shangeth/libriasr-mimi-codes

LibriSpeech ASR — Kyutai Mimi Encoded

LibriSpeech ASR (train.clean.100) pre-encoded with the Kyutai Mimi neural audio codec.

Instead of raw waveforms, every utterance is stored as a compact matrix of discrete codec tokens. This format is ready to use directly in any language-model-style audio generation pipeline without needing a GPU encoder at training time.

What's inside

manifest.jsonl       # metadata — one JSON record per utterance
spk_index.json       # { "speaker_id": [idx, idx, ...] } — speaker-to-utterance index
shards/
├── shard_0000.pt    # packed dict of { idx -> (8, L) int16 code tensor }
├── shard_0001.pt
└── ...

Each manifest.jsonl record:

{
  "idx": 0,
  "text": "He was in a confused state of mind.",
  "codes_file": "shards/shard_0000.pt:0",
  "speaker_id": "1234",
  "n_frames": 198
}

spk_index.json maps each speaker ID to the list of utterance indices for that speaker, useful for sampling reference audio in speaker-conditioned tasks.

Dataset details

Source	LibriSpeech ASR train.clean.100
Speakers	~251
Utterances	~28,000
Total duration	~100 hours
Codec	Kyutai Mimi
Codec sample rate	24,000 Hz
Codec frame rate	12.5 fps
Codebooks	8
Token dtype	int16
License	CC BY 4.0

What you can use this for

• Multi-speaker / voice-cloning TTS research
• Speaker-conditioned codec language models
• Speaker representation learning
• Audio tokenization benchmarks
• Any task that benefits from a diverse, multi-speaker English speech corpus in discrete token form