per-row token count** = `min(ceil(n_audio_samples * T / max_padded_samples), T)`
where `T` is the codec's per-batch output length, so the count tracks the model's
own time grid and is exact for fixed-stride encoders.

## License

CC-BY-4.0 (matches LibriTTS-R source). Mimi model weights: CC-BY-4.0.