zheedong's picture
Upload README.md with huggingface_hub
0ecbcdd verified
|
Raw
History Blame Contribute Delete
1.99 kB
metadata
license: cc-by-nc-4.0
task_categories:
  - automatic-speech-recognition
  - text-to-speech
language:
  - en

LibriTTS (train, English) — DualCodec pre-tokenized

DualCodec (12 Hz) pre-tokenized speech for omni speech–vision MLLM training. Source: LibriTTS train.

Contents (train split)

Total samples ≈ 356,000 (audio–transcript pairs)
Total audio ≈ 556 hours
WebDataset shards see repo (.tar)
Mean / median duration 5.63s / 4.42s
p90 / max duration 11.58s / 29.83s
Length cap 30s

Counts/hours are computed by sampling shards (per-shard sample count × #shards); duration stats from a scanned subset.

Duration distribution

bucket share
0-5s 55.7%
5-10s 29.6%
10-15s 10.0%
15-20s 3.5%
>20s 1.2%

Format (WebDataset .tar)

Each sample shares a key and consists of {key}.sem.npy, {key}.ac.npy, {key}.txt:

  • .sem.npy — DualCodec semantic codes, int16, shape (T,), vocab 16384.
  • .ac.npy — DualCodec acoustic codes, int16, shape (7, T), vocab 4096 per codebook.
  • .txt — transcript.

Frame rate is 12 Hz, so duration in seconds = T / 12.

Length & padding

Codes are stored at their true variable length (no padding baked in). Samples longer than the training grid are handled at load time. During training, sequences are padded to a fixed 240-frame (20 s) grid by appending the DualCodec encoded-silence column (semantic code 3716, with its matching acoustic column) — i.e. padding is applied as code-level silence, not waveform zeros, so the padded region matches the codec's silence distribution. Samples exceeding the grid (> 240 frames) are skipped rather than cropped to preserve audio–text alignment.

Intended use

ASR (speech→text) and TTS (text→speech) pretraining. Semantic + acoustic codes reconstruct waveforms via the DualCodec decoder.